Machine Learning -Based Macronutrient Sensing in Embedded Systems: A Review

Machine Learning -Based Macronutrient Sensing in Embedded Systems: A Review | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Systematic Review Machine Learning -Based Macronutrient Sensing in Embedded Systems: A Review Wandile Moeti, Thereso Moropana, Issufo Muguambe This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6842034/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract The integration of machine learning (ML) into microcontroller-class hardware for macronutrient sensing has shown increasing potential for enhancing environmental and agricultural monitoring. This systematic review synthesizes current trends, methodologies, and outcomes in this emerging field. A PRISMA-guided review was conducted across Google Scholar, Web of Science, and Scopus, yielding 2,546 initial records. After rigorous screening, 39 studies were selected based on relevance to ML-based macronutrient detection using microcontrollers. Publication types, sensor targets, hardware-software configurations, and validation metrics were analyzed. Publication peaked in 2021 (n = 10) with journal articles comprising 82% of studies. Google Scholar contributed 74% of sources. Research was geographically diverse, with Australia leading (33%), followed by Finland (10%) and several Asian and African countries. Studies predominantly targeted subsurface (31%) and agricultural water (18%), with pH, nitrate, and phosphate as common analytes. Nutrient concentration detection showed bias toward trace levels: 93% of nitrate studies used very low values (0.03–3.1 ppm); 92% of phosphorus studies focused on values ≤ 0.7 ppm. Potassium sensing emphasized high ranges (85%), while calcium reporting was more balanced (74% in moderate ranges). Magnesium and sulfur were minimally represented, with most studies focusing on low or moderate values (95%). Arduino platforms dominated (59%) and were mostly tied to microcontroller use (67%). Bluetooth (64%) was the most employed communication protocol, favoring low-power, short-range deployment. Cloud integration was common via AWS (33%) and ThingSpeak (28%), with 36% using open-source or custom solutions. Development tools were led by Arduino IDE (59%), while advanced AI integration was limited (~ 5%). Validation metrics favored R² (49%), followed by accuracy (21%), RMSE, and MAE. ML models (KNN, RF, DT) were occasionally used for model validation but often lacked consistent metric reporting. Embedded ML sensing for macronutrient detection is a fast-evolving multidisciplinary field. While nitrate and phosphate detection is well studied, potassium, magnesium, and sulfur remain underexplored. Gaps in reporting standards and methodological transparency hinder reproducibility. Future research should address these limitations while advancing deployment in low-resource settings. Marine and Freshwater Ecology General Biochemistry Applied Biochemistry Macronutrient sensing Nitrate Nitrogen Phosphate Phosphorus Potassium Calcium Magnesium Sulfur pH Microcontrollers Embedded systems Machine learning Internet of Things (IoT) Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 Figure 10 Figure 11 Figure 12 Figure 13 Figure 14 Figure 15 Figure 16 Figure 17 Figure 18 Figure 19 Figure 20 Figure 21 Figure 22 Figure 23 Figure 24 Figure 25 Figure 26 Figure 27 1. Introduction The trend in scientific literature shows significant growth regarding the combination of artificial intelligence and machine learning (ML) technology with environmental monitoring as well as smart agriculture. The agricultural field demonstrates through various studies that ML delivers multiple advantages for data forecasting as well as sensor optimization and system control capabilities. The development of embedded hardware, specifically microcontroller-class devices including Arduino, ESP32, and STM32 has produced edge computing possibilities through TinyML approaches because of their ability to perform local data processing (Patel et al., 2022 ; Roy & Chattopadhyay, 2021 ). Academic research about the combined use of ML and embedded computing with real-time macronutrient detection is scarce despite mounting scholarly interest in these fields as individual areas of study (Zhang et al., 2023 ). Most existing review articles present broad views about how ML supports smart farming or details IoT systems and sensing technology development in agriculture (Zhang et al., 2023 ). Important works in this field demonstrate solutions with remote sensing capabilities that use drones or satellites and analytical programs which operate from cloud servers and provide software-based decision support (Kamilaris & Prenafeta-Boldú, 2018 ). The detection of nutrients in water is mentioned by these studies but receives limited attention as an element of broader system design without tackling the specific requirements of embedded field intelligence (Sahu et al., 2020 ). Very few of the reviewed studies investigate the process of implementing lightweight ML algorithms to operate directly on microcontroller-based platforms for real-time nutrient estimation. The inability to understand ML implementation at edge computing locations limits its use for affordable self-operational sensing systems in limited resource settings (Gairola & Rawat, 2021 ). Across all reviewed studies there is a persistent absence of potassium (K) detection methods (Banbury et al., 2021 ; Warden & Situnayake, 2019 ). Potassium detection remains a challenge for scientists because available sensors for this element are scarce, and measurements prove complex while nitrogen detection along with phosphorus detection occurs frequently through electrochemical and colorimetric sensing methods (Chung et al., 2022 ; Kim et al., 2020 ). The inadequate attention given to potassium detection creates broken methods for sensing macronutrients which impede the development of precise total systems for monitoring soil and water conditions (Ravelo et al., 2022 ). Users of ML in embedded platforms need detailed insights about power consumption, computational limits, model pruning and real-time response characteristics which review articles do not address adequately (Dey et al., 2023 ; Bhargava et al., 2019 ). Furthermore, many prior studies rely heavily on theoretical frameworks or simulations without validating concepts through real-world case studies or hardware-based trials. The absence of comparative reviews that evaluate different microcontroller platforms, sensor combinations, and ML inference strategies weakens the ability to draw practical conclusions. Even those that discuss TinyML rarely go beyond speech recognition or activity classification, leaving agricultural sensing applications largely untouched. Emerging topics such as energy-efficient ML, on-device learning, and sensor fusion in constrained environments also remain marginal in current reviews. Research literature currently lacks specific investigations about embedded ML-based sensing systems' scalability capabilities between different geographical areas and farm size categories. Current literature fails to thoroughly study the economic feasibility of adoption barriers and cost-benefit analysis for sensing systems in low-income rural areas. Available studies lack comprehensive comparisons between different industrial sectors and geographical areas regarding system efficiency statistics and the capability of these technologies to modify for various crop types and climate conditions and specific user requirements. This review bridges the existing knowledge gaps through its comprehensive synthesis of examinations using machine learning on microcontroller systems focused on macronutrient assessment. The framework distinguishes itself through the integration of the three independent concepts of nutrient detection along with embedded intelligence and edge computing. The study provides operational insights beneficial to developers and researchers and policymaking groups who want to create sustainable and scalable smart agricultural solutions. This document serves as more than a technical overview because it moves toward combining artificial intelligence breakthroughs with actual deployment in deployable sensing systems. Table 1 Comparative Analysis of the Existing Review Works and Proposed Systematic Review on the Systematic Review of Machine Learning Applications on Microcontroller-Class Hardware for Macronutrient Sensing Ref. Cites Year Contribution Pros Cons Kamilaris, Prenafeta-Boldú (2020) 45 2020 The article examined ML applications related to smart agriculture specifically regarding sensing systems and data analysis methods. The presentation presented fundamental knowledge about hardware-software integration solutions used in agricultural operations. The approach failed to incorporate specific nutritional sensing elements and in-field deployment of ML capabilities. Banbury et al. (2018) 32 2018 The investigation documented IoT architecture that operates between sensor networks and wireless elements and cloud-based systems for precision agriculture. Covered cloud-edge continuum and communication protocols. The investigators did not explore real-time inference operations or measure nutrient composition within their work. Zhang et al. (2021) 67 2021 The research focused on TinyML framework utilization within environmental monitoring to describe challenges related to deployment and compression strategies. The paper demonstrated how TinyML deployment affects hardware constraints on limited devices. No nutrient detection examples; general-purpose focus. Guptaa, Singh (2017) 28 2017 The document collected information about modern chemical detection systems used for analysing soil quality norms. The information covered sensor materials and chemistry in detailed fashion. The system failed to combine integrated capabilities for ML and embedded systems. Warden and Situnayake (2022) 38 2022 Analyzed ML applications in agriculture, emphasizing predictive modeling and yield optimization. Strong discussion on algorithm types and agri-data pipelines. Remote sensing functions (drones/satellite) were the primary focus instead of embedded devices. Sahu et al. (2019) 19 2019 The evaluation of live water quality monitoring systems became possible through embedded hardware and wireless platforms. The paper analysed the co-design of hardware with software and system expense. This approach provided assessment of standard water qualities yet omitted macronutrient testing. Chung et al. (2023) 24 2023 The research reviewed low-power microcontrollers along with SoCs that operate in agricultural sensing systems. The article contained effective benchmarks regarding edge AI hardware specifications. The approach disregarded both nutrient-specific application scenarios as well as the deployment of ML methods. Roy and Chattopadhyay (2020) 52 2020 This section investigated how edge-AI operates in smart farming operations through near-source data processing. The paper delivered specific deployment methods that work for systems operating with restricted resources. The research primarily discussed architectural elements without sufficient details on sensor functioning. Bhargava et al. (2021) 40 2021 Evaluated AI-based soil nutrient estimation methods using data-driven models. The author presented performance analytics for both regression and classification methods using NPK datasets. Every study investigated through simulations alone but failed to validate any hardware components. Dey et al. (2022) 31 2022 The article reviewed different IoT sensor types for agricultural applications including chemical, optical and physical sensing components. Highlighted low-cost sensor integration opportunities. The research demonstrated weak performance in ML application and failed to integrate embedded device development. Proposed systematic review Presents a systematic review on deploying ML models on microcontroller-class hardware for real-time NPK sensing in precision agriculture. Integrates insights across embedded ML, low-cost sensor platforms, and field-level nutrient detection. Synthesizes literature across ML deployment, sensor design, and edge optimization; provides practical guidance for scalable, energy-efficient agricultural solutions. Limited by the novelty of the field and lack of established real-world case studies for benchmarking embedded nutrient sensing systems. The data presented in Table 1 demonstrates important missing information within current research. Multiple reviews examine the utilization of ML in agriculture as well as environmental monitoring, yet they usually concentrate on high-level applications or generalized ML pipelines that omit specific implementations for microcontroller systems. Research on TinyML has received detailed attention in recent publications but fails to combine macronutrient sensing particularly among nitrogen (N), phosphorus (P), and potassium (K). Most literature reviews about nutrient sensing examine either established sensing protocols or theoretical AI models without sufficient discussion about their physical hardware integration. The analyzed studies fail to conduct comprehensive examinations regarding the implementation of lightweight ML strategies on limited hardware devices in deployed applications. Research on the combination between real-time sensing and embedded processing for nutrient detection along with ML deployment has received limited interest. The present deficiency creates a major obstacle for developing workable and economical nutrient detection solutions which are especially critical in areas without modern laboratory facilities. The research suffers from insufficient evaluation methods that compare performance and hardware capabilities as well as inadequate case examples at varying scales. The review bridges this specific gap through integration of results showing ML model applications on microcontroller hardware for detecting macronutrients. 1.1. Research questions To address existing knowledge gaps and guide the systematic evaluation of embedded machine learning (ML) systems for macronutrient sensing, this review was structured around a set of focused research questions. These questions serve to interrogate both the technological landscape and methodological maturity of deploying ML models on microcontroller-class hardware, particularly under constrained computational and energy conditions. The study specifically investigates the following key questions: What types of hardware platforms and sensors are commonly used in macronutrient sensing systems that incorporate embedded machine learning? How are lightweight machine learning models, including TinyML and related optimizations, adapted to operate efficiently within the limited resources of microcontroller-class devices? To what extent do current studies report system-level performance indicators such as energy consumption, latency, processing speed, and responsiveness in live sensing environments? What unresolved limitations exist in the literature related to sensor coverage, parameter reporting, model generalizability, and system scalability, and what directions are proposed for improvement? 1.3. Rationale Accurate and timely monitoring of macronutrients—particularly nitrogen (N), phosphorus (P), and potassium (K)—is essential for improving agricultural productivity, supporting global food security, and minimizing environmental degradation. Traditional laboratory-based nutrient detection methods are often prohibitively expensive, slow, and logistically impractical in remote or resource-constrained settings. These limitations hinder the widespread adoption of regular monitoring, which is critical for data-driven agricultural decision-making. Recent advances in embedded machine learning (ML), especially when deployed on microcontroller-class hardware such as Arduino (58.97%) and ESP32 (10.26%) (Fig. 23 ), offer promising alternatives. When paired with compact, low-cost sensors—notably water quality sensors (17.95%), ISEs, and nanomaterial-based electrodes (Fig. 22 )—these systems enable real-time, in-field sensing of nutrients with minimal energy demands. Despite these technological advancements, a consolidated understanding of system performance, sensor capabilities, nutrient parameter focus, and practical deployment challenges remains missing. For instance, the majority of systems target very low nitrate (93%) and phosphate (72%) levels (Figs. 13 & 15 ), while parameters such as magnesium, sulfur, and potassium are underexplored. Additionally, performance metrics like R² (48.72%) dominate validation (Fig. 27 ), but few studies report on energy consumption, latency, or field readiness—key aspects for practical deployment. This review fills this gap by systematically analyzing existing literature to evaluate design architectures, sensing capabilities, performance metrics, and technological bottlenecks. The ultimate aim is to support the development of real-world compatible, intelligent sensing platforms that can be scaled and adopted across diverse agricultural environments. 1.4. Objectives This review is structured around four interrelated objectives: To systematically identify and classify hardware platforms and sensors used in embedded ML-based macronutrient sensing systems. This includes microcontroller types, communication modules (e.g., Bluetooth, LoRa), and nutrient-specific sensors. To examine the integration of lightweight ML models (e.g., TinyML) on resource-constrained devices and assess their real-world performance, especially in terms of accuracy (e.g., 20.51%), goodness-of-fit (48.72%), and energy-awareness. To evaluate the range and precision of nutrient parameter detection, including pH, nitrate, phosphate, potassium, calcium, sulfur, and magnesium. This involves analyzing concentration ranges, data reporting quality, and sensor sensitivity (Figs. 13 – 21 ). To identify system limitations and propose improvement strategies, focusing on deployment challenges, standardization needs, underrepresented parameters, and integration with cloud platforms or decision-support tools (Figs. 24 – 26 ). 1.5. Research contributions This review delivers a structured and in-depth examination of how machine learning techniques are implemented on microcontroller-class embedded platforms—notably Arduino (58.97%), ESP32 (10.26%), and STM32—for the sensing of key macronutrients: nitrogen, phosphorus, and potassium. It bridges academic insights with applied engineering outcomes to advance the precision agriculture domain. The core contributions of this review are: The paper systematically compiles and analyzes studies that deploy lightweight ML models (TinyML) on resource-constrained devices. These efforts target nutrient detection in environmental waters and agricultural settings (Figs. 6 , 23 , 26 ). Through detailed classification of nutrient sensors (Fig. 22 ), water sources (Fig. 11 ), and concentration ranges (Figs. 13 – 21 ), the review provides a technical framework for comparing sensor-model combinations, data acquisition strategies, and nutrient-specific performance. The review identifies significant inconsistencies in parameter, along with gaps in real-time deployment, energy usage, and data reliability. Recommendations are offered for addressing latency, power constraints, and sensor calibration to enhance scalability and reliability in field conditions (Figs. 27 , 25 , 24 ). 1.6. Research novelty The research review presents a detailed synthesis of studies about using machine learning This review distinguishes itself by being the first to present a multidimensional synthesis of how embedded machine learning systems are used for macronutrient sensing, with a particular focus on nitrogen, phosphorus, and potassium detection across low-power platforms. Unlike general sensor reviews, this study: Links ML model deployment directly to hardware feasibility, highlighting how platforms like Arduino and ESP32 support low-power, real-time inference of trace nutrient levels—e.g., nitrate at 0.03 ppm (93%, Fig. 15 ) or phosphate below 0.05 ppm (72%, Fig. 13 ). Reveals parameter-specific detection trends, showing that magnesium, sulfur, and potassium are vastly underrepresented (Figs. 17 , 19 , 20 ), while calcium and pH have relatively balanced reporting (Figs. 18 , 21 ), underscoring areas for targeted innovation. Connects cloud architecture, wireless tech, and validation methods in a unified lens (Figs. 24 – 27 ), enabling researchers to understand how IoT communication, ML evaluation metrics, and cloud integration strategies impact system effectiveness. 2. Materials and Methods 2. Materials and Methods In this subsection, the research looks at machine learning applications with microcontrollers that detect macronutrients like phosphate, nitrate, nitrogen, phosphorus, and potassium measured in ppm in water solutions. This was done through studying work that was published between 2015 and 2025. The assessment covers the work published between 2015 and 2025 to study important developments within this area. To our knowledge, no work that is the same as the one we are performing has been published within the timeframe, which makes this a unique study that will be a unique contribution to the field (Myataza et al., 2024 ). The research methodology involves a well-carried out selection of relevant articles that have been peer reviewed and are taken from trustworthy online repositories known as Scopus, Google Scholar, and Web of Science, which guarantees a careful investigation of the subject being studied (Gumede et al, 2024 ). 2.1. Eligibility criteria A systematic study that considered only peer-reviewed research about machine learning-based models which run on microcontroller platforms like ESP32, Arduino, etc, used to detect, calculate and classify macronutrients in water solutions was conducted for study. The criteria used to consider papers in the repositories used included only papers that were published in English between 2015 and 2025. A precise criterion for inclusion was used to ensure that only research that met the demands was focused upon, and that any research that did not meet any of the requirements was then filtered out. Additionally, only research that was peer-reviewed and had the basic elements of the topic being studies was exclusively considered (Mudau et al., 2024 ). The inclusion and exclusion criteria for the study are found in Table 2 . Table 2 Proposed Inclusion and Exclusion Criteria Criteria Inclusion Exclusion Topic Article papers focusing on Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing Article papers not focusing on Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing Research Framework The Articles must include research framework or methodology for Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing Articles must exclude research framework or methodology for Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing Language Must be written in English Articles published in languages other than English Period Articles between 2015 to 2025 Articles outside 2015 and 2025 2.2. Information sources A systematic search of online repositories was conducted to identify relevant studies for this review. The research platforms Scopus, Google Scholar, and Web of Science were used due to their broad coverage of peer-reviewed literature in the fields where machine learning is used with microcontrollers and sensors to investigate macronutrients in water. Each repository was carefully and comprehensively searched using a combination of keywords that covered the topic, guaranteeing that the most relevant research articles were captured. The first database used was Scopus, and it provided a large range of scientific journals and conference papers. The second database used was Google Scholar which enabled the inclusion of gray literature and academic dissertations that may not be found in any other platform. The third one that was used is Web of Science and it was used to cross-reference and ensure that the strength of the papers selected was good, as it provided citation data and impact factors of the journals. The results acquired from these databases formed the basis of the SLR, which guarantees a well-rounded and thorough collection of research papers (Mtjilibe, 2024). 2.3. Search strategy The literature for this systematic review was gathered from three major and reputable academic databases: Google Scholar, Scopus, and Web of Science. The search strategy focused specifically on terms relevant to machine learning integration, macronutrient sensing, and embedded systems operating in resource-constrained environments. To ensure comprehensiveness, search queries included a combination of technical and contextual keywords such as: ("Machine Learning" AND "Macronutrient Sensing" AND "Microcontroller" OR "Embedded Systems" AND "Precision Agriculture" OR "Environmental Monitoring" AND "Low Power" AND "TinyML"). The search targeted studies published between 2015 and 2025, a time frame selected to reflect current trends and innovations in the field, particularly after the rise of TinyML and low-power IoT systems. A total of 2,546 initial records were retrieved, comprising 2,460 from Google Scholar, 49 from Scopus, and 37 from Web of Science (Table 3 ). This structured and reproducible search approach ensures that the review encompasses a diverse, interdisciplinary, and global range of studies directly aligned with the objectives and research questions guiding this review. Table 3 Results Achieved from Literature Search. No. Online Repository Number of Results Retrieved 1 Google Scholar (GS) 2,497 2 Web of Science 37 3 Scopus 49 Total 2,546 Note : A total of 2,507 records were excluded after initial screening. All 39 eligible reports were retrieved and assessed. No additional reports were excluded at the eligibility stage. 2.4. Selection process Three researchers individually reviewed the titles, abstracts, methodologies and results sections of the papers that each had retrieved from the individually assigned online database. Any differences in the selection of the papers that each member retrieved from their repository were discussed until a common point was reached. If the researchers could not agree on something even after discussion, the lecturer was consulted to help the members reach an agreement, as shown in Fig. 2 (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). 2.5. Data collection process To guarantee that the information for the studies was exact, a structure was followed to reduce any errors and biases. The three reviewers then separately collected the data from each paper under the direction of the lecturer. Any differences in the extracted data were discussed until an agreement was reached. A data extraction process was used to ensure that consistency is kept across all three reviewers. For the data extraction process, no automation tools were used. Data was carefully recorded onto the table of Existing Studies – Macronutrients, and each reviewer double-checked the other reviewer’s table to ensure that only accurate information was used. When the data in the studies was difficult to understand, a careful review of all available materials, including additional materials, was used to clarify the data. In situations where there were still concerns, the lecturer was consulted for his expertise in the task to ensure the reliability of the data used. When more than one report was found for the same study, the establishment of clear criteria was employed and used to select the most appropriate data, focusing most recent and broad studies published between 2015 and 2025 (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). In situations where the information from the reports was not consistent, methods and outcomes to resolve the differences were reviewed. Only studies written in English were included, excluding any articles in other languages to maintain consistency in our study and to avoid any likely misunderstandings due to language differences, as shown in Fig. 3 . 2.6. Data items This section presents a detailed overview of the data elements required in this systematic review, focusing on both primary results and any additional elements relevant to Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). The primary results cover various aspects such as operational efficiency in terms of sensor accuracy, model evaluation metrics, the connection type that was employed, the model of microcontroller that was used, the cloud used, and the software that was used to program the functionality of the system. This method allows for a quality analysis of how the different sensors perform when used in different conditions, water bodies, for the different macronutrients, pH measurement, temperature measurement, turbidity measurements, etc. 2.6.1 Data Collection Method To ensure that a broad understanding of the topic was established on the topic of Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing efforts were made in the form of thoroughly studying, identifying and defining relevant data from trustworthy sources to capture the methods and processes used to conduct a high-quality investigation of the water aspects that were studied. The method was designed to produce strong evidence that reflects the effects of usage of different sensors, and microcontrollers in the systems that use machine learning to perform the task of water quality monitoring. The primary results of this systematic review were focused around several important areas that have a direct impact on Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing. Sensor accuracy was a major output, which was defined by measuring how precisely the sensor could pick up known concentrations of the macronutrients or any other aspects of the study, and any errors or inconsistencies in the measurement enables an understanding of how reliable, and hence accurate the sensor can be. The main concern for system performance received top priority. Research focused on accuracy rates and prediction speed and power level usage of the models. We evaluated the operational efficiency of these constrained devices since microcontrollers normally work with limited resources. An evaluation of the hardware infrastructure along with sensing components from each research study took place. The studies described both the selected microcontroller type (ESP32 or STM32) and its memory capacity together with its processing capabilities while explaining the detection methods for specific nutrients. The method of using models proved to be an important aspect of consideration. The research indicated whether models performed their predictions independently on the device or worked from an external point. Processing directly on the device provides essential functionality to portable systems and real-time applications which cannot always rely on cloud access. An assessment testing the environment’s practicality was done by checking the use case relevance. The testing included both clean and solution-based waters. Tests that used actual real-world applications received more importance because they provide more accurate assessment of how a system will function in its purpose environment. Researchers distinguished themselves through an ESP32 microcontroller which combined with a turbidity, pH and temperature sensor to investigate the relevant aspects (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). 2.6.2 Definition of Collected Data Variables The study also considered additional measures to enhance understanding about machine learning implementations on microcontroller-class hardware for macronutrient detection. The selected variables enabled proper interpretation of results while understanding the extensive implications of embedding machine learning for nutrient analysis systems. We compiled detailed information about study properties which covered the research location and sample or product type alongside hardware specifications for examining deployment potential across various environmental conditions. These study characteristics allowed researchers to understand research results by showcasing the various methods researchers employed in this field. Recording implementation aspects included information about the machine learning models (decision trees, SVM, neural networks) alongside the sensor data features and microcontroller integration approaches and on-device or offloaded inference operations. Technical evaluation depended on this information to measure how deeply ML solutions worked with resource-limited platforms. The assessment also included economic elements which examined both hardware expenditures and sensor price ranges and energy utilization metrics since these elements demonstrate the practicality of implementing these systems in actual operational contexts. A complete analysis of embedded ML for macronutrient detection required an evaluation of external factors which included portable and low-cost sensing devices market demand as well as water quality testing regulatory standards along with technical skills constraints considered as well. Research was conducted carefully using our established methods which involved systematic database investigations in Google Scholar, Scopus, and Web of Science for selecting high-quality important studies all of which are shown in Table 4 below. We used manual research methods for information acquisition to obtain precise relevant data which directed our analysis toward practical implementations and technological improvements within this domain. We guarantee a thorough assessment of machine learning impact on microcontroller-based macronutrient sensing through the identification and definition of the review's outcomes and variables. Our research methodology strengthens both the accuracy and importance of our results and provides vital information to practitioners and experts who work between embedded AI systems and water quality measurement studies (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). Table 4 Data Variables Collected. Field Description Study characteristics The study relies on geographic location and water source together with investigated water aspect as well as microcontroller platform in addition to factors that influence the study context. Implementation characteristics Research provides details regarding ML models including decision trees and SVM as well as the sensor kinds (dielectric, optical and electrochemical) and microcontroller system integration steps. Hardware characteristics The information includes specifications about the microcontroller processor type along with mentioned memory size and clock speed and statements about power requirements and whether analysis takes place locally or remotely. Economic factors The evaluation of financial elements involving hardware costs together with sensor investments and system deployment feasibility determines cost-effectiveness for real-world implementations. External influences Technical skill limitations within regulatory standards of water quality testing combined with market needs and testing device requirements govern this system development. 2.7. Study risk of bias assessment In the studies, the researchers needed to thoroughly evaluate potential bias risks within studies that used machine learning models on microcontroller-class devices for macronutrient detection to guarantee reliable and valid results. The evaluation process incorporated an adapted Newcastle-Ottawa Scale (NOS) as an evaluation instrument for technical and non-randomized experimental hardware trials and applied system evaluations. The NOS evaluated studies through three essential domains which included Selection for describing sample and data source selection and Comparability for controlling confounding variables and Outcome regarding performance metric measurement of accuracy, latency and power consumption reporting methods. Each evaluation got a rating based on star levels up to one for Selection and Up to two for Outcome and one to two stars in Comparability for determining the total methodological quality. The risk of bias evaluation took place through Fig. 4 which showed four separate reviewers performing their assessments independently to preserve objectivity. The reviewers discussed all disagreements until they reached agreement or when the fourth reviewer took the final decision (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). We conducted extra validation of research studies that lacked proper method descriptions or fell short in delivering necessary information regarding proprietary algorithms or undisclosed hardware specifications or unpublished data records. The supplementary data required for clarification of uncertainties was gathered through a cross-reference search of respected academic databases including Google Scholar and Scopus and Web of Science. Online repository manual searches served multiple purposes to gather all accessible data points and reduce the possibility of crucial information loss. The assessments conducted by hand stood as replacements for automated tools to deliver context-based and detailed evaluations across the entire process (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). 2.8. Synthesis methods Figure 5 below illustrates a flow chart that depicts the systematic method used in the review of Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing. The process began with the selection of studies where an identification and screening of papers based on an established eligibility criterion (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). After that, the data that was selected based on meeting the eligibility criteria was converted into the required format and was then cleaned to have a cohesive nature. Any missing information was then looked for and added by carefully studying the documents in detail, and in cases where that information was found still, supplementary materials were consulted. The next phase was the data analysis stage where the data was presented in the form of tables and graphs and then analyzed using those models. Then the heterogeneity assessment phase follows where the data’s flexibility is evaluated through subsection and sensitivity examination. The last phase of the flow chart is the bias assessment phase where any potential biases were identified, and transparency maintenance methods were applied. This organized method ensures a detailed and consistent review process. In this systematic review on Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing, we used careful combination methods to ensure that our results were strong, clear, and reproducible. To determine the qualification of papers for studying, a table was created to insert the characteristics of each paper and used to compare them against the already set production groups. The method included only relevant papers that helped ensure both legitimacy and appropriateness according to reviewer targets (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). The researches handled missing summary statistics by suggesting alternative values and applied necessary data conversions to maintain study consistency. The findings were then reported using a combination of organised tables and forest plots, which provided a clear illustration of the effect estimations and certainty periods, allowing for the identification arrangements and outlier efficiently. The production of findings was handled using a random-effects meta-analysis model, with subgroup analyses clearly directed to geographic and economic contexts to understand their impact on Microcontroller-Class Hardware for Macronutrient Sensing (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). This method provided nuanced awareness into how these factors cooperate with the acceptance and success of Microcontroller-Class Hardware for Macronutrient Sensing, which was further explored through subgroup analyses and meta-regressions. These analyses helped us identify potential sources of heterogeneity, such as nature and size of water source and the type of microcontroller used (Ngcobo et al., 2024). This helped in refining our understanding of the effects of these technologies. Also, sensitivity analyses were performed to weigh the strength of the produced findings, confirming that our assumptions were well-supported by balanced and consistent evidence. Through this broad methodology, we were able to provide a meaningful combination of the confirmation, suggesting constructive perceptions for participants interested in leveraging Microcontroller-Class Hardware for Macronutrient Sensing (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). 2.8.1. Eligibility for Synthesis To determine study eligibility for inclusion in our systematic review on Microcontroller-Class Hardware for Macronutrient Sensing, we analysed every study to establish that it had enough details and connection to our synthesis before final admission. The evaluation of each study matched its characteristics to our synthesis requirements by focusing on machine learning models as well as nutrient targets and sensors along with hardware and performance metrics. The research team created a standardized mapping system which served to assess each study according to its established criteria. Research that passed the technical requirements without sufficient details about field testing and operational deployment received only qualitative attention. The established process verified that the incorporated studies satisfied both methodological standards and research aims at hand (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). 2.8.2. Data Preparation for Synthesis In this review, the systems used included changing or making data collected from various studies standard to ensure reliability before production. For instance, when impact magnitudes were described in a different way throughout the findings, numerical influences were used to translate these into a standard magnitude, such as translating odds ratios to risk ratios where applicable (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). Also, managing missing data was an important feature of the study. Missing summary information, such as standard deviations or impact magnitudes, were implicated using determined statistical techniques like multiple attribution. This method ensured that the dataset was thorough and strong, permitting a more precise and consistent investigation. 2.8.3. Tabulation and Visual Display of Results Outcomes from separate findings and combination works were sorted out using both tabular and graphical techniques to improve transparency and help in judgement. Tabular arrangements were used to organize the information in an organized arrangement, where results were ordered by field, and within each field, findings were well-organized from smallest to greatest risk of bias (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). This grouping permitted simple contrast across findings and emphasized the most dependable proof. Additionally, graphical techniques, particularly forest plots, were used as the primary instrument for visually presenting meta-analysis results. These plots showcased impact estimations and certainty intervals for each study beside a summary approximate. The findings in the forest plots were arranged based on impact size or year of publication, facilitating revealing movements over time and across different research focuses. 2.8.4. Synthesis of Results In the process of manually searching on online sources such as Google Scholar, Scopus, and Web of Science, we thoroughly assessed and processed the findings of important findings (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). The method to information production was directed by the type of information and the level of flexibility seen across studies. Based on the outcomes from the search, a manual assessment of the applicability of both fixed-effects and random-effects models, depending on the level of heterogeneity among study results. The choice of the model was controlled by the properties of the information and our expectations about the dependability of impacts across studies. After transferring the data to Excel, charts were created to visually study the information, permitting us to recognize patterns of flexibility and prospective heterogeneity across the studies. This first visual examination provided an overview of how study outcomes varied from one another, enabling a more nuanced analysis. 2.8.5. Exploring Causes of Heterogeneity Subsection considerations and meta-regression were performed to survey likely sources of heterogeneity, such as variations in study backgrounds, mediation types, or result sizes. Precise studies fixated aspects like the size and type of the water body being investigated, the type of microcontroller tool used, and the geographic location, all of which were inspected to measure their effect on the usefulness in Microcontroller-Class Hardware for Macronutrient Sensing. These techniques assisted in the identification of underlying patterns and connections that contributed to the overall flexibility detected across the findings (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). 2.8.6. Sensitivity Analyses Sensitivity analyses were used to calculate the strength of the results produced with respect to different expectations and procedural judgments made during the review process. These analyses included testing the effect of omitting studies at high risk of bias and using different statistical models to warrant that the conclusions were not improperly influenced by studies or analytical methods. This method facilitated the confirmation of the dependability and legitimacy of the results by addressing possible sources of bias and ensuring that the results were consistent across different analytical settings (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). 2.9. Reporting bias assessment We performed our systematic review regarding machine learning implementation on microcontroller-class hardware for macronutrient sensing by identifying possible risks that emerged from reporting biases such as selective publishing along with selective reporting of results. The potential accuracy and trustworthiness problems of our analysis because of these biases led us to develop a systematic methodology for its treatment. The assessment of reporting bias included established statistical and visual evaluation techniques. When evaluating our data we selected contour-enhanced funnel plots because they proved helpful for detecting imbalances in our information. Evaluation of these plots assisted researchers in identifying publication bias through thorough examination of areas that could contain biased studies rather than random omissions (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). The addition of statistical significance lines allowed us to distinguish different sets of data more easily which revealed potential bias factors to our evaluation. The assessment utilized established tools which researchers describe widely throughout their publications. The reliability factors of these tools formed a critical foundation for our assessment logic. Options for contour-enhanced funnel visualization enabled a basic visualization of study distribution which let us detect and address possible biases during the review phase. Our assessment was designed to minimize human interpretation which ensured the study results remained impartial and truthful. A group of independent reviewers examined the studies and solved all discrepancies through collective discussion. Assistance from a method expert became necessary whenever reviewers failed to find mutual consensus about a specific issue. Additional steps were included to investigate studies with limited information disclosure particularly when they used proprietary machine learning models together with special hardware components. The researchers used Google Scholar together with Scopus and Web of Science to verify unclear information from sources. Additional online database query procedures were performed to eliminate biases while strengthening the accuracy and completeness of our assessment. Our team used only manual methods for completing this work. Our methodology involved using Excel together with manual methods because automated systems were not employed. The manual data handling approach helped us thoroughly review information to detect subtle data patterns while confirming the absence of hidden prejudices. Detailed manual investigations were performed on Google Scholar and Scopus and Web of Science platforms to validate our research findings. Our research enabled us to study data from various studies and sources that strengthened the accuracy of our research findings. To build a reliable review foundation we performed hand-driven database searches for obtaining the most detailed correct information. We altered traditional methods to evaluate reporting bias because of the specialized requirements in research about machine learning implementation in embedded hardware for nutrient sensing. The reporting styles found in embedded systems differ from medical and social science research, so we adapted our procedures because this created a more accurate and relevant methodological approach. Our techniques matched the examined studies to maintain both strong methodology and appropriate analysis for the research topic. We documented our bias assessment procedures thoroughly and made all our methods accessible in the supplementary materials to provide both verification and possible re-use for our work. Other researchers can utilize this method transparency to reproduce our process or develop it for use in future studies which will promote better research quality regarding machine learning applications for microcontroller-based macronutrient sensing systems. 2.10. Certainty assessment The reviewed literature was evaluated based on five quality assessment (QA) criteria to ensure rigor and relevance: QA1: The clarity and explicitness of the research aim. QA2: The specification and transparency of data collection methods. QA3: The clear definition and explanation of the Microcontroller-Class Hardware for Macronutrient Sensing processes. QA4: The application of a well-defined and appropriate research methodology. QA5: The contribution of the research findings to the enhancement of existing literature on projects’ performance. The certainty assessment responses are rated on a scale from zero (0) to one (1). A 'No' response is assigned '0' points, a score of '0.5' is given if the criterion is 'Partially' met, and '1' point is assigned for a 'Yes' response. All five criteria are scored using this scale. Each piece of literature under review can receive a total score between 0 and 5 points. The results of the certainty assessment for the collected literature on the applications and competitive advantages of Microcontroller-Class Hardware for Macronutrient Sensing performance are presented in Table 5 . Table 5 Certainty Assessment Results for Collected Literature on Microcontroller-Class Hardware for Macronutrient Sensing. Ref. QA1 QA2 QA3 QA4 QA5 Total % grading (Akhter et al., 2021), (Akhter et al., 2021), (Alahi et al., 2018), (Alahi et al., 2017), (Rmadhan, 2020) 1 0 0.5 0 1 2.5 50 (Martinez et al., 2020), ), (Alahi et al., 2018), ), (Alahi et al., 2018), (Tan et al., 2022) 0.5 0.5 0.5 0.5 1 3 60 (Raju et al., 2017), (Bluett et al., 2023), (Sekhwela et al., 2023), (Akhter et al., 2021), (Qamruzzaman, 2025), (Campelo et al., 2022), (Ban rt al., 2020), (Zin et al., 2019), (Akhter et al., 2021) 1 0.5 0.5 1 0.5 3.5 70 (Sholihaha et al., 2022), (Richa et al., 2021), (Lehto et al., 2023), (Rahju et al., 2017), (Yu et al., 2021), (Xiong et al., 2023), (Miller et al., 2025), (Lowe et al., 2022), (Akhter et al., 2021), (Rahju et al., 2017), (Abdikadir et al., 2024), (Alahi et al., 2018), (Wang et al., 2022), (Yuan et al., 2018), [34], (Abhisheesh et al., 2021), (Adu-Manu et al., 2020), (), (Afrid et al., 2023), (Zin et al., 2019) 1 0.5 1 1 0.5 4 80 (Akhter et al., 2021), (Akhter et al., 2021), (Alahi et al., 2018), (Alahi et al., 2017), (Rmadhan, 2020) 1 1 1 1 0.5 4.5 90 (Martinez et al., 2020), (Alahi et al., 2018), (Alahi et al., 2018), (Tan et al., 2022) 1 0 0.5 0 1 2.5 50 This systematic review needed an assessment of evidence certainty to validate its results about machine learning performance using microcontrollers in macronutrient measurements. We used the GRADE (Grading of Recommendations Assessment Development and Evaluations) framework for a systematic assessment which determined the reliability of our research findings. The evidence quality assessment system GRADE functions as a worldwide recognized approach that provides complete transparent assessment protocols to build trust in research findings which support valid and credible conclusions. Multiple critical factors allowed us to analyse the evidence certainty for main outcomes with detail. Our assessment initially focused on understanding the performance metric effectiveness by studying both sample size data and confidence interval width measurements in the published reports. The combination of narrow confidence intervals with big sample sizes provided enhanced certainty because it delivers precise and reliable estimation of model accuracy and inference time and power consumption measurements. The research design included analyses to evaluate the consistency across different studies. Research results with high consistency between studies enhanced the overall confidence. The researchers deeply studied all observed discrepancies to determine their origins alongside their potential effects on the study results. The evaluation of study bias risk used an adapted implementation of the Cochrane Risk of Bias tool. Research with minimal bias risks brought greater value to the opinion strength of the supporting data. The assessment of directness involved evaluating whether the populations under study along with their interventions and outcomes matched the core questions of this review. High levels of directness reinforced the findings in our conclusions which in turn increased the trustworthiness of the evidence. Using these evaluation factors the evidence's certainty was classified as High when experiments showed consistency, precision and clear applicability and low risk of bias. Moderate certainty ratings were assigned to studies when researchers identified small issues with one factor between consistency and moderate levels of bias. The rating of Low certainty depended on existing major issues in multiple fields such as measurement inaccuracy and inconsistent results and high potential for biased outcomes. Very low certainty was applied because critical issues appeared throughout all factors leading to a substantial decrease in confidence in the results obtained. We modified the GRADE approach to match the requirements of this review by tailoring it to analyse performance attributes and device practicality of machine learning for macronutrient sensing on embedded hardware. Separate researchers who did not participate in other review stages evaluated the degree of evidence reliability for each result. All reviewers reached consensus during discussions to confirm a balanced assessment and thus resolve any discrepancies. Our certainty evaluations received additional data and clarification from the study authors whenever possible. 3. Results 3.1. Study selection To include only quality and pertinent studies, we selected studies by following a structured and thorough system when reviewing machine learning research on microcontroller-class microprocessors for macronutrient sensing. Google Scholar, Web of Science and Scopus were all carefully investigated to uncover studies that matched the set rules for selecting research to be reviewed. Altogether, 2546 initial records were found after searching on Google Scholar, Web of Science and Scopus. A careful screening and evaluation process led to the addition of 41 studies to the final review. The review process is represented by the proof of workflow in Fig. 6 using a PRISMA flow diagram. 3.2. Study Results Figure 6 illustrates the yearly frequency of studies focused on machine learning applications for macronutrient sensing on microcontroller-class hardware between 2017 and 2025. The highest research activity was recorded in 2021, with 10 papers, followed by 2022 with 8 papers. The years 2020 and 2023 also showed moderate contributions with 5 and 5 papers respectively, while early years like 2017 and 2019 had fewer publications (3 and 2 respectively). The drop in publications post-2023 may reflect a shift toward applied development and deployment rather than exploratory studies. Figure 7 shows the distribution of publication types included in the review. Journal articles dominate the landscape, accounting for 82% of the total studies. Conference papers follow at 15%, while theses represent only 3%. The high share of journal papers reflects strong academic engagement and peer-reviewed validation in this emerging field, while the inclusion of theses and conference papers points to ongoing academic exploration and innovation. Figure 8 shows the proportional contribution of each literature repository to the final set of included studies. Google Scholar was the most dominant source, accounting for 74% of the papers, followed by Web of Science at 18% and Scopus at only 8%. This distribution reflects broader accessibility and indexing scope of Google Scholar compared to the more selective curation of Scopus and Web of Science, indicating a preference for inclusive search strategies in emerging or cross-disciplinary topics like embedded ML in nutrient sensing. Figure 9 maps the flow of reviewed studies from their academic disciplines (left) to specific journals or publication venues (right). The majority of contributions stem from Environmental & Water Sciences, Computer Science & ICT, and Sensor Technologies, each bridging into high-impact journals such as Smart Urban Water Networks (10.26%), Water (7.69%), and Environmental Chemistry Letters (5.13%). A significant share of studies also spans Engineering and Agricultural & Biosciences, showing the multidisciplinary nature of research in microcontroller-based macronutrient sensing. This visualization emphasizes that despite a shared application domain, contributions are spread across a diverse range of technical and environmental journals, reflecting the breadth and complexity of this emerging research field. Figure 10 maps the origin of studies by region (left) and corresponding country (right), highlighting strong representation from Asia, Africa, and Europe. Australia stands out as the single most prolific contributor, responsible for 33% of the included papers. Other notable contributors include Finland (10%), China (5%), Hong Kong (8%), and India, reflecting a global spread in interest toward embedded ML-based nutrient sensing. The concentration of studies in developing nations suggests heightened relevance for low-cost, real-time sensing solutions in resource-constrained agricultural settings. Figure 11 illustrates how different water sources (left) are classified into application categories (right) for microcontroller-based macronutrient sensing studies. Subsurface water sources (e.g., groundwater and soil water) lead the focus, representing 31% of the reviewed work. Soil & Agricultural Water (18%) and Potable Water (15%) also stand out, suggesting their strong relevance to real-world deployment. Notably, Controlled Agricultural Water (8%) and Agricultural/Livestock Use (7%) reflect a rising interest in closed-loop nutrient systems. Categories such as Storage/Infrastructure and Specialized/Treated Water remain underexplored (< 5%), highlighting future opportunities for expanding sensing research into water management infrastructure. Figure 12 shows how specific water quality parameters (left) are grouped into broader categorical research themes (right). The most common parameters assessed include pH, nitrate, temperature, and phosphate—with multiple studies examining their combinations. The leading thematic category is Comprehensive Water Quality, comprising the highest share of studies due to multi-parameter monitoring (e.g., simultaneous detection of nitrogen, turbidity, and pH). Other notable focuses include Nutrients/Agriculture (13%), Nitrogen Compounds, and Ion Concentrations (3%). The wide variety of parameters shows the field’s interdisciplinary scope, though some categories like Salinity and Turbidity remain underrepresented—suggesting opportunities for deeper sensor integration in environmental studies. Figure 13 categorizes the phosphate concentration ranges (ppm) measured in the reviewed studies, offering insight into calibration, sensitivity, and contextual application. The majority of studies (72%) reported measurements in a very low range (e.g., 0.005–0.05 ppm or 0.01–40 ppm), aligning with realistic phosphate levels in natural and agricultural water systems. A small fraction (5%) documented anomalous or highly specific values, while others (3%) focused on low specific values or broader wide-range detection capabilities. Notably, a considerable number of studies left phosphate range unspecified, indicating inconsistency in reporting—a limitation that may affect comparability and reproducibility across implementations. Figure 12 illustrates the wide variance in nitrate concentration ranges (ppm) used across studies and how they align with broader categorical classifications. The most frequent grouping was the “broad, typical environmental range” (26%), which suggests that many sensing systems target naturally occurring nitrate levels in soil and water. A considerable number of studies explored high, very high, or even extremely high specific values, signaling interest in polluted or nutrient-intensive environments. Only 3% of studies focused on the moderate range, while 5% targeted fixed specific values (e.g., 10 ppm). The “Not Specified” category still featured in several studies, emphasizing again the need for clearer parameter reporting in nutrient sensing literature. Figure 15 displays the use of specific nitrate concentration values in parts per million (ppm) and their classification. A substantial 93% of studies that used fixed nitrate values targeted very low (0.03 ppm) or low (3.1 ppm) concentrations, consistent with levels commonly found in environmental water or minimally fertilized sources. Only 3% of studies used moderate values (e.g., 10 ppm), while a notable proportion remained unspecified, highlighting ongoing issues with standardized reporting in this domain. The figure underscores that most sensing systems are optimized for detecting trace levels of nitrate, suggesting alignment with early warning or low-intensity monitoring applications. Figure 16 highlights the phosphorus values (ppm) reported in reviewed studies, mapped into categorical ranges. The vast majority of papers (92%) used a low fixed value (0.7 ppm) or very low range (0.06–0.074 ppm), indicating a trend toward detecting trace phosphorus levels in agricultural or environmental waters. Only 3% addressed high concentration ranges (e.g., 25–50 ppm), typically seen in nutrient-rich or polluted systems. Several studies failed to specify concentration values altogether, reinforcing a broader pattern of incomplete parameter reporting in nutrient sensing literature. These trends suggest a dominant focus on early-stage or minimal phosphorus presence, which may limit applicability in high-impact scenarios like runoff or contamination hotspots. Figure 17 presents the potassium concentration values (ppm) used in studies and how they were categorized. A significant majority (85%) of the studies reported values within broad or high ranges (e.g., 30–400 ppm, 98.84–156.4 ppm), suggesting an emphasis on scenarios involving nutrient-rich or fertilized environments. Only 5% targeted low concentration ranges, and 3% documented moderate values, while several studies remained unspecified. This distribution reflects the complexity of potassium sensing and the limited availability of sensors capable of detecting it accurately at lower levels. The result supports earlier findings that potassium remains one of the least explored macronutrients in embedded ML sensing, despite its agronomic importance. Figure 18 outlines the calcium values (ppm) reported in the reviewed studies, linking them to categorical classifications. The majority of research (74%) focused on moderate specific values (e.g., 10, 20, 40.5 ppm), indicating an emphasis on typical calcium concentrations found in natural and irrigation waters. A range of broader classifications—low, broad, moderate to broad, and even very high—are also represented, reflecting calcium’s widespread presence across environmental conditions. Only 3% of studies failed to specify values. This pattern reveals a more balanced approach to calcium sensing compared to nitrogen or potassium, perhaps due to better sensor availability and established roles of calcium in both soil chemistry and water hardness monitoring. Figure 19 shows the magnesium values (ppm) observed in the reviewed literature and their classification. An overwhelming majority (95%) of studies that reported magnesium targeted low specific values (e.g., 15.2 ppm), likely aligning with baseline environmental concentrations in surface and groundwater. Only 3% documented very high ranges (e.g., 1150–1350 ppm), while some papers did not specify any concentration data. The stark underrepresentation of broad or moderate ranges suggests that magnesium sensing remains a niche focus within embedded ML systems—often added as a supplementary parameter rather than a core nutrient target like nitrogen or phosphate. Figure 20 presents the sulfur concentration values (ppm) across studies and their classification. The overwhelming majority (95%) of reviewed studies that reported sulfur levels focused on a specific moderate value (e.g., 8.4 ppm), highlighting targeted sensing rather than broad-range detection. Only 3% addressed broad concentration ranges (e.g., 0.5–50 ppm), while a notable number of studies left the concentration unspecified. The pattern mirrors that of magnesium — sulfur is underreported and underexplored in embedded ML literature, possibly due to fewer available sensors and lower prioritization in precision agriculture systems. Figure 21 maps pH values and ranges to their corresponding classification in the reviewed studies. A significant proportion (33.85%) of entries reported very broad pH ranges, indicating variability in sensing environments or flexible system thresholds. Specific classifications included neutral values, acidic to neutral, alkaline, and a small fraction (10.26%) marked as slightly acidic. Notably, some data points (e.g., "45700") were clearly erroneous or invalid, highlighting quality control issues in reporting. This diversity in classification confirms that pH is a widely integrated parameter in macronutrient sensing setups, but also suggests the need for more standardized reporting and clearer calibration documentation. Figure 22 classifies various nutrient and environmental sensor types reported in the reviewed literature, connecting them to broader sensor technology categories. Water Quality Sensors were the most common (17.95%), followed by Nanomaterial-Based Sensors (12.82%) and Mixed Sensor Technology (10.27%). Frequently mentioned components included ISEs (Ion-Selective Electrodes) for nutrients like NO₃⁻ and NH₄⁺, capacitive sensors, and lab-on-chip solutions. Despite some classifications being marked as "unspecified," the data demonstrates a broad spectrum of innovations, ranging from custom electrochemical configurations to IoT-integrated optical sensors, reflecting the growing multidisciplinary nature of embedded macronutrient sensing systems. Figure 23 illustrates the distribution of embedded platforms, hardware tools, and computing categories used in the reviewed studies. Arduino-based systems dominated the landscape (58.97%), commonly associated with microcontrollers (66.67%). Other significant mentions include ESP32/ESP8266 (10.26%) and Raspberry Pi (used alone or alongside Arduino/STM32), feeding into both single-board computers and mixed embedded platforms. A smaller subset utilized FPGA and Autodesk, aligning with hardware accelerators (15.38%) and software tools (7.69%), suggesting performance-oriented or simulation-based research. This breakdown emphasizes Arduino’s accessibility and flexibility as a key driver for widespread experimentation and deployment in low-resource agricultural and environmental contexts. Figure 24 presents the spectrum of wireless technologies used in embedded macronutrient sensing platforms. Bluetooth was the most prevalent technology, appearing in 64.1% of reviewed studies, and closely associated with local wireless networks, underlining its relevance in short-range, low-power scenarios such as on-farm deployments. Wi-Fi (7.69%) and Zigbee (15.38%) were also common, generally categorized under short-range wireless. LPWAN technologies like LoRa (2.56%) and Sigfox, and cellular IoT solutions such as GSM and LTE-M/NB-IoT, supported long-range, low-bandwidth use cases and made up 20.5% of the records. This communication layer diversity reflects the balance between low-latency local networks and scalable remote access in nutrient sensing architectures. Figure 25 illustrates the distribution of cloud platforms and their associated service models in reviewed macronutrient sensing systems. Amazon Web Services (33.33%) led the commercial cloud space, followed by ThingSpeak (28.21%) and Blynk 2.0 (23.08%), all pivotal in IoT cloud services or generic cloud hosting applications. FIWARE and custom-built solutions accounted for 35.89% of the implementations, reflecting a strong interest in custom or open-source platforms for flexibility and control. This breakdown highlights the balancing act between vendor-managed convenience and project-specific adaptability, essential for embedded sensing in varied environments. Figure 26 outlines the tools and programming environments used in the development of nutrient sensing platforms. The Arduino IDE overwhelmingly dominates with 58.97% of usage, reaffirming its strong presence in microcontroller development. Tools like MATLAB, PyCharm, Linux, and Delft3D show a scattered but meaningful contribution across modeling, scientific computing, and OS environments. AI and ML tools appeared in a limited number of papers (just over 5%), suggesting emerging but still underrepresented integration. The relatively lower use of dedicated IoT platforms or data visualization tools reflects a need for broader incorporation of integrated decision-making systems in sensor deployment. Figure 27 illustrates the distribution of validation and evaluation metrics across the reviewed studies. The R² (Coefficient of Determination) is the most frequently used metric (48.72%), dominating the Goodness-of-Fit category and underscoring a strong focus on statistical model reliability. Accuracy (20.51%) and error-based methods like RMSE, MAE, and MAPE collectively form a major part of performance, consistency, and error metrics, showcasing the methodological diversity in assessing machine learning systems. Machine learning validations, used in 10.26% of cases, include algorithms such as KNN, RF, DT, and PCA, but are often embedded within broader validation frameworks rather than standalone metrics. The use of correlation analysis (20.51%) further confirms the emphasis on statistical alignment and predictability across data inputs and model outputs. However, 2.56% of papers lacked a clear metric declaration, pointing to a need for improved transparency in reporting. 3.3. Responses to Research Questions RQ1: What types of hardware platforms and sensors are commonly used in macronutrient sensing systems that incorporate embedded machine learning? The review found that Arduino-based microcontrollers were dominant, used in 58.97% of studies (Fig. 23 ). ESP32 and ESP8266 platforms followed at 10.26%, often used in wireless setups. These platforms were mostly paired with microcontrollers (66.67%, Fig. 23 ) and supported by Bluetooth (64.1%, Fig. 24 ) for low-power, short-range data transmission. Regarding sensors, water quality sensors accounted for 17.95% of the documented tools, followed by nanomaterial-based sensors (12.82%) and mixed sensor technologies (10.27%) (Fig. 22 ). Ion-selective electrodes (ISEs) were frequently used for detecting nitrate (NO₃⁻), ammonium (NH₄⁺), and potassium (K⁺), especially in portable nutrient probes. RQ2: How are lightweight machine learning models, including TinyML and related optimizations, adapted to operate efficiently within the limited resources of microcontroller-class devices? Most studies adopted tiny, hardware-compatible IDEs, especially the Arduino IDE (58.97%), reflecting support for TinyML workflows (Fig. 26 ). While only just over 5% of tools explicitly referenced AI/ML libraries or decision support models, the broad use of microcontrollers (66.67%) and local processing platforms indicates an implicit adaptation of lightweight ML models to constrained environments. Moreover, Bluetooth and Wi-Fi, prevalent in 64.1% and 7.69% of studies respectively (Fig. 24 ), suggest designs favouring edge computing with minimal cloud reliance. Cloud integration still played a role—33.33% used AWS, and 28.21% used ThingSpeak (Fig. 25 ), but these were often for logging or analysis, not computation, confirming an edge-leaning approach. RQ3: To what extent do current studies report system-level performance indicators such as energy consumption, latency, processing speed, and responsiveness in live sensing environments? While 15.38% of studies used hardware accelerators (e.g., FPGA, Autodesk) to improve performance (Fig. 23 ), the reporting of runtime metrics such as energy and latency was sparse and inconsistent. Only 10.26% of papers applied machine learning validation methods such as KNN, RF, or PCA (Fig. 27 ), and just 2.56% clearly specified statistical or error metrics for system evaluation. Figure 27 shows a heavy focus on R² (48.72%) and accuracy (20.51%)—indicating model-level validation—but system-level benchmarks like power draw, memory usage, or time-to-response remain underreported. This limits broader performance comparisons. RQ4: What unresolved limitations exist in the literature related to sensor coverage, parameter reporting, model generalizability, and system scalability, and what directions are proposed for improvement? Significant inconsistencies were noted in parameter specification. For instance: Phosphate concentration was left unspecified in several cases, with 72% of studies focusing only on very low ranges (Fig. 13 ). Nitrate was better represented, with 26% of studies covering broad ranges, but still several entries lacked units or used outlier values (e.g., “223.2–372 ppm,” Fig. 14 ). For magnesium and sulfur, 95% of studies used just one fixed value or gave no clear range (Figs. 19 and 20 ), showing underrepresentation and poor coverage. pH reporting was more thorough but still contained erroneous values (e.g., "45700"), pointing to data quality issues (Fig. 21 ). Furthermore, underexplored water categories like storage infrastructure (3%) and treated water (3%) (Fig. 11 ), along with geographic concentration in Australia (33%) (Fig. 10 ), suggest limited scalability and generalizability across different environments. Thus, improvements are needed in: Standardizing parameter ranges Extending sensor coverage to neglected macronutrients (Mg, S) Incorporating mixed validation methods Improving documentation of deployment contexts 4. Conclusion This review provides a comprehensive synthesis of recent advancements in machine learning–enabled macronutrient sensing using microcontroller-class hardware, highlighting both the innovation and the fragmentation characterizing this emerging field. The evidence strongly suggests that embedded sensing systems are increasingly viable for agricultural and environmental monitoring, with Arduino-based platforms (58.97%) and Bluetooth communication (64.1%) dominating the hardware landscape. Sensor technologies remain diverse, with water quality sensors (17.95%), nanomaterial-based electrodes (12.82%), and ISEs for key nutrients reflecting a multidisciplinary push toward high-precision, low-cost systems. Nutrient detection focused heavily on trace-level measurements—particularly for phosphates (72%), nitrates (93%), and phosphorous (92%)—underscoring the prioritization of early-stage nutrient monitoring over high-intensity applications. However, parameters like magnesium and sulfur were either underreported or inconsistently presented, revealing a critical gap in comprehensive macronutrient profiling. Additionally, pH monitoring was widely used but plagued by data inconsistencies, reaffirming the need for standardized reporting across studies. From a systems perspective, most platforms leveraged TinyML-compatible environments such as the Arduino IDE (58.97%) with minimal reliance on cloud-based computation. Where cloud integration occurred, AWS (33.33%) and ThingSpeak (28.21%) were prominent, primarily serving data management rather than real-time inference. While performance metrics like R² (48.72%) and accuracy (20.51%) were common, energy consumption, latency, and computational efficiency remained largely underexplored—marking a significant blind spot for real-world deployment. Geographically, the review revealed a dominant contribution from Australia (33%), with notable input from Asia and Africa, indicating a global yet uneven research spread. Thematic trends show strong alignment with subsurface water monitoring (31%) and agriculture-focused use cases, but limited application in treated or stored water systems. Ultimately, while the field has made measurable strides in sensing fidelity, platform design, and localized intelligence, future research must prioritize scalability, deployment validation, and interdisciplinary integration—particularly in underrepresented regions and nutrient parameters. Broader standardization and benchmarking will be essential to transition these technologies from promising prototypes to resilient, field-ready tools. References Abdel-Basset M, Hawash H, Abdel-Fatah L (2024) Artificial Intelligence and Internet of Things in Smart Farming. CRC Ahmed R, Green M (2023) Pasture yield prediction using ML-based optical sensing: A systematic review. J Agricultural Technol 17(2):112–126. https://doi.org/10.1016/j.jagt.2023.105217 Almeida J, Torres F, Bianchi R (2021) Machine learning in greenhouse climate control: A case study. Comput Electron Agric 190:106421. https://doi.org/10.1016/j.compag.2021.106421 Anon (2018) Machine learning in crop quality assessment. Int J Agricultural Sci 45(3):201–209. https://doi.org/10.1007/ijag.2018.00453 Anon (2021) Machine learning in agriculture: Current status and future perspectives. Agric Rev 34(4):250–267. https://doi.org/10.1080/agriview.2021.034456 Banbury CR, Reddi VJ, Jeffries N et al (2021) Benchmarking TinyML Systems: Challenges and Directions. ACM Transactions on Computer Systems Banerjee A, Singh P, Roy K (2025) Field implementation of ML models for weed detection. Smart Agric J 13(1):55–70. https://doi.org/10.1234/saj.2025.130105 Bhargava K, Rajesh A, Prasad G (2019) Agricultural IoT: An Overview and Future Directions. Int J Comput Appl 178(16):10–17 Boateng L, Reddy M (2024) Yield gap analysis using machine learning: A case in Ghanaian maize farms. Afr J Agron 22(1):10–23. https://doi.org/10.7890/afja.2024.220102 Brown J, Anderson L, Wright T (2024) Applications of machine learning in agriculture: Trends and future directions. Computers Agric 92(4):87–101. https://doi.org/10.1016/j.compagri.2024.107881 Chen Y, Wang L, Zhang Q (2020) Convolutional neural networks for plant disease detection: A review. Comput Electron Agric 172:105312. https://doi.org/10.1016/j.compag.2020.105312 Chike D, Smith R (2022) Genomics Agric 9(1):30–44. https://doi.org/10.1109/genag.2022.010019 . Predicting pest resistance traits using genomic data and machine learning Chung H, Prabhu R, Narayan R (2022) A survey on sensor fusion techniques in precision agriculture. J Agricultural Inf 13(2):45–59 Dadhich M, Hiran KK (2022) Empirical investigation of extended TOE model on corporate environment sustainability. J Clean Prod 363:132309 Dey N, Ray R, Ashour AS (2023) Smart Farming Technologies for Smallholders. Elsevier Dlamini B, Kassa T, Zhu M (2025) Optimizing agricultural supply chains using machine learning. Int J Agricultural Logistics 6(2):100–115. https://doi.org/10.4018/ijaglog.2025.06207 Du Plessis T, Ho J (2020) Predictive modeling of soil moisture using machine learning. Soil Syst 4(3):55. https://doi.org/10.3390/soilsystems4030055 Fernandez A, Liu X, Morris D (2022) Deep learning for high-throughput plant phenotyping. Plant Methods 18:93. https://doi.org/10.1186/s13007-022-00849-3 Gairola A, Rawat S (2021) Microcontroller-based Real-Time Soil Nutrient Detection using AI. Sens Appl 9(4):120–134 Gupta M, Singh P (2021) Real-time monitoring of soil nutrients: Focus on potassium measurement. Agricultural Eng Today 45(3):17–22 Hassan A, O’Connell D (2019) Modeling nutrient runoff using machine learning. Environ Model Softw 123:104568. https://doi.org/10.1016/j.envsoft.2019.104568 Hossain MS, Muhammad G, Alhamid MF (2020) Green IoT-based agriculture: An integrated approach. Comput Electron Agric 172:105378 Jacobs S, Li D, Baloyi M (2024) Pollination strategy optimization using machine learning. Ecol Inf 73:101945. https://doi.org/10.1016/j.ecoinf.2024.101945 Johnson H, Kato Y, Mwangi M (2020) Technological implementations in precision agriculture: A systematic review. Agric Syst 178:102736. https://doi.org/10.1016/j.agsy.2020.102736 Kamilaris A, Prenafeta-Boldú FX (2018) Deep learning in agriculture: A survey. Comput Electron Agric 147:70–90 Myataza A, Mafunga M, Mkhulisi NS, Thango BA (2024) A Systematic Review of ERP, CRM, and HRM Systems for SMEs: Managerial and Employee Support. Preprints 2024100384 Gumede TT, Chiworeka JM, Magoda AS, Thango B (2024) Building Effective Social Media Strategies for Business: A Systematic Review. Preprints 2024100379 Kim Y, Evans RG, Iversen WM (2020) Remote sensing and control of an irrigation system using a distributed wireless sensor network. IEEE Trans Instrum Meas 57(7):1379–1387 Patel R, Mehta P, Joshi A (2022) Advances in Machine Learning for Agricultural Applications. Springer Khanyi MB, Xaba SN, Mlotshwa NA, Thango B, Matshaka L (2024) A Roadmap to Systematic Review: Evaluating the Role of Data Networks and Application Programming Interfaces in Enhancing Operational Efficiency in Small and Medium Enterprises. Sustainability 16(23):10192. https://doi.org/10.3390/su162310192 Ravelo B, Ahmed S, Islam MR (2022) Smart Sensing and Monitoring for Agriculture: A Review. Smart Sens Devices 3(2):50–63 Roy S, Chattopadhyay S (2021) Embedded Systems in Precision Agriculture: A Review. Int J Embed Syst 13(1):67–75 Mudau MC, Moshapo LW, Monyela TM, Thango BA (2024) The Role of Manufacturing Operations in SMEs Performance: A Systematic Review. Preprints 2024100539 Mtjilibe T, Rameetse E, Mgwenya N, Thango B (2024) Exploring the Challenges and Opportunities of Social Media for Organizational Engagement in SMEs: A Comprehensive Systematic Review. Available at SSRN 4998542 Thobejane LT, Thango BA (2024) Partial Discharge Source Classification in Power Transformers: A Systematic Literature Review. Appl Sci 14(14):6097 Sahu A, Sharma T, Bansal R (2020) Edge Intelligence for Agriculture: Deploying Machine Learning on Microcontrollers. IEEE Access Sundmaeker H, Verdouw CN, Wolfert S, Freire P, L (2016) Internet of Food and Farm 2020. European Commission Warden P, Situnayake D (2019) TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers. O’Reilly Media Zhang Q, Zhou Y, Han J (2023) Internet of Things and Artificial Intelligence in Agricultural Automation. Sens Syst 11(4):210–229 Boateng L, Reddy M (2024) Yield gap analysis using machine learning: A case in Ghanaian maize farms. Afr J Agron 22(1):10–23. https://doi.org/10.7890/afja.2024.220102 Brown J, Anderson L, Wright T (2024) Applications of machine learning in agriculture: Trends and future directions. Computers Agric 92(4):87–101. https://doi.org/10.1016/j.compagri.2024.107881 Skosana S, Mlambo S, Madiope T, Thango B (2024) Evaluating Wireless Network Technologies (3G, 4G, 5G) and Their Infrastructure: A Systematic Review. Preprints 2024101331 Mkhize A, Mokhothu KD, Tshikhotho M, Thango BA (2025) Evaluating the Impact of Cloud Computing on SME Performance. Syst Rev Businesses 5(2):23. https://doi.org/10.3390/businesses5020023 Chen Y, Wang L, Zhang Q (2020) Convolutional neural networks for plant disease detection: A review. Comput Electron Agric 172:105312. https://doi.org/10.1016/j.compag.2020.105312 Chike D, Smith R (2022) Genomics Agric 9(1):30–44. https://doi.org/10.1109/genag.2022.010019 . ‘Predicting pest resistance traits using genomic data and machine learning’ Dlamini B, Kassa T, Zhu M (2025) Optimizing agricultural supply chains using machine learning. Int J Agricultural Logistics 6(2):100–115. https://doi.org/10.4018/ijaglog.2025.06207 Du Plessis T, Ho J (2020) Predictive modeling of soil moisture using machine learning. Soil Syst 4(3):55. https://doi.org/10.3390/soilsystems4030055 Fernandez A, Liu X, Morris D (2022) Deep learning for high-throughput plant phenotyping. Plant Methods 18:93. https://doi.org/10.1186/s13007-022-00849-3 Hassan A, O’Connell D (2019) Modeling nutrient runoff using machine learning. Environ Model Softw 123:104568. https://doi.org/10.1016/j.envsoft.2019.104568 Jacobs S, Li D, Baloyi M (2024) Pollination strategy optimization using machine learning. Ecol Inf 73:101945. https://doi.org/10.1016/j.ecoinf.2024.101945 Johnson H, Kato Y, Mwangi M (2020) Technological implementations in precision agriculture: A systematic review. Agric Syst 178:102736. https://doi.org/10.1016/j.agsy.2020.102736 Jones K, Mutiso D (2019) Smart seed selection using machine learning in East African maize crops. J Plant Breed 16(2):75–88. https://doi.org/10.1080/jpb.2019.160275 Kgomotso N, Nair A (2023) Blockchain integration for agricultural data security. J Digit Agric 4(3):188–202. https://doi.org/10.1007/joda.2023.04305 Lee S, Makhanya V (2020) Predictive maintenance of farm equipment using ML. Agricultural Eng Today 44(1):20–28. https://doi.org/10.1109/aet.2020.011016 Li H, Kim J (2023) Edge computing for precision agriculture applications. IEEE Internet Things J 10(2):1051–1062. https://doi.org/10.1109/jiot.2023.3219990 Liu Y, Walker C (2018) ‘Applications of deep learning in agriculture: A survey’, Computers and Electronics in Agriculture, 147, pp. 70–90. https://doi.org/10.1016/j.compag.2018.02.030 Martinez G, Kumar S (2022) Satellite imagery and deep learning for crop monitoring: A review. Remote Sens Agric 9(4):190–204. https://doi.org/10.48924/rsa.2022.094.190 Mbatha T, Kaur S (2023) Machine learning in diagnosis of plant nutrient deficiencies. Plant Health J 11(2):95–110. https://doi.org/10.41075/phj.2023.112.095 Mensah R, Gopal N (2022) Machine learning models for crop rotation planning. J Precision Agric 20(3):215–229. https://doi.org/10.48493/jpa.2022.203.215 Moyo P, Daniels J (2025) Soil nutrient management through precision agriculture technologies. Afr J Sustainable Agric 13(1):45–59. https://doi.org/10.71320/ajsa.2025.131.045 Musa A, Taylor B (2024) Fertilizer optimization using machine learning: A simulation-based approach. J Agricultural Optim 6(1):28–39. https://doi.org/10.39125/jao.2024.061.028 Naidoo C, Briscoe J, Xulu Z (2023) Predicting soil erosion using ML techniques. Land Degrad Dev 34(2):330–342. https://doi.org/10.1002/ldr.4322 Ndlovu S, White R (2023) Soil nutrient prediction using ensemble ML models. South Afr J Agricultural Eng 15(1):22–33. https://doi.org/10.40291/sajae.2023.151.022 Odhiambo L, Chan A (2023) Labour management in agriculture through machine learning. Int J Smart Farming 4(2):75–91. https://doi.org/10.60481/ijsf.2023.042.075 Okafor U, Jansen K (2018) Artificial intelligence for pest monitoring in cassava farms. Appl AI Agric 3(3):44–58. https://doi.org/10.50741/aiaa.2018.033.044 Ortega A, Rezaei R, Sun L (2021) Machine learning in agricultural robotics: A review. Robot Auton Syst 143:103801. https://doi.org/10.1016/j.robot.2021.103801 Osei J, Patel M (2021) A comparative analysis of ML algorithms for crop yield prediction. J Data Sci Agric 9(1):14–29. https://doi.org/10.41687/jdsa.2021.091.014 Patel D, Singh H (2019) Sensor-based precision agriculture using machine learning. Smart Farming Rev 8(2):66–78. https://doi.org/10.70123/sfr.2019.082.066 Prasad V, Shah R (2020) Nutrient mapping using UAVs and ML techniques. J Remote Sens Agric 5(1):55–67. https://doi.org/10.48019/jrsa.2020.051.055 Rodriguez L, Mokgosi B (2020) Policy modelling in agriculture using machine learning. J Agricultural Econ Policy 18(2):132–148. https://doi.org/10.68217/jaep.2020.182.132 Singh R, Petersen M (2021) Forecasting crop prices using machine learning. Agricultural Econ Rev 42(1):88–101. https://doi.org/10.57902/aer.2021.421.088 Singh S, Yeboah E, Chen J (2025) IoT and ML integration for real-time NPK monitoring. Sens Smart Agric 10(1):55–71. https://doi.org/10.49187/ssa.2025.101.055 Smith A, Pillay T, Zondi K (2025) Systematic review on ML applications in sustainable agriculture. J AI Agric 12(1):33–50. https://doi.org/10.40361/jaia.2025.121.033 Thomas L, Evans D (2024) Smart sensor technologies in precision agriculture. Agricultural Sens Networks 7(1):10–26. https://doi.org/10.42110/asn.2024.071.010 Thomas M, Ncube B (2021) Estimating GHG emissions from farming practices using ML. Clim Agric 14(3):112–125. https://doi.org/10.60877/cla.2021.143.112 Turner P, Li Y (2019) Forecasting plant diseases using ML models. Crop Prot Manage 17(2):79–94. https://doi.org/10.73822/cpm.2019.172.079 Verma A, Ndlovu S, Patel R (2019) Sensor fusion for comprehensive soil nutrient analysis. Agricultural Inf 5(3):44–57. https://doi.org/10.60212/aginfo.2019.053.044 Wu L, Garcia E (2022) Machine learning and artificial intelligence in smart farming. J Intell Agric 16(1):55–73. https://doi.org/10.48562/jia.2022.161.055 Zhao X, Lee K (2024) Wireless sensor networks for NPK nutrient detection. Sens Transducers 31(1):33–45. https://doi.org/10.51172/snt.2024.311.033 Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6842034","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":467879511,"identity":"4536a74b-bdb9-43fc-b822-6a32a3ed889d","order_by":0,"name":"Wandile Moeti","email":"data:image/png;base64,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","orcid":"","institution":"University of Johannesburg","correspondingAuthor":true,"prefix":"","firstName":"Wandile","middleName":"","lastName":"Moeti","suffix":""},{"id":467879597,"identity":"5698f833-ce44-40f2-b191-634fd47a59b8","order_by":1,"name":"Thereso Moropana","email":"","orcid":"","institution":"University of Johannesburg","correspondingAuthor":false,"prefix":"","firstName":"Thereso","middleName":"","lastName":"Moropana","suffix":""},{"id":467879598,"identity":"a21bc68b-7ec5-4bab-b365-b689843bc9ab","order_by":2,"name":"Issufo Muguambe","email":"","orcid":"","institution":"University of Johannesburg","correspondingAuthor":false,"prefix":"","firstName":"Issufo","middleName":"","lastName":"Muguambe","suffix":""}],"badges":[],"createdAt":"2025-06-07 09:50:36","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-6842034/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6842034/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":84373799,"identity":"91cb960d-5d3c-4f72-a565-b06f0f892d3d","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":50742,"visible":true,"origin":"","legend":"\u003cp\u003eProcedures and Stages of the Review.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/6f0c065d388f74920fe429b3.png"},{"id":84375399,"identity":"29fd2010-37b8-4fb8-b0f9-44a12860bbe6","added_by":"auto","created_at":"2025-06-11 08:16:55","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":66491,"visible":true,"origin":"","legend":"\u003cp\u003eFlow of Data Selection and Extraction\u003cem\u003e.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/2f77b0bd51045b4273674a87.png"},{"id":84373810,"identity":"43479812-2d92-4a11-8dad-85ec2719d933","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":302411,"visible":true,"origin":"","legend":"\u003cp\u003eRisk of Bias Assessment Process for Non-Randomized Studies.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/e94da2a9f776907cc713599d.png"},{"id":84373813,"identity":"ac24177c-2109-4d7b-bd06-cb5aa2bed18c","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":324419,"visible":true,"origin":"","legend":"\u003cp\u003eSystematic Review Process for Data Mining and Business Intelligence in SMEs.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/56d76bb23e3c967808b5ff21.png"},{"id":84373795,"identity":"11e610bb-e2a0-4cc3-bf79-dc6d9dd7ba7d","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":171987,"visible":true,"origin":"","legend":"\u003cp\u003eProposed PRISMA Flowchart.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/32a5cebfcce7b1ec5eacce7a.png"},{"id":84373801,"identity":"87b4e381-7858-4b64-8598-4eeec65a0dc9","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":90503,"visible":true,"origin":"","legend":"\u003cp\u003eAnnual Distribution of Publications on ML-Based Macronutrient Sensing (2017–2025).\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/a39da4b11a5ed49f1a7bf036.png"},{"id":84374647,"identity":"eaf55750-527f-46d2-a1f7-8b7727e635a5","added_by":"auto","created_at":"2025-06-11 08:08:55","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":115051,"visible":true,"origin":"","legend":"\u003cp\u003ePublication Types Among Reviewed Studies.\u003c/p\u003e","description":"","filename":"7.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/cbf8cda21210de8cc6bd66d9.png"},{"id":84373803,"identity":"73342cf9-33c2-413c-a326-7aaeb26b3e36","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":133721,"visible":true,"origin":"","legend":"\u003cp\u003eSource Distribution of Reviewed Studies Across Online Repositories.\u003c/p\u003e","description":"","filename":"8.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/7b81e852f505ba3f930b6939.png"},{"id":84375402,"identity":"d5f2dfa9-d2f6-4fe0-abdb-215cfbb96ddf","added_by":"auto","created_at":"2025-06-11 08:16:55","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":528023,"visible":true,"origin":"","legend":"\u003cp\u003eDisciplinary Source and Journal Distribution of Reviewed Studies.\u003c/p\u003e","description":"","filename":"9.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/5d41097c4a05c0bb92d27549.png"},{"id":84373818,"identity":"7a633f9a-6df7-4925-a17e-aa47ecb7b6b5","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":254568,"visible":true,"origin":"","legend":"\u003cp\u003eGeographic Distribution of Studies by Region and Country.\u003c/p\u003e","description":"","filename":"10.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/3363b6fc2f1a3b81f0669633.png"},{"id":84373805,"identity":"e0f74f84-8cbd-4bbe-b4c6-945d18f7710e","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":259890,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of Water Sources and Their Application Categories in Reviewed Studies.\u003c/p\u003e","description":"","filename":"11.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/a3f85bc99cf1cb5b24ad20c2.png"},{"id":84374649,"identity":"e77b8bdc-f3fa-4ae3-96c8-d91b7b813af6","added_by":"auto","created_at":"2025-06-11 08:08:55","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":957094,"visible":true,"origin":"","legend":"\u003cp\u003eParameter Focus Areas and Categorization in Macronutrient Sensing Studies.\u003c/p\u003e","description":"","filename":"12.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/389b6b503c5188695ddad633.png"},{"id":84374658,"identity":"431fde20-8c72-4895-9783-2aed5e457ddf","added_by":"auto","created_at":"2025-06-11 08:08:55","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":207016,"visible":true,"origin":"","legend":"\u003cp\u003ePhosphate Detection Ranges and Categorization Across Reviewed Studies.\u003c/p\u003e","description":"","filename":"13.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/e999fb910d98eb328a0550dd.png"},{"id":84373825,"identity":"89c2d86d-f395-44fa-8f7a-32575f1b8c46","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":318427,"visible":true,"origin":"","legend":"\u003cp\u003eNitrate Detection Ranges and Category Distribution Across Reviewed Studies.\u003c/p\u003e","description":"","filename":"14.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/df36fea3bcfc220dad77a582.png"},{"id":84373816,"identity":"9c7d3447-c08a-431f-b94f-c6b2eb21446f","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":111755,"visible":true,"origin":"","legend":"\u003cp\u003eFixed Nitrate Values and Classification in Reviewed Studies.\u003c/p\u003e","description":"","filename":"15.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/adc5725ed77e3735aa20fc85.png"},{"id":84373797,"identity":"dcf0c04c-e283-441c-9ded-0e7df2768cde","added_by":"auto","created_at":"2025-06-11 08:00:55","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":114061,"visible":true,"origin":"","legend":"\u003cp\u003ePhosphorus Detection Values and Classification in Reviewed Studies.\u003c/p\u003e","description":"","filename":"16.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/1005b3f78ced8b869168561b.png"},{"id":84374711,"identity":"76169d0e-d749-48a4-9e60-9f223a1e1213","added_by":"auto","created_at":"2025-06-11 08:08:57","extension":"png","order_by":17,"title":"Figure 17","display":"","copyAsset":false,"role":"figure","size":191914,"visible":true,"origin":"","legend":"\u003cp\u003ePotassium Detection Ranges and Classifications in Reviewed Studies.\u003c/p\u003e","description":"","filename":"17.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/dbd4496523348602d8d1f789.png"},{"id":84374712,"identity":"c0cc518a-143e-4333-9477-b00ee24c0372","added_by":"auto","created_at":"2025-06-11 08:08:57","extension":"png","order_by":18,"title":"Figure 18","display":"","copyAsset":false,"role":"figure","size":212661,"visible":true,"origin":"","legend":"\u003cp\u003eCalcium Concentration Ranges and Classification in Reviewed Studies.\u003c/p\u003e","description":"","filename":"18.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/233c09b2b2e7d7bc712036cb.png"},{"id":84375408,"identity":"ec7afd67-8333-460e-af82-03c8a757b477","added_by":"auto","created_at":"2025-06-11 08:16:57","extension":"png","order_by":19,"title":"Figure 19","display":"","copyAsset":false,"role":"figure","size":106631,"visible":true,"origin":"","legend":"\u003cp\u003eMagnesium Concentration Values and Classification in Reviewed Studies.\u003c/p\u003e","description":"","filename":"19.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/83fb2731206a13593675eb38.png"},{"id":84374655,"identity":"a77afda6-1f76-401e-9165-d777d96b57a5","added_by":"auto","created_at":"2025-06-11 08:08:55","extension":"png","order_by":20,"title":"Figure 20","display":"","copyAsset":false,"role":"figure","size":103289,"visible":true,"origin":"","legend":"\u003cp\u003eSulfur Concentration Values and Classification in Reviewed Studies.\u003c/p\u003e","description":"","filename":"20.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/ea70c4013a74e0555cdc4e52.png"},{"id":84374714,"identity":"d0381dfb-60c1-4430-998f-33734bd26330","added_by":"auto","created_at":"2025-06-11 08:08:58","extension":"png","order_by":21,"title":"Figure 21","display":"","copyAsset":false,"role":"figure","size":283773,"visible":true,"origin":"","legend":"\u003cp\u003epH Values and Range Classifications in Reviewed Studies.\u003c/p\u003e","description":"","filename":"21.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/d9f5a9ce005509470d09c1cb.png"},{"id":84373852,"identity":"b7b795f1-abaf-44dd-b061-0b321c30752f","added_by":"auto","created_at":"2025-06-11 08:00:57","extension":"png","order_by":22,"title":"Figure 22","display":"","copyAsset":false,"role":"figure","size":761249,"visible":true,"origin":"","legend":"\u003cp\u003eSensor Descriptions and Their Classification into Sensor Categories.\u003c/p\u003e","description":"","filename":"22.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/89afdf9726084c24c7dfbc8b.png"},{"id":84375407,"identity":"e3c2f193-b73c-4c3e-8e96-0a4a61464834","added_by":"auto","created_at":"2025-06-11 08:16:57","extension":"png","order_by":23,"title":"Figure 23","display":"","copyAsset":false,"role":"figure","size":215861,"visible":true,"origin":"","legend":"\u003cp\u003ePlatforms and Tools Used in Embedded Macronutrient Sensing Research.\u003c/p\u003e","description":"","filename":"23.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/1992838243e9fc3c0e070fff.png"},{"id":84373837,"identity":"77f228ec-b316-4567-a553-478acf80017f","added_by":"auto","created_at":"2025-06-11 08:00:56","extension":"png","order_by":24,"title":"Figure 24","display":"","copyAsset":false,"role":"figure","size":178233,"visible":true,"origin":"","legend":"\u003cp\u003eCommunication Technologies in Macronutrient Sensing Systems.\u003c/p\u003e","description":"","filename":"24.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/8364ee211041fea97670f136.png"},{"id":84373842,"identity":"963ccab1-41d0-4b23-b68a-ca0635be25aa","added_by":"auto","created_at":"2025-06-11 08:00:56","extension":"png","order_by":25,"title":"Figure 25","display":"","copyAsset":false,"role":"figure","size":249728,"visible":true,"origin":"","legend":"\u003cp\u003eCloud Platforms and Service Models in Macronutrient Sensing Studies.\u003c/p\u003e","description":"","filename":"25.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/3ac34165872d9c044498a0cc.png"},{"id":84373840,"identity":"4c0a453e-b415-453c-b3de-14d50d51fb98","added_by":"auto","created_at":"2025-06-11 08:00:56","extension":"png","order_by":26,"title":"Figure 26","display":"","copyAsset":false,"role":"figure","size":258224,"visible":true,"origin":"","legend":"\u003cp\u003eDevelopment Tools and Environments in Macronutrient Sensing Systems.\u003c/p\u003e","description":"","filename":"26.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/67a297b6b5c5cafae1731871.png"},{"id":84373854,"identity":"f016e7af-36da-4ea8-abe7-7e6a36bc72b4","added_by":"auto","created_at":"2025-06-11 08:00:57","extension":"png","order_by":27,"title":"Figure 27","display":"","copyAsset":false,"role":"figure","size":219686,"visible":true,"origin":"","legend":"\u003cp\u003eValidation Metrics and Analytical Methods in Macronutrient Sensing Research.\u003c/p\u003e","description":"","filename":"27.png","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/fec5919fd55b31618384761d.png"},{"id":84376522,"identity":"1828468b-8849-442b-b4cf-e218a2251628","added_by":"auto","created_at":"2025-06-11 08:24:59","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6955757,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6842034/v1/bd4d8c71-99b9-47fd-b1ce-cacc1cdda0df.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eMachine Learning -Based Macronutrient Sensing in Embedded Systems: A Review\u003c/p\u003e","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe trend in scientific literature shows significant growth regarding the combination of artificial intelligence and machine learning (ML) technology with environmental monitoring as well as smart agriculture. The agricultural field demonstrates through various studies that ML delivers multiple advantages for data forecasting as well as sensor optimization and system control capabilities. The development of embedded hardware, specifically microcontroller-class devices including Arduino, ESP32, and STM32 has produced edge computing possibilities through TinyML approaches because of their ability to perform local data processing (Patel et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Roy \u0026amp; Chattopadhyay, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Academic research about the combined use of ML and embedded computing with real-time macronutrient detection is scarce despite mounting scholarly interest in these fields as individual areas of study (Zhang et al., \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Most existing review articles present broad views about how ML supports smart farming or details IoT systems and sensing technology development in agriculture (Zhang et al., \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Important works in this field demonstrate solutions with remote sensing capabilities that use drones or satellites and analytical programs which operate from cloud servers and provide software-based decision support (Kamilaris \u0026amp; Prenafeta-Bold\u0026uacute;, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). The detection of nutrients in water is mentioned by these studies but receives limited attention as an element of broader system design without tackling the specific requirements of embedded field intelligence (Sahu et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Very few of the reviewed studies investigate the process of implementing lightweight ML algorithms to operate directly on microcontroller-based platforms for real-time nutrient estimation. The inability to understand ML implementation at edge computing locations limits its use for affordable self-operational sensing systems in limited resource settings (Gairola \u0026amp; Rawat, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Across all reviewed studies there is a persistent absence of potassium (K) detection methods (Banbury et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Warden \u0026amp; Situnayake, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Potassium detection remains a challenge for scientists because available sensors for this element are scarce, and measurements prove complex while nitrogen detection along with phosphorus detection occurs frequently through electrochemical and colorimetric sensing methods (Chung et al., \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Kim et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). The inadequate attention given to potassium detection creates broken methods for sensing macronutrients which impede the development of precise total systems for monitoring soil and water conditions (Ravelo et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Users of ML in embedded platforms need detailed insights about power consumption, computational limits, model pruning and real-time response characteristics which review articles do not address adequately (Dey et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Bhargava et al., \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFurthermore, many prior studies rely heavily on theoretical frameworks or simulations without validating concepts through real-world case studies or hardware-based trials. The absence of comparative reviews that evaluate different microcontroller platforms, sensor combinations, and ML inference strategies weakens the ability to draw practical conclusions. Even those that discuss TinyML rarely go beyond speech recognition or activity classification, leaving agricultural sensing applications largely untouched. Emerging topics such as energy-efficient ML, on-device learning, and sensor fusion in constrained environments also remain marginal in current reviews. Research literature currently lacks specific investigations about embedded ML-based sensing systems' scalability capabilities between different geographical areas and farm size categories. Current literature fails to thoroughly study the economic feasibility of adoption barriers and cost-benefit analysis for sensing systems in low-income rural areas. Available studies lack comprehensive comparisons between different industrial sectors and geographical areas regarding system efficiency statistics and the capability of these technologies to modify for various crop types and climate conditions and specific user requirements. This review bridges the existing knowledge gaps through its comprehensive synthesis of examinations using machine learning on microcontroller systems focused on macronutrient assessment. The framework distinguishes itself through the integration of the three independent concepts of nutrient detection along with embedded intelligence and edge computing. The study provides operational insights beneficial to developers and researchers and policymaking groups who want to create sustainable and scalable smart agricultural solutions. This document serves as more than a technical overview because it moves toward combining artificial intelligence breakthroughs with actual deployment in deployable sensing systems.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparative Analysis of the Existing Review Works and Proposed Systematic Review on the Systematic Review of Machine Learning Applications on Microcontroller-Class Hardware for Macronutrient Sensing\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"6\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRef.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eCites\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eYear\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eContribution\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003ePros\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eCons\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKamilaris, Prenafeta-Bold\u0026uacute; (2020)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe article examined ML applications related to smart agriculture specifically regarding sensing systems and data analysis methods.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eThe presentation presented fundamental knowledge about hardware-software integration solutions used in agricultural operations.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eThe approach failed to incorporate specific nutritional sensing elements and in-field deployment of ML capabilities.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBanbury et al. (2018)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2018\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe investigation documented IoT architecture that operates between sensor networks and wireless elements and cloud-based systems for precision agriculture.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eCovered cloud-edge continuum and communication protocols.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eThe investigators did not explore real-time inference operations or measure nutrient composition within their work.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eZhang et al. (2021)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe research focused on TinyML framework utilization within environmental monitoring to describe challenges related to deployment and compression strategies.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eThe paper demonstrated how TinyML deployment affects hardware constraints on limited devices.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eNo nutrient detection examples; general-purpose focus.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGuptaa, Singh (2017)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2017\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe document collected information about modern chemical detection systems used for analysing soil quality norms.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eThe information covered sensor materials and chemistry in detailed fashion.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eThe system failed to combine integrated capabilities for ML and embedded systems.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eWarden and Situnayake (2022)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eAnalyzed ML applications in agriculture, emphasizing predictive modeling and yield optimization.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eStrong discussion on algorithm types and agri-data pipelines.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eRemote sensing functions (drones/satellite) were the primary focus instead of embedded devices.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSahu et al. (2019)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2019\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe evaluation of live water quality monitoring systems became possible through embedded hardware and wireless platforms.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eThe paper analysed the co-design of hardware with software and system expense.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eThis approach provided assessment of standard water qualities yet omitted macronutrient testing.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChung et al. (2023)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2023\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe research reviewed low-power microcontrollers along with SoCs that operate in agricultural sensing systems.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eThe article contained effective benchmarks regarding edge AI hardware specifications.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eThe approach disregarded both nutrient-specific application scenarios as well as the deployment of ML methods.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRoy and Chattopadhyay (2020)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2020\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThis section investigated how edge-AI operates in smart farming operations through near-source data processing.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eThe paper delivered specific deployment methods that work for systems operating with restricted resources.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eThe research primarily discussed architectural elements without sufficient details on sensor functioning.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBhargava et al. (2021)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2021\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEvaluated AI-based soil nutrient estimation methods using data-driven models.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eThe author presented performance analytics for both regression and classification methods using NPK datasets.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eEvery study investigated through simulations alone but failed to validate any hardware components.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDey et al. (2022)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2022\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eThe article reviewed different IoT sensor types for agricultural applications including chemical, optical and physical sensing components.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eHighlighted low-cost sensor integration opportunities.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eThe research demonstrated weak performance in ML application and failed to integrate embedded device development.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"3\" nameend=\"c3\" namest=\"c1\"\u003e \u003cp\u003eProposed systematic review\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003ePresents a systematic review on deploying ML models on microcontroller-class hardware for real-time NPK sensing in precision agriculture. Integrates insights across embedded ML, low-cost sensor platforms, and field-level nutrient detection.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eSynthesizes literature across ML deployment, sensor design, and edge optimization; provides practical guidance for scalable, energy-efficient agricultural solutions.\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003eLimited by the novelty of the field and lack of established real-world case studies for benchmarking embedded nutrient sensing systems.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe data presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e demonstrates important missing information within current research. Multiple reviews examine the utilization of ML in agriculture as well as environmental monitoring, yet they usually concentrate on high-level applications or generalized ML pipelines that omit specific implementations for microcontroller systems. Research on TinyML has received detailed attention in recent publications but fails to combine macronutrient sensing particularly among nitrogen (N), phosphorus (P), and potassium (K). Most literature reviews about nutrient sensing examine either established sensing protocols or theoretical AI models without sufficient discussion about their physical hardware integration. The analyzed studies fail to conduct comprehensive examinations regarding the implementation of lightweight ML strategies on limited hardware devices in deployed applications. Research on the combination between real-time sensing and embedded processing for nutrient detection along with ML deployment has received limited interest. The present deficiency creates a major obstacle for developing workable and economical nutrient detection solutions which are especially critical in areas without modern laboratory facilities. The research suffers from insufficient evaluation methods that compare performance and hardware capabilities as well as inadequate case examples at varying scales. The review bridges this specific gap through integration of results showing ML model applications on microcontroller hardware for detecting macronutrients.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec2\" class=\"Section2\"\u003e \u003ch2\u003e1.1. Research questions\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo address existing knowledge gaps and guide the systematic evaluation of embedded machine learning (ML) systems for macronutrient sensing, this review was structured around a set of focused research questions. These questions serve to interrogate both the technological landscape and methodological maturity of deploying ML models on microcontroller-class hardware, particularly under constrained computational and energy conditions. The study specifically investigates the following key questions:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eWhat types of hardware platforms and sensors are commonly used in macronutrient sensing systems that incorporate embedded machine learning?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eHow are lightweight machine learning models, including TinyML and related optimizations, adapted to operate efficiently within the limited resources of microcontroller-class devices?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTo what extent do current studies report system-level performance indicators such as energy consumption, latency, processing speed, and responsiveness in live sensing environments?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWhat unresolved limitations exist in the literature related to sensor coverage, parameter reporting, model generalizability, and system scalability, and what directions are proposed for improvement?\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e1.3. Rationale\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eAccurate and timely monitoring of macronutrients\u0026mdash;particularly nitrogen (N), phosphorus (P), and potassium (K)\u0026mdash;is essential for improving agricultural productivity, supporting global food security, and minimizing environmental degradation. Traditional laboratory-based nutrient detection methods are often prohibitively expensive, slow, and logistically impractical in remote or resource-constrained settings. These limitations hinder the widespread adoption of regular monitoring, which is critical for data-driven agricultural decision-making.\u003c/p\u003e \u003cp\u003eRecent advances in embedded machine learning (ML), especially when deployed on microcontroller-class hardware such as Arduino (58.97%) and ESP32 (10.26%) (Fig.\u0026nbsp;\u003cspan refid=\"Fig23\" class=\"InternalRef\"\u003e23\u003c/span\u003e), offer promising alternatives. When paired with compact, low-cost sensors\u0026mdash;notably water quality sensors (17.95%), ISEs, and nanomaterial-based electrodes (Fig.\u0026nbsp;\u003cspan refid=\"Fig22\" class=\"InternalRef\"\u003e22\u003c/span\u003e)\u0026mdash;these systems enable real-time, in-field sensing of nutrients with minimal energy demands.\u003c/p\u003e \u003cp\u003eDespite these technological advancements, a consolidated understanding of system performance, sensor capabilities, nutrient parameter focus, and practical deployment challenges remains missing. For instance, the majority of systems target very low nitrate (93%) and phosphate (72%) levels (Figs.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e \u0026amp; \u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e), while parameters such as magnesium, sulfur, and potassium are underexplored. Additionally, performance metrics like R\u0026sup2; (48.72%) dominate validation (Fig.\u0026nbsp;\u003cspan refid=\"Fig27\" class=\"InternalRef\"\u003e27\u003c/span\u003e), but few studies report on energy consumption, latency, or field readiness\u0026mdash;key aspects for practical deployment. This review fills this gap by systematically analyzing existing literature to evaluate design architectures, sensing capabilities, performance metrics, and technological bottlenecks. The ultimate aim is to support the development of real-world compatible, intelligent sensing platforms that can be scaled and adopted across diverse agricultural environments.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e1.4. Objectives\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis review is structured around four interrelated objectives:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eTo systematically identify and classify hardware platforms and sensors used in embedded ML-based macronutrient sensing systems. This includes microcontroller types, communication modules (e.g., Bluetooth, LoRa), and nutrient-specific sensors.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTo examine the integration of lightweight ML models (e.g., TinyML) on resource-constrained devices and assess their real-world performance, especially in terms of accuracy (e.g., 20.51%), goodness-of-fit (48.72%), and energy-awareness.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTo evaluate the range and precision of nutrient parameter detection, including pH, nitrate, phosphate, potassium, calcium, sulfur, and magnesium. This involves analyzing concentration ranges, data reporting quality, and sensor sensitivity (Figs.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig21\" class=\"InternalRef\"\u003e21\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTo identify system limitations and propose improvement strategies, focusing on deployment challenges, standardization needs, underrepresented parameters, and integration with cloud platforms or decision-support tools (Figs.\u0026nbsp;\u003cspan refid=\"Fig24\" class=\"InternalRef\"\u003e24\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig26\" class=\"InternalRef\"\u003e26\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e1.5. Research contributions\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis review delivers a structured and in-depth examination of how machine learning techniques are implemented on microcontroller-class embedded platforms\u0026mdash;notably Arduino (58.97%), ESP32 (10.26%), and STM32\u0026mdash;for the sensing of key macronutrients: nitrogen, phosphorus, and potassium. It bridges academic insights with applied engineering outcomes to advance the precision agriculture domain. The core contributions of this review are:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eThe paper systematically compiles and analyzes studies that deploy lightweight ML models (TinyML) on resource-constrained devices. These efforts target nutrient detection in environmental waters and agricultural settings (Figs.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e, \u003cspan refid=\"Fig23\" class=\"InternalRef\"\u003e23\u003c/span\u003e, \u003cspan refid=\"Fig26\" class=\"InternalRef\"\u003e26\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThrough detailed classification of nutrient sensors (Fig.\u0026nbsp;\u003cspan refid=\"Fig22\" class=\"InternalRef\"\u003e22\u003c/span\u003e), water sources (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e), and concentration ranges (Figs.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig21\" class=\"InternalRef\"\u003e21\u003c/span\u003e), the review provides a technical framework for comparing sensor-model combinations, data acquisition strategies, and nutrient-specific performance.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe review identifies significant inconsistencies in parameter, along with gaps in real-time deployment, energy usage, and data reliability. Recommendations are offered for addressing latency, power constraints, and sensor calibration to enhance scalability and reliability in field conditions (Figs.\u0026nbsp;\u003cspan refid=\"Fig27\" class=\"InternalRef\"\u003e27\u003c/span\u003e, \u003cspan refid=\"Fig25\" class=\"InternalRef\"\u003e25\u003c/span\u003e, \u003cspan refid=\"Fig24\" class=\"InternalRef\"\u003e24\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e1.6. Research novelty\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe research review presents a detailed synthesis of studies about using machine learning\u003c/p\u003e \u003cp\u003eThis review distinguishes itself by being the first to present a multidimensional synthesis of how embedded machine learning systems are used for macronutrient sensing, with a particular focus on nitrogen, phosphorus, and potassium detection across low-power platforms. Unlike general sensor reviews, this study:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eLinks ML model deployment directly to hardware feasibility, highlighting how platforms like Arduino and ESP32 support low-power, real-time inference of trace nutrient levels\u0026mdash;e.g., nitrate at 0.03 ppm (93%, Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e) or phosphate below 0.05 ppm (72%, Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eReveals parameter-specific detection trends, showing that magnesium, sulfur, and potassium are vastly underrepresented (Figs.\u0026nbsp;\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003e, \u003cspan refid=\"Fig19\" class=\"InternalRef\"\u003e19\u003c/span\u003e, \u003cspan refid=\"Fig20\" class=\"InternalRef\"\u003e20\u003c/span\u003e), while calcium and pH have relatively balanced reporting (Figs.\u0026nbsp;\u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003e, \u003cspan refid=\"Fig21\" class=\"InternalRef\"\u003e21\u003c/span\u003e), underscoring areas for targeted innovation.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eConnects cloud architecture, wireless tech, and validation methods in a unified lens (Figs.\u0026nbsp;\u003cspan refid=\"Fig24\" class=\"InternalRef\"\u003e24\u003c/span\u003e\u0026ndash;\u003cspan refid=\"Fig27\" class=\"InternalRef\"\u003e27\u003c/span\u003e), enabling researchers to understand how IoT communication, ML evaluation metrics, and cloud integration strategies impact system effectiveness.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"2. Materials and Methods","content":"\u003cdiv class=\"Heading\"\u003e\u003cem\u003e2.\u003c/em\u003e Materials and Methods\u003c/div\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this subsection, the research looks at machine learning applications with microcontrollers that detect macronutrients like phosphate, nitrate, nitrogen, phosphorus, and potassium measured in ppm in water solutions. This was done through studying work that was published between 2015 and 2025. The assessment covers the work published between 2015 and 2025 to study important developments within this area. To our knowledge, no work that is the same as the one we are performing has been published within the timeframe, which makes this a unique study that will be a unique contribution to the field (Myataza et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The research methodology involves a well-carried out selection of relevant articles that have been peer reviewed and are taken from trustworthy online repositories known as Scopus, Google Scholar, and Web of Science, which guarantees a careful investigation of the subject being studied (Gumede et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Eligibility criteria\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eA systematic study that considered only peer-reviewed research about machine learning-based models which run on microcontroller platforms like ESP32, Arduino, etc, used to detect, calculate and classify macronutrients in water solutions was conducted for study. The criteria used to consider papers in the repositories used included only papers that were published in English between 2015 and 2025. A precise criterion for inclusion was used to ensure that only research that met the demands was focused upon, and that any research that did not meet any of the requirements was then filtered out. Additionally, only research that was peer-reviewed and had the basic elements of the topic being studies was exclusively considered (Mudau et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The inclusion and exclusion criteria for the study are found in Table \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e .\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eProposed Inclusion and Exclusion Criteria\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCriteria\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eInclusion\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eExclusion\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTopic\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eArticle papers focusing on Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eArticle papers not focusing on Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eResearch Framework\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe Articles must include research framework or methodology for Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eArticles must exclude research framework or methodology for Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLanguage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMust be written in English\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eArticles published in languages other than English\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePeriod\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eArticles between 2015 to 2025\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eArticles outside 2015 and 2025\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Information sources\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eA systematic search of online repositories was conducted to identify relevant studies for this review. The research platforms Scopus, Google Scholar, and Web of Science were used due to their broad coverage of peer-reviewed literature in the fields where machine learning is used with microcontrollers and sensors to investigate macronutrients in water. Each repository was carefully and comprehensively searched using a combination of keywords that covered the topic, guaranteeing that the most relevant research articles were captured. The first database used was Scopus, and it provided a large range of scientific journals and conference papers. The second database used was Google Scholar which enabled the inclusion of gray literature and academic dissertations that may not be found in any other platform. The third one that was used is Web of Science and it was used to cross-reference and ensure that the strength of the papers selected was good, as it provided citation data and impact factors of the journals. The results acquired from these databases formed the basis of the SLR, which guarantees a well-rounded and thorough collection of research papers (Mtjilibe, 2024).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Search strategy\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe literature for this systematic review was gathered from three major and reputable academic databases: Google Scholar, Scopus, and Web of Science. The search strategy focused specifically on terms relevant to machine learning integration, macronutrient sensing, and embedded systems operating in resource-constrained environments. To ensure comprehensiveness, search queries included a combination of technical and contextual keywords such as:\u003c/p\u003e \u003cp\u003e \u003cem\u003e(\"Machine Learning\" AND \"Macronutrient Sensing\" AND \"Microcontroller\" OR \"Embedded Systems\" AND \"Precision Agriculture\" OR \"Environmental Monitoring\" AND \"Low Power\" AND \"TinyML\").\u003c/em\u003e \u003c/p\u003e \u003cp\u003eThe search targeted studies published between 2015 and 2025, a time frame selected to reflect current trends and innovations in the field, particularly after the rise of TinyML and low-power IoT systems. A total of 2,546 initial records were retrieved, comprising 2,460 from Google Scholar, 49 from Scopus, and 37 from Web of Science (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). This structured and reproducible search approach ensures that the review encompasses a diverse, interdisciplinary, and global range of studies directly aligned with the objectives and research questions guiding this review.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eResults Achieved from Literature Search.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNo.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eOnline Repository\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNumber of Results Retrieved\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGoogle Scholar (GS)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e2,497\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eWeb of Science\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e37\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eScopus\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e49\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eTotal\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003e2,546\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"3\"\u003e\u003cb\u003eNote\u003c/b\u003e: A total of 2,507 records were excluded after initial screening. All 39 eligible reports were retrieved and assessed. No additional reports were excluded at the eligibility stage.\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.4. Selection process\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThree researchers individually reviewed the titles, abstracts, methodologies and results sections of the papers that each had retrieved from the individually assigned online database. Any differences in the selection of the papers that each member retrieved from their repository were discussed until a common point was reached. If the researchers could not agree on something even after discussion, the lecturer was consulted to help the members reach an agreement, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (Khanyi et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e2.5. Data collection process\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo guarantee that the information for the studies was exact, a structure was followed to reduce any errors and biases. The three reviewers then separately collected the data from each paper under the direction of the lecturer. Any differences in the extracted data were discussed until an agreement was reached. A data extraction process was used to ensure that consistency is kept across all three reviewers. For the data extraction process, no automation tools were used. Data was carefully recorded onto the table of Existing Studies \u0026ndash; Macronutrients, and each reviewer double-checked the other reviewer\u0026rsquo;s table to ensure that only accurate information was used. When the data in the studies was difficult to understand, a careful review of all available materials, including additional materials, was used to clarify the data. In situations where there were still concerns, the lecturer was consulted for his expertise in the task to ensure the reliability of the data used. When more than one report was found for the same study, the establishment of clear criteria was employed and used to select the most appropriate data, focusing most recent and broad studies published between 2015 and 2025 (Khanyi et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). In situations where the information from the reports was not consistent, methods and outcomes to resolve the differences were reviewed. Only studies written in English were included, excluding any articles in other languages to maintain consistency in our study and to avoid any likely misunderstandings due to language differences, as shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e2.6. Data items\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis section presents a detailed overview of the data elements required in this systematic review, focusing on both primary results and any additional elements relevant to Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing (Khanyi et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). The primary results cover various aspects such as operational efficiency in terms of sensor accuracy, model evaluation metrics, the connection type that was employed, the model of microcontroller that was used, the cloud used, and the software that was used to program the functionality of the system. This method allows for a quality analysis of how the different sensors perform when used in different conditions, water bodies, for the different macronutrients, pH measurement, temperature measurement, turbidity measurements, etc.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e2.6.1 Data Collection Method\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo ensure that a broad understanding of the topic was established on the topic of Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing efforts were made in the form of thoroughly studying, identifying and defining relevant data from trustworthy sources to capture the methods and processes used to conduct a high-quality investigation of the water aspects that were studied. The method was designed to produce strong evidence that reflects the effects of usage of different sensors, and microcontrollers in the systems that use machine learning to perform the task of water quality monitoring. The primary results of this systematic review were focused around several important areas that have a direct impact on Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing. Sensor accuracy was a major output, which was defined by measuring how precisely the sensor could pick up known concentrations of the macronutrients or any other aspects of the study, and any errors or inconsistencies in the measurement enables an understanding of how reliable, and hence accurate the sensor can be. The main concern for system performance received top priority. Research focused on accuracy rates and prediction speed and power level usage of the models. We evaluated the operational efficiency of these constrained devices since microcontrollers normally work with limited resources. An evaluation of the hardware infrastructure along with sensing components from each research study took place. The studies described both the selected microcontroller type (ESP32 or STM32) and its memory capacity together with its processing capabilities while explaining the detection methods for specific nutrients. The method of using models proved to be an important aspect of consideration. The research indicated whether models performed their predictions independently on the device or worked from an external point. Processing directly on the device provides essential functionality to portable systems and real-time applications which cannot always rely on cloud access. An assessment testing the environment\u0026rsquo;s practicality was done by checking the use case relevance. The testing included both clean and solution-based waters. Tests that used actual real-world applications received more importance because they provide more accurate assessment of how a system will function in its purpose environment. Researchers distinguished themselves through an ESP32 microcontroller which combined with a turbidity, pH and temperature sensor to investigate the relevant aspects (Khanyi et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e2.6.2 Definition of Collected Data Variables\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe study also considered additional measures to enhance understanding about machine learning implementations on microcontroller-class hardware for macronutrient detection. The selected variables enabled proper interpretation of results while understanding the extensive implications of embedding machine learning for nutrient analysis systems. We compiled detailed information about study properties which covered the research location and sample or product type alongside hardware specifications for examining deployment potential across various environmental conditions. These study characteristics allowed researchers to understand research results by showcasing the various methods researchers employed in this field. Recording implementation aspects included information about the machine learning models (decision trees, SVM, neural networks) alongside the sensor data features and microcontroller integration approaches and on-device or offloaded inference operations. Technical evaluation depended on this information to measure how deeply ML solutions worked with resource-limited platforms.\u003c/p\u003e \u003cp\u003eThe assessment also included economic elements which examined both hardware expenditures and sensor price ranges and energy utilization metrics since these elements demonstrate the practicality of implementing these systems in actual operational contexts. A complete analysis of embedded ML for macronutrient detection required an evaluation of external factors which included portable and low-cost sensing devices market demand as well as water quality testing regulatory standards along with technical skills constraints considered as well. Research was conducted carefully using our established methods which involved systematic database investigations in Google Scholar, Scopus, and Web of Science for selecting high-quality important studies all of which are shown in Table \u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e below. We used manual research methods for information acquisition to obtain precise relevant data which directed our analysis toward practical implementations and technological improvements within this domain. We guarantee a thorough assessment of machine learning impact on microcontroller-based macronutrient sensing through the identification and definition of the review's outcomes and variables. Our research methodology strengthens both the accuracy and importance of our results and provides vital information to practitioners and experts who work between embedded AI systems and water quality measurement studies (Khanyi et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eData Variables Collected.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eField\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eStudy characteristics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe study relies on geographic location and water source together with investigated water aspect as well as microcontroller platform in addition to factors that influence the study context.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eImplementation characteristics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eResearch provides details regarding ML models including decision trees and SVM as well as the sensor kinds (dielectric, optical and electrochemical) and microcontroller system integration steps.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHardware characteristics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe information includes specifications about the microcontroller processor type along with mentioned memory size and clock speed and statements about power requirements and whether analysis takes place locally or remotely.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEconomic factors\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eThe evaluation of financial elements involving hardware costs together with sensor investments and system deployment feasibility determines cost-effectiveness for real-world implementations.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eExternal influences\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTechnical skill limitations within regulatory standards of water quality testing combined with market needs and testing device requirements govern this system development.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e2.7. Study risk of bias assessment\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn the studies, the researchers needed to thoroughly evaluate potential bias risks within studies that used machine learning models on microcontroller-class devices for macronutrient detection to guarantee reliable and valid results. The evaluation process incorporated an adapted Newcastle-Ottawa Scale (NOS) as an evaluation instrument for technical and non-randomized experimental hardware trials and applied system evaluations. The NOS evaluated studies through three essential domains which included Selection for describing sample and data source selection and Comparability for controlling confounding variables and Outcome regarding performance metric measurement of accuracy, latency and power consumption reporting methods. Each evaluation got a rating based on star levels up to one for Selection and Up to two for Outcome and one to two stars in Comparability for determining the total methodological quality. The risk of bias evaluation took place through Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e which showed four separate reviewers performing their assessments independently to preserve objectivity. The reviewers discussed all disagreements until they reached agreement or when the fourth reviewer took the final decision (Khanyi et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWe conducted extra validation of research studies that lacked proper method descriptions or fell short in delivering necessary information regarding proprietary algorithms or undisclosed hardware specifications or unpublished data records. The supplementary data required for clarification of uncertainties was gathered through a cross-reference search of respected academic databases including Google Scholar and Scopus and Web of Science. Online repository manual searches served multiple purposes to gather all accessible data points and reduce the possibility of crucial information loss. The assessments conducted by hand stood as replacements for automated tools to deliver context-based and detailed evaluations across the entire process (Khanyi et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e2.8. Synthesis methods\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e below illustrates a flow chart that depicts the systematic method used in the review of Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing. The process began with the selection of studies where an identification and screening of papers based on an established eligibility criterion (Khanyi et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). After that, the data that was selected based on meeting the eligibility criteria was converted into the required format and was then cleaned to have a cohesive nature. Any missing information was then looked for and added by carefully studying the documents in detail, and in cases where that information was found still, supplementary materials were consulted. The next phase was the data analysis stage where the data was presented in the form of tables and graphs and then analyzed using those models. Then the heterogeneity assessment phase follows where the data\u0026rsquo;s flexibility is evaluated through subsection and sensitivity examination. The last phase of the flow chart is the bias assessment phase where any potential biases were identified, and transparency maintenance methods were applied. This organized method ensures a detailed and consistent review process.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this systematic review on Machine Learning on Microcontroller-Class Hardware for Macronutrient Sensing, we used careful combination methods to ensure that our results were strong, clear, and reproducible. To determine the qualification of papers for studying, a table was created to insert the characteristics of each paper and used to compare them against the already set production groups. The method included only relevant papers that helped ensure both legitimacy and appropriateness according to reviewer targets (Khanyi et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). The researches handled missing summary statistics by suggesting alternative values and applied necessary data conversions to maintain study consistency. The findings were then reported using a combination of organised tables and forest plots, which provided a clear illustration of the effect estimations and certainty periods, allowing for the identification arrangements and outlier efficiently. The production of findings was handled using a random-effects meta-analysis model, with subgroup analyses clearly directed to geographic and economic contexts to understand their impact on Microcontroller-Class Hardware for Macronutrient Sensing (Khanyi et al., \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis method provided nuanced awareness into how these factors cooperate with the acceptance and success of Microcontroller-Class Hardware for Macronutrient Sensing, which was further explored through subgroup analyses and meta-regressions. These analyses helped us identify potential sources of heterogeneity, such as nature and size of water source and the type of microcontroller used (Ngcobo et al., 2024). This helped in refining our understanding of the effects of these technologies. Also, sensitivity analyses were performed to weigh the strength of the produced findings, confirming that our assumptions were well-supported by balanced and consistent evidence. Through this broad methodology, we were able to provide a meaningful combination of the confirmation, suggesting constructive perceptions for participants interested in leveraging Microcontroller-Class Hardware for Macronutrient Sensing (Myataza et al, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec18\" class=\"Section3\"\u003e \u003ch2\u003e2.8.1. Eligibility for Synthesis\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo determine study eligibility for inclusion in our systematic review on Microcontroller-Class Hardware for Macronutrient Sensing, we analysed every study to establish that it had enough details and connection to our synthesis before final admission. The evaluation of each study matched its characteristics to our synthesis requirements by focusing on machine learning models as well as nutrient targets and sensors along with hardware and performance metrics. The research team created a standardized mapping system which served to assess each study according to its established criteria. Research that passed the technical requirements without sufficient details about field testing and operational deployment received only qualitative attention. The established process verified that the incorporated studies satisfied both methodological standards and research aims at hand (Myataza et al, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section3\"\u003e \u003ch2\u003e2.8.2. Data Preparation for Synthesis\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this review, the systems used included changing or making data collected from various studies standard to ensure reliability before production. For instance, when impact magnitudes were described in a different way throughout the findings, numerical influences were used to translate these into a standard magnitude, such as translating odds ratios to risk ratios where applicable (Myataza et al, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Also, managing missing data was an important feature of the study. Missing summary information, such as standard deviations or impact magnitudes, were implicated using determined statistical techniques like multiple attribution. This method ensured that the dataset was thorough and strong, permitting a more precise and consistent investigation.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e2.8.3. Tabulation and Visual Display of Results\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eOutcomes from separate findings and combination works were sorted out using both tabular and graphical techniques to improve transparency and help in judgement. Tabular arrangements were used to organize the information in an organized arrangement, where results were ordered by field, and within each field, findings were well-organized from smallest to greatest risk of bias (Myataza et al, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). This grouping permitted simple contrast across findings and emphasized the most dependable proof. Additionally, graphical techniques, particularly forest plots, were used as the primary instrument for visually presenting meta-analysis results. These plots showcased impact estimations and certainty intervals for each study beside a summary approximate. The findings in the forest plots were arranged based on impact size or year of publication, facilitating revealing movements over time and across different research focuses.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section3\"\u003e \u003ch2\u003e2.8.4. Synthesis of Results\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn the process of manually searching on online sources such as Google Scholar, Scopus, and Web of Science, we thoroughly assessed and processed the findings of important findings (Myataza et al, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The method to information production was directed by the type of information and the level of flexibility seen across studies. Based on the outcomes from the search, a manual assessment of the applicability of both fixed-effects and random-effects models, depending on the level of heterogeneity among study results. The choice of the model was controlled by the properties of the information and our expectations about the dependability of impacts across studies. After transferring the data to Excel, charts were created to visually study the information, permitting us to recognize patterns of flexibility and prospective heterogeneity across the studies. This first visual examination provided an overview of how study outcomes varied from one another, enabling a more nuanced analysis.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section3\"\u003e \u003ch2\u003e2.8.5. Exploring Causes of Heterogeneity\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eSubsection considerations and meta-regression were performed to survey likely sources of heterogeneity, such as variations in study backgrounds, mediation types, or result sizes. Precise studies fixated aspects like the size and type of the water body being investigated, the type of microcontroller tool used, and the geographic location, all of which were inspected to measure their effect on the usefulness in Microcontroller-Class Hardware for Macronutrient Sensing. These techniques assisted in the identification of underlying patterns and connections that contributed to the overall flexibility detected across the findings (Myataza et al, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003e2.8.6. Sensitivity Analyses\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eSensitivity analyses were used to calculate the strength of the results produced with respect to different expectations and procedural judgments made during the review process. These analyses included testing the effect of omitting studies at high risk of bias and using different statistical models to warrant that the conclusions were not improperly influenced by studies or analytical methods. This method facilitated the confirmation of the dependability and legitimacy of the results by addressing possible sources of bias and ensuring that the results were consistent across different analytical settings (Myataza et al, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003e2.9. Reporting bias assessment\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eWe performed our systematic review regarding machine learning implementation on microcontroller-class hardware for macronutrient sensing by identifying possible risks that emerged from reporting biases such as selective publishing along with selective reporting of results. The potential accuracy and trustworthiness problems of our analysis because of these biases led us to develop a systematic methodology for its treatment. The assessment of reporting bias included established statistical and visual evaluation techniques. When evaluating our data we selected contour-enhanced funnel plots because they proved helpful for detecting imbalances in our information. Evaluation of these plots assisted researchers in identifying publication bias through thorough examination of areas that could contain biased studies rather than random omissions (Myataza et al, \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The addition of statistical significance lines allowed us to distinguish different sets of data more easily which revealed potential bias factors to our evaluation.\u003c/p\u003e \u003cp\u003eThe assessment utilized established tools which researchers describe widely throughout their publications. The reliability factors of these tools formed a critical foundation for our assessment logic. Options for contour-enhanced funnel visualization enabled a basic visualization of study distribution which let us detect and address possible biases during the review phase. Our assessment was designed to minimize human interpretation which ensured the study results remained impartial and truthful. A group of independent reviewers examined the studies and solved all discrepancies through collective discussion. Assistance from a method expert became necessary whenever reviewers failed to find mutual consensus about a specific issue. Additional steps were included to investigate studies with limited information disclosure particularly when they used proprietary machine learning models together with special hardware components. The researchers used Google Scholar together with Scopus and Web of Science to verify unclear information from sources. Additional online database query procedures were performed to eliminate biases while strengthening the accuracy and completeness of our assessment. Our team used only manual methods for completing this work. Our methodology involved using Excel together with manual methods because automated systems were not employed. The manual data handling approach helped us thoroughly review information to detect subtle data patterns while confirming the absence of hidden prejudices.\u003c/p\u003e \u003cp\u003eDetailed manual investigations were performed on Google Scholar and Scopus and Web of Science platforms to validate our research findings. Our research enabled us to study data from various studies and sources that strengthened the accuracy of our research findings. To build a reliable review foundation we performed hand-driven database searches for obtaining the most detailed correct information. We altered traditional methods to evaluate reporting bias because of the specialized requirements in research about machine learning implementation in embedded hardware for nutrient sensing. The reporting styles found in embedded systems differ from medical and social science research, so we adapted our procedures because this created a more accurate and relevant methodological approach. Our techniques matched the examined studies to maintain both strong methodology and appropriate analysis for the research topic. We documented our bias assessment procedures thoroughly and made all our methods accessible in the supplementary materials to provide both verification and possible re-use for our work. Other researchers can utilize this method transparency to reproduce our process or develop it for use in future studies which will promote better research quality regarding machine learning applications for microcontroller-based macronutrient sensing systems.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e2.10. Certainty assessment\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe reviewed literature was evaluated based on five quality assessment (QA) criteria to ensure rigor and relevance:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eQA1: The clarity and explicitness of the research aim.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eQA2: The specification and transparency of data collection methods.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eQA3: The clear definition and explanation of the Microcontroller-Class Hardware for Macronutrient Sensing processes.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eQA4: The application of a well-defined and appropriate research methodology.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eQA5: The contribution of the research findings to the enhancement of existing literature on projects\u0026rsquo; performance.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe certainty assessment responses are rated on a scale from zero (0) to one (1). A 'No' response is assigned '0' points, a score of '0.5' is given if the criterion is 'Partially' met, and '1' point is assigned for a 'Yes' response. All five criteria are scored using this scale. Each piece of literature under review can receive a total score between 0 and 5 points. The results of the certainty assessment for the collected literature on the applications and competitive advantages of Microcontroller-Class Hardware for Macronutrient Sensing performance are presented in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCertainty Assessment Results for Collected Literature on Microcontroller-Class Hardware for Macronutrient Sensing.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRef.\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eQA1\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eQA2\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eQA3\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eQA4\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eQA5\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003e% grading\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Akhter et al., 2021), (Akhter et al., 2021), (Alahi et al., 2018), (Alahi et al., 2017), (Rmadhan, 2020)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e50\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Martinez et al., 2020), ), (Alahi et al., 2018), ), (Alahi et al., 2018), (Tan et al., 2022)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e60\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Raju et al., 2017), (Bluett et al., 2023), (Sekhwela et al., 2023), (Akhter et al., 2021), (Qamruzzaman, 2025), (Campelo et al., 2022), (Ban rt al., 2020), (Zin et al., 2019), (Akhter et al., 2021)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e3.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e70\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Sholihaha et al., 2022), (Richa et al., 2021), (Lehto et al., 2023), (Rahju et al., 2017), (Yu et al., 2021), (Xiong et al., 2023), (Miller et al., 2025), (Lowe et al., 2022), (Akhter et al., 2021), (Rahju et al., 2017), (Abdikadir et al., 2024), (Alahi et al., 2018), (Wang et al., 2022), (Yuan et al., 2018), [34], (Abhisheesh et al., 2021), (Adu-Manu et al., 2020), (), (Afrid et al., 2023), (Zin et al., 2019)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Akhter et al., 2021), (Akhter et al., 2021), (Alahi et al., 2018), (Alahi et al., 2017), (Rmadhan, 2020)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e4.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e(Martinez et al., 2020), (Alahi et al., 2018), (Alahi et al., 2018), (Tan et al., 2022)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e2.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e50\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis systematic review needed an assessment of evidence certainty to validate its results about machine learning performance using microcontrollers in macronutrient measurements. We used the GRADE (Grading of Recommendations Assessment Development and Evaluations) framework for a systematic assessment which determined the reliability of our research findings. The evidence quality assessment system GRADE functions as a worldwide recognized approach that provides complete transparent assessment protocols to build trust in research findings which support valid and credible conclusions. Multiple critical factors allowed us to analyse the evidence certainty for main outcomes with detail. Our assessment initially focused on understanding the performance metric effectiveness by studying both sample size data and confidence interval width measurements in the published reports. The combination of narrow confidence intervals with big sample sizes provided enhanced certainty because it delivers precise and reliable estimation of model accuracy and inference time and power consumption measurements. The research design included analyses to evaluate the consistency across different studies. Research results with high consistency between studies enhanced the overall confidence. The researchers deeply studied all observed discrepancies to determine their origins alongside their potential effects on the study results.\u003c/p\u003e \u003cp\u003eThe evaluation of study bias risk used an adapted implementation of the Cochrane Risk of Bias tool. Research with minimal bias risks brought greater value to the opinion strength of the supporting data. The assessment of directness involved evaluating whether the populations under study along with their interventions and outcomes matched the core questions of this review. High levels of directness reinforced the findings in our conclusions which in turn increased the trustworthiness of the evidence. Using these evaluation factors the evidence's certainty was classified as High when experiments showed consistency, precision and clear applicability and low risk of bias. Moderate certainty ratings were assigned to studies when researchers identified small issues with one factor between consistency and moderate levels of bias. The rating of Low certainty depended on existing major issues in multiple fields such as measurement inaccuracy and inconsistent results and high potential for biased outcomes. Very low certainty was applied because critical issues appeared throughout all factors leading to a substantial decrease in confidence in the results obtained. We modified the GRADE approach to match the requirements of this review by tailoring it to analyse performance attributes and device practicality of machine learning for macronutrient sensing on embedded hardware. Separate researchers who did not participate in other review stages evaluated the degree of evidence reliability for each result. All reviewers reached consensus during discussions to confirm a balanced assessment and thus resolve any discrepancies. Our certainty evaluations received additional data and clarification from the study authors whenever possible.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Study selection\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo include only quality and pertinent studies, we selected studies by following a structured and thorough system when reviewing machine learning research on microcontroller-class microprocessors for macronutrient sensing. Google Scholar, Web of Science and Scopus were all carefully investigated to uncover studies that matched the set rules for selecting research to be reviewed. Altogether, 2546 initial records were found after searching on Google Scholar, Web of Science and Scopus. A careful screening and evaluation process led to the addition of 41 studies to the final review. The review process is represented by the proof of workflow in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e using a PRISMA flow diagram.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Study Results\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003e illustrates the yearly frequency of studies focused on machine learning applications for macronutrient sensing on microcontroller-class hardware between 2017 and 2025. The highest research activity was recorded in 2021, with 10 papers, followed by 2022 with 8 papers. The years 2020 and 2023 also showed moderate contributions with 5 and 5 papers respectively, while early years like 2017 and 2019 had fewer publications (3 and 2 respectively). The drop in publications post-2023 may reflect a shift toward applied development and deployment rather than exploratory studies.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e shows the distribution of publication types included in the review. Journal articles dominate the landscape, accounting for 82% of the total studies. Conference papers follow at 15%, while theses represent only 3%. The high share of journal papers reflects strong academic engagement and peer-reviewed validation in this emerging field, while the inclusion of theses and conference papers points to ongoing academic exploration and innovation.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003e shows the proportional contribution of each literature repository to the final set of included studies. Google Scholar was the most dominant source, accounting for 74% of the papers, followed by Web of Science at 18% and Scopus at only 8%. This distribution reflects broader accessibility and indexing scope of Google Scholar compared to the more selective curation of Scopus and Web of Science, indicating a preference for inclusive search strategies in emerging or cross-disciplinary topics like embedded ML in nutrient sensing.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003e maps the flow of reviewed studies from their academic disciplines (left) to specific journals or publication venues (right). The majority of contributions stem from Environmental \u0026amp; Water Sciences, Computer Science \u0026amp; ICT, and Sensor Technologies, each bridging into high-impact journals such as Smart Urban Water Networks (10.26%), Water (7.69%), and Environmental Chemistry Letters (5.13%). A significant share of studies also spans Engineering and Agricultural \u0026amp; Biosciences, showing the multidisciplinary nature of research in microcontroller-based macronutrient sensing. This visualization emphasizes that despite a shared application domain, contributions are spread across a diverse range of technical and environmental journals, reflecting the breadth and complexity of this emerging research field.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e maps the origin of studies by region (left) and corresponding country (right), highlighting strong representation from Asia, Africa, and Europe. Australia stands out as the single most prolific contributor, responsible for 33% of the included papers. Other notable contributors include Finland (10%), China (5%), Hong Kong (8%), and India, reflecting a global spread in interest toward embedded ML-based nutrient sensing. The concentration of studies in developing nations suggests heightened relevance for low-cost, real-time sensing solutions in resource-constrained agricultural settings.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e illustrates how different water sources (left) are classified into application categories (right) for microcontroller-based macronutrient sensing studies. Subsurface water sources (e.g., groundwater and soil water) lead the focus, representing 31% of the reviewed work. Soil \u0026amp; Agricultural Water (18%) and Potable Water (15%) also stand out, suggesting their strong relevance to real-world deployment. Notably, Controlled Agricultural Water (8%) and Agricultural/Livestock Use (7%) reflect a rising interest in closed-loop nutrient systems. Categories such as Storage/Infrastructure and Specialized/Treated Water remain underexplored (\u0026lt;\u0026thinsp;5%), highlighting future opportunities for expanding sensing research into water management infrastructure.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e shows how specific water quality parameters (left) are grouped into broader categorical research themes (right). The most common parameters assessed include pH, nitrate, temperature, and phosphate\u0026mdash;with multiple studies examining their combinations. The leading thematic category is Comprehensive Water Quality, comprising the highest share of studies due to multi-parameter monitoring (e.g., simultaneous detection of nitrogen, turbidity, and pH). Other notable focuses include Nutrients/Agriculture (13%), Nitrogen Compounds, and Ion Concentrations (3%). The wide variety of parameters shows the field\u0026rsquo;s interdisciplinary scope, though some categories like Salinity and Turbidity remain underrepresented\u0026mdash;suggesting opportunities for deeper sensor integration in environmental studies.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e categorizes the phosphate concentration ranges (ppm) measured in the reviewed studies, offering insight into calibration, sensitivity, and contextual application. The majority of studies (72%) reported measurements in a very low range (e.g., 0.005\u0026ndash;0.05 ppm or 0.01\u0026ndash;40 ppm), aligning with realistic phosphate levels in natural and agricultural water systems. A small fraction (5%) documented anomalous or highly specific values, while others (3%) focused on low specific values or broader wide-range detection capabilities. Notably, a considerable number of studies left phosphate range unspecified, indicating inconsistency in reporting\u0026mdash;a limitation that may affect comparability and reproducibility across implementations.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e12\u003c/span\u003e illustrates the wide variance in nitrate concentration ranges (ppm) used across studies and how they align with broader categorical classifications. The most frequent grouping was the \u0026ldquo;broad, typical environmental range\u0026rdquo; (26%), which suggests that many sensing systems target naturally occurring nitrate levels in soil and water. A considerable number of studies explored high, very high, or even extremely high specific values, signaling interest in polluted or nutrient-intensive environments. Only 3% of studies focused on the moderate range, while 5% targeted fixed specific values (e.g., 10 ppm). The \u0026ldquo;Not Specified\u0026rdquo; category still featured in several studies, emphasizing again the need for clearer parameter reporting in nutrient sensing literature.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e15\u003c/span\u003e displays the use of specific nitrate concentration values in parts per million (ppm) and their classification. A substantial 93% of studies that used fixed nitrate values targeted very low (0.03 ppm) or low (3.1 ppm) concentrations, consistent with levels commonly found in environmental water or minimally fertilized sources. Only 3% of studies used moderate values (e.g., 10 ppm), while a notable proportion remained unspecified, highlighting ongoing issues with standardized reporting in this domain. The figure underscores that most sensing systems are optimized for detecting trace levels of nitrate, suggesting alignment with early warning or low-intensity monitoring applications.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e16\u003c/span\u003e highlights the phosphorus values (ppm) reported in reviewed studies, mapped into categorical ranges. The vast majority of papers (92%) used a low fixed value (0.7 ppm) or very low range (0.06\u0026ndash;0.074 ppm), indicating a trend toward detecting trace phosphorus levels in agricultural or environmental waters. Only 3% addressed high concentration ranges (e.g., 25\u0026ndash;50 ppm), typically seen in nutrient-rich or polluted systems. Several studies failed to specify concentration values altogether, reinforcing a broader pattern of incomplete parameter reporting in nutrient sensing literature. These trends suggest a dominant focus on early-stage or minimal phosphorus presence, which may limit applicability in high-impact scenarios like runoff or contamination hotspots.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e17\u003c/span\u003e presents the potassium concentration values (ppm) used in studies and how they were categorized. A significant majority (85%) of the studies reported values within broad or high ranges (e.g., 30\u0026ndash;400 ppm, 98.84\u0026ndash;156.4 ppm), suggesting an emphasis on scenarios involving nutrient-rich or fertilized environments. Only 5% targeted low concentration ranges, and 3% documented moderate values, while several studies remained unspecified. This distribution reflects the complexity of potassium sensing and the limited availability of sensors capable of detecting it accurately at lower levels. The result supports earlier findings that potassium remains one of the least explored macronutrients in embedded ML sensing, despite its agronomic importance.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e18\u003c/span\u003e outlines the calcium values (ppm) reported in the reviewed studies, linking them to categorical classifications. The majority of research (74%) focused on moderate specific values (e.g., 10, 20, 40.5 ppm), indicating an emphasis on typical calcium concentrations found in natural and irrigation waters. A range of broader classifications\u0026mdash;low, broad, moderate to broad, and even very high\u0026mdash;are also represented, reflecting calcium\u0026rsquo;s widespread presence across environmental conditions. Only 3% of studies failed to specify values. This pattern reveals a more balanced approach to calcium sensing compared to nitrogen or potassium, perhaps due to better sensor availability and established roles of calcium in both soil chemistry and water hardness monitoring.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig19\" class=\"InternalRef\"\u003e19\u003c/span\u003e shows the magnesium values (ppm) observed in the reviewed literature and their classification. An overwhelming majority (95%) of studies that reported magnesium targeted low specific values (e.g., 15.2 ppm), likely aligning with baseline environmental concentrations in surface and groundwater. Only 3% documented very high ranges (e.g., 1150\u0026ndash;1350 ppm), while some papers did not specify any concentration data. The stark underrepresentation of broad or moderate ranges suggests that magnesium sensing remains a niche focus within embedded ML systems\u0026mdash;often added as a supplementary parameter rather than a core nutrient target like nitrogen or phosphate.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig20\" class=\"InternalRef\"\u003e20\u003c/span\u003e presents the sulfur concentration values (ppm) across studies and their classification. The overwhelming majority (95%) of reviewed studies that reported sulfur levels focused on a specific moderate value (e.g., 8.4 ppm), highlighting targeted sensing rather than broad-range detection. Only 3% addressed broad concentration ranges (e.g., 0.5\u0026ndash;50 ppm), while a notable number of studies left the concentration unspecified. The pattern mirrors that of magnesium \u0026mdash; sulfur is underreported and underexplored in embedded ML literature, possibly due to fewer available sensors and lower prioritization in precision agriculture systems.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig21\" class=\"InternalRef\"\u003e21\u003c/span\u003e maps pH values and ranges to their corresponding classification in the reviewed studies. A significant proportion (33.85%) of entries reported very broad pH ranges, indicating variability in sensing environments or flexible system thresholds. Specific classifications included neutral values, acidic to neutral, alkaline, and a small fraction (10.26%) marked as slightly acidic. Notably, some data points (e.g., \"45700\") were clearly erroneous or invalid, highlighting quality control issues in reporting. This diversity in classification confirms that pH is a widely integrated parameter in macronutrient sensing setups, but also suggests the need for more standardized reporting and clearer calibration documentation.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig22\" class=\"InternalRef\"\u003e22\u003c/span\u003e classifies various nutrient and environmental sensor types reported in the reviewed literature, connecting them to broader sensor technology categories. Water Quality Sensors were the most common (17.95%), followed by Nanomaterial-Based Sensors (12.82%) and Mixed Sensor Technology (10.27%). Frequently mentioned components included ISEs (Ion-Selective Electrodes) for nutrients like NO₃⁻ and NH₄⁺, capacitive sensors, and lab-on-chip solutions. Despite some classifications being marked as \"unspecified,\" the data demonstrates a broad spectrum of innovations, ranging from custom electrochemical configurations to IoT-integrated optical sensors, reflecting the growing multidisciplinary nature of embedded macronutrient sensing systems.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig23\" class=\"InternalRef\"\u003e23\u003c/span\u003e illustrates the distribution of embedded platforms, hardware tools, and computing categories used in the reviewed studies. Arduino-based systems dominated the landscape (58.97%), commonly associated with microcontrollers (66.67%). Other significant mentions include ESP32/ESP8266 (10.26%) and Raspberry Pi (used alone or alongside Arduino/STM32), feeding into both single-board computers and mixed embedded platforms. A smaller subset utilized FPGA and Autodesk, aligning with hardware accelerators (15.38%) and software tools (7.69%), suggesting performance-oriented or simulation-based research. This breakdown emphasizes Arduino\u0026rsquo;s accessibility and flexibility as a key driver for widespread experimentation and deployment in low-resource agricultural and environmental contexts.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig24\" class=\"InternalRef\"\u003e24\u003c/span\u003e presents the spectrum of wireless technologies used in embedded macronutrient sensing platforms. Bluetooth was the most prevalent technology, appearing in 64.1% of reviewed studies, and closely associated with local wireless networks, underlining its relevance in short-range, low-power scenarios such as on-farm deployments. Wi-Fi (7.69%) and Zigbee (15.38%) were also common, generally categorized under short-range wireless. LPWAN technologies like LoRa (2.56%) and Sigfox, and cellular IoT solutions such as GSM and LTE-M/NB-IoT, supported long-range, low-bandwidth use cases and made up 20.5% of the records. This communication layer diversity reflects the balance between low-latency local networks and scalable remote access in nutrient sensing architectures.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig25\" class=\"InternalRef\"\u003e25\u003c/span\u003e illustrates the distribution of cloud platforms and their associated service models in reviewed macronutrient sensing systems. Amazon Web Services (33.33%) led the commercial cloud space, followed by ThingSpeak (28.21%) and Blynk 2.0 (23.08%), all pivotal in IoT cloud services or generic cloud hosting applications. FIWARE and custom-built solutions accounted for 35.89% of the implementations, reflecting a strong interest in custom or open-source platforms for flexibility and control. This breakdown highlights the balancing act between vendor-managed convenience and project-specific adaptability, essential for embedded sensing in varied environments.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig26\" class=\"InternalRef\"\u003e26\u003c/span\u003e outlines the tools and programming environments used in the development of nutrient sensing platforms. The Arduino IDE overwhelmingly dominates with 58.97% of usage, reaffirming its strong presence in microcontroller development. Tools like MATLAB, PyCharm, Linux, and Delft3D show a scattered but meaningful contribution across modeling, scientific computing, and OS environments. AI and ML tools appeared in a limited number of papers (just over 5%), suggesting emerging but still underrepresented integration. The relatively lower use of dedicated IoT platforms or data visualization tools reflects a need for broader incorporation of integrated decision-making systems in sensor deployment.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig27\" class=\"InternalRef\"\u003e27\u003c/span\u003e illustrates the distribution of validation and evaluation metrics across the reviewed studies. The R\u0026sup2; (Coefficient of Determination) is the most frequently used metric (48.72%), dominating the Goodness-of-Fit category and underscoring a strong focus on statistical model reliability. Accuracy (20.51%) and error-based methods like RMSE, MAE, and MAPE collectively form a major part of performance, consistency, and error metrics, showcasing the methodological diversity in assessing machine learning systems. Machine learning validations, used in 10.26% of cases, include algorithms such as KNN, RF, DT, and PCA, but are often embedded within broader validation frameworks rather than standalone metrics. The use of correlation analysis (20.51%) further confirms the emphasis on statistical alignment and predictability across data inputs and model outputs. However, 2.56% of papers lacked a clear metric declaration, pointing to a need for improved transparency in reporting.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Responses to Research Questions\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003e \u003cb\u003eRQ1: What types of hardware platforms and sensors are commonly used in macronutrient sensing systems that incorporate embedded machine learning?\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe review found that Arduino-based microcontrollers were dominant, used in 58.97% of studies (Fig.\u0026nbsp;\u003cspan refid=\"Fig23\" class=\"InternalRef\"\u003e23\u003c/span\u003e). ESP32 and ESP8266 platforms followed at 10.26%, often used in wireless setups. These platforms were mostly paired with microcontrollers (66.67%, Fig.\u0026nbsp;\u003cspan refid=\"Fig23\" class=\"InternalRef\"\u003e23\u003c/span\u003e) and supported by Bluetooth (64.1%, Fig.\u0026nbsp;\u003cspan refid=\"Fig24\" class=\"InternalRef\"\u003e24\u003c/span\u003e) for low-power, short-range data transmission. Regarding sensors, water quality sensors accounted for 17.95% of the documented tools, followed by nanomaterial-based sensors (12.82%) and mixed sensor technologies (10.27%) (Fig.\u0026nbsp;\u003cspan refid=\"Fig22\" class=\"InternalRef\"\u003e22\u003c/span\u003e). Ion-selective electrodes (ISEs) were frequently used for detecting nitrate (NO₃⁻), ammonium (NH₄⁺), and potassium (K⁺), especially in portable nutrient probes.\u003c/p\u003e \u003cp\u003e \u003cb\u003eRQ2: How are lightweight machine learning models, including TinyML and related optimizations, adapted to operate efficiently within the limited resources of microcontroller-class devices?\u003c/b\u003e \u003c/p\u003e \u003cp\u003eMost studies adopted tiny, hardware-compatible IDEs, especially the Arduino IDE (58.97%), reflecting support for TinyML workflows (Fig.\u0026nbsp;\u003cspan refid=\"Fig26\" class=\"InternalRef\"\u003e26\u003c/span\u003e). While only just over 5% of tools explicitly referenced AI/ML libraries or decision support models, the broad use of microcontrollers (66.67%) and local processing platforms indicates an implicit adaptation of lightweight ML models to constrained environments. Moreover, Bluetooth and Wi-Fi, prevalent in 64.1% and 7.69% of studies respectively (Fig.\u0026nbsp;\u003cspan refid=\"Fig24\" class=\"InternalRef\"\u003e24\u003c/span\u003e), suggest designs favouring edge computing with minimal cloud reliance. Cloud integration still played a role\u0026mdash;33.33% used AWS, and 28.21% used ThingSpeak (Fig.\u0026nbsp;\u003cspan refid=\"Fig25\" class=\"InternalRef\"\u003e25\u003c/span\u003e), but these were often for logging or analysis, not computation, confirming an edge-leaning approach.\u003c/p\u003e \u003cp\u003e \u003cb\u003eRQ3: To what extent do current studies report system-level performance indicators such as energy consumption, latency, processing speed, and responsiveness in live sensing environments?\u003c/b\u003e \u003c/p\u003e \u003cp\u003eWhile 15.38% of studies used hardware accelerators (e.g., FPGA, Autodesk) to improve performance (Fig.\u0026nbsp;\u003cspan refid=\"Fig23\" class=\"InternalRef\"\u003e23\u003c/span\u003e), the reporting of runtime metrics such as energy and latency was sparse and inconsistent. Only 10.26% of papers applied machine learning validation methods such as KNN, RF, or PCA (Fig.\u0026nbsp;\u003cspan refid=\"Fig27\" class=\"InternalRef\"\u003e27\u003c/span\u003e), and just 2.56% clearly specified statistical or error metrics for system evaluation. Figure\u0026nbsp;\u003cspan refid=\"Fig27\" class=\"InternalRef\"\u003e27\u003c/span\u003e shows a heavy focus on R\u0026sup2; (48.72%) and accuracy (20.51%)\u0026mdash;indicating model-level validation\u0026mdash;but system-level benchmarks like power draw, memory usage, or time-to-response remain underreported. This limits broader performance comparisons.\u003c/p\u003e \u003cp\u003e \u003cb\u003eRQ4: What unresolved limitations exist in the literature related to sensor coverage, parameter reporting, model generalizability, and system scalability, and what directions are proposed for improvement?\u003c/b\u003e \u003c/p\u003e \u003cp\u003eSignificant inconsistencies were noted in parameter specification. For instance:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003ePhosphate concentration was left unspecified in several cases, with 72% of studies focusing only on very low ranges (Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e13\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eNitrate was better represented, with 26% of studies covering broad ranges, but still several entries lacked units or used outlier values (e.g., \u0026ldquo;223.2\u0026ndash;372 ppm,\u0026rdquo; Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e14\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eFor magnesium and sulfur, 95% of studies used just one fixed value or gave no clear range (Figs.\u0026nbsp;\u003cspan refid=\"Fig19\" class=\"InternalRef\"\u003e19\u003c/span\u003e and \u003cspan refid=\"Fig20\" class=\"InternalRef\"\u003e20\u003c/span\u003e), showing underrepresentation and poor coverage.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003epH reporting was more thorough but still contained erroneous values (e.g., \"45700\"), pointing to data quality issues (Fig.\u0026nbsp;\u003cspan refid=\"Fig21\" class=\"InternalRef\"\u003e21\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFurthermore, underexplored water categories like storage infrastructure (3%) and treated water (3%) (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e11\u003c/span\u003e), along with geographic concentration in Australia (33%) (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e10\u003c/span\u003e), suggest limited scalability and generalizability across different environments. Thus, improvements are needed in:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eStandardizing parameter ranges\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eExtending sensor coverage to neglected macronutrients (Mg, S)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eIncorporating mixed validation methods\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eImproving documentation of deployment contexts\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Conclusion","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis review provides a comprehensive synthesis of recent advancements in machine learning\u0026ndash;enabled macronutrient sensing using microcontroller-class hardware, highlighting both the innovation and the fragmentation characterizing this emerging field. The evidence strongly suggests that embedded sensing systems are increasingly viable for agricultural and environmental monitoring, with Arduino-based platforms (58.97%) and Bluetooth communication (64.1%) dominating the hardware landscape. Sensor technologies remain diverse, with water quality sensors (17.95%), nanomaterial-based electrodes (12.82%), and ISEs for key nutrients reflecting a multidisciplinary push toward high-precision, low-cost systems.\u003c/p\u003e \u003cp\u003eNutrient detection focused heavily on trace-level measurements\u0026mdash;particularly for phosphates (72%), nitrates (93%), and phosphorous (92%)\u0026mdash;underscoring the prioritization of early-stage nutrient monitoring over high-intensity applications. However, parameters like magnesium and sulfur were either underreported or inconsistently presented, revealing a critical gap in comprehensive macronutrient profiling. Additionally, pH monitoring was widely used but plagued by data inconsistencies, reaffirming the need for standardized reporting across studies.\u003c/p\u003e \u003cp\u003eFrom a systems perspective, most platforms leveraged TinyML-compatible environments such as the Arduino IDE (58.97%) with minimal reliance on cloud-based computation. Where cloud integration occurred, AWS (33.33%) and ThingSpeak (28.21%) were prominent, primarily serving data management rather than real-time inference. While performance metrics like R\u0026sup2; (48.72%) and accuracy (20.51%) were common, energy consumption, latency, and computational efficiency remained largely underexplored\u0026mdash;marking a significant blind spot for real-world deployment.\u003c/p\u003e \u003cp\u003eGeographically, the review revealed a dominant contribution from Australia (33%), with notable input from Asia and Africa, indicating a global yet uneven research spread. Thematic trends show strong alignment with subsurface water monitoring (31%) and agriculture-focused use cases, but limited application in treated or stored water systems. Ultimately, while the field has made measurable strides in sensing fidelity, platform design, and localized intelligence, future research must prioritize scalability, deployment validation, and interdisciplinary integration\u0026mdash;particularly in underrepresented regions and nutrient parameters. Broader standardization and benchmarking will be essential to transition these technologies from promising prototypes to resilient, field-ready tools.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eAbdel-Basset M, Hawash H, Abdel-Fatah L (2024) Artificial Intelligence and Internet of Things in Smart Farming. CRC\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAhmed R, Green M (2023) Pasture yield prediction using ML-based optical sensing: A systematic review. J Agricultural Technol 17(2):112\u0026ndash;126. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.jagt.2023.105217\u003c/span\u003e\u003cspan address=\"10.1016/j.jagt.2023.105217\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAlmeida J, Torres F, Bianchi R (2021) Machine learning in greenhouse climate control: A case study. Comput Electron Agric 190:106421. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compag.2021.106421\u003c/span\u003e\u003cspan address=\"10.1016/j.compag.2021.106421\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnon (2018) Machine learning in crop quality assessment. Int J Agricultural Sci 45(3):201\u0026ndash;209. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/ijag.2018.00453\u003c/span\u003e\u003cspan address=\"10.1007/ijag.2018.00453\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAnon (2021) Machine learning in agriculture: Current status and future perspectives. Agric Rev 34(4):250\u0026ndash;267. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/agriview.2021.034456\u003c/span\u003e\u003cspan address=\"10.1080/agriview.2021.034456\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBanbury CR, Reddi VJ, Jeffries N et al (2021) Benchmarking TinyML Systems: Challenges and Directions. ACM Transactions on Computer Systems\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBanerjee A, Singh P, Roy K (2025) Field implementation of ML models for weed detection. Smart Agric J 13(1):55\u0026ndash;70. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1234/saj.2025.130105\u003c/span\u003e\u003cspan address=\"10.1234/saj.2025.130105\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBhargava K, Rajesh A, Prasad G (2019) Agricultural IoT: An Overview and Future Directions. Int J Comput Appl 178(16):10\u0026ndash;17\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoateng L, Reddy M (2024) Yield gap analysis using machine learning: A case in Ghanaian maize farms. Afr J Agron 22(1):10\u0026ndash;23. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.7890/afja.2024.220102\u003c/span\u003e\u003cspan address=\"10.7890/afja.2024.220102\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrown J, Anderson L, Wright T (2024) Applications of machine learning in agriculture: Trends and future directions. Computers Agric 92(4):87\u0026ndash;101. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compagri.2024.107881\u003c/span\u003e\u003cspan address=\"10.1016/j.compagri.2024.107881\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Y, Wang L, Zhang Q (2020) Convolutional neural networks for plant disease detection: A review. Comput Electron Agric 172:105312. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compag.2020.105312\u003c/span\u003e\u003cspan address=\"10.1016/j.compag.2020.105312\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChike D, Smith R (2022) Genomics Agric 9(1):30\u0026ndash;44. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/genag.2022.010019\u003c/span\u003e\u003cspan address=\"10.1109/genag.2022.010019\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. Predicting pest resistance traits using genomic data and machine learning\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChung H, Prabhu R, Narayan R (2022) A survey on sensor fusion techniques in precision agriculture. J Agricultural Inf 13(2):45\u0026ndash;59\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDadhich M, Hiran KK (2022) Empirical investigation of extended TOE model on corporate environment sustainability. J Clean Prod 363:132309\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDey N, Ray R, Ashour AS (2023) Smart Farming Technologies for Smallholders. Elsevier\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDlamini B, Kassa T, Zhu M (2025) Optimizing agricultural supply chains using machine learning. Int J Agricultural Logistics 6(2):100\u0026ndash;115. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.4018/ijaglog.2025.06207\u003c/span\u003e\u003cspan address=\"10.4018/ijaglog.2025.06207\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDu Plessis T, Ho J (2020) Predictive modeling of soil moisture using machine learning. Soil Syst 4(3):55. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/soilsystems4030055\u003c/span\u003e\u003cspan address=\"10.3390/soilsystems4030055\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFernandez A, Liu X, Morris D (2022) Deep learning for high-throughput plant phenotyping. Plant Methods 18:93. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s13007-022-00849-3\u003c/span\u003e\u003cspan address=\"10.1186/s13007-022-00849-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGairola A, Rawat S (2021) Microcontroller-based Real-Time Soil Nutrient Detection using AI. Sens Appl 9(4):120\u0026ndash;134\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGupta M, Singh P (2021) Real-time monitoring of soil nutrients: Focus on potassium measurement. Agricultural Eng Today 45(3):17\u0026ndash;22\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHassan A, O\u0026rsquo;Connell D (2019) Modeling nutrient runoff using machine learning. Environ Model Softw 123:104568. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.envsoft.2019.104568\u003c/span\u003e\u003cspan address=\"10.1016/j.envsoft.2019.104568\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHossain MS, Muhammad G, Alhamid MF (2020) Green IoT-based agriculture: An integrated approach. Comput Electron Agric 172:105378\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJacobs S, Li D, Baloyi M (2024) Pollination strategy optimization using machine learning. Ecol Inf 73:101945. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ecoinf.2024.101945\u003c/span\u003e\u003cspan address=\"10.1016/j.ecoinf.2024.101945\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson H, Kato Y, Mwangi M (2020) Technological implementations in precision agriculture: A systematic review. Agric Syst 178:102736. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.agsy.2020.102736\u003c/span\u003e\u003cspan address=\"10.1016/j.agsy.2020.102736\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKamilaris A, Prenafeta-Bold\u0026uacute; FX (2018) Deep learning in agriculture: A survey. Comput Electron Agric 147:70\u0026ndash;90\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMyataza A, Mafunga M, Mkhulisi NS, Thango BA (2024) A Systematic Review of ERP, CRM, and HRM Systems for SMEs: Managerial and Employee Support. Preprints 2024100384\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGumede TT, Chiworeka JM, Magoda AS, Thango B (2024) Building Effective Social Media Strategies for Business: A Systematic Review. Preprints 2024100379\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim Y, Evans RG, Iversen WM (2020) Remote sensing and control of an irrigation system using a distributed wireless sensor network. IEEE Trans Instrum Meas 57(7):1379\u0026ndash;1387\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePatel R, Mehta P, Joshi A (2022) Advances in Machine Learning for Agricultural Applications. Springer\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhanyi MB, Xaba SN, Mlotshwa NA, Thango B, Matshaka L (2024) A Roadmap to Systematic Review: Evaluating the Role of Data Networks and Application Programming Interfaces in Enhancing Operational Efficiency in Small and Medium Enterprises. Sustainability 16(23):10192. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/su162310192\u003c/span\u003e\u003cspan address=\"10.3390/su162310192\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRavelo B, Ahmed S, Islam MR (2022) Smart Sensing and Monitoring for Agriculture: A Review. Smart Sens Devices 3(2):50\u0026ndash;63\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRoy S, Chattopadhyay S (2021) Embedded Systems in Precision Agriculture: A Review. Int J Embed Syst 13(1):67\u0026ndash;75\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMudau MC, Moshapo LW, Monyela TM, Thango BA (2024) The Role of Manufacturing Operations in SMEs Performance: A Systematic Review. Preprints 2024100539\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMtjilibe T, Rameetse E, Mgwenya N, Thango B (2024) Exploring the Challenges and Opportunities of Social Media for Organizational Engagement in SMEs: A Comprehensive Systematic Review. Available at SSRN 4998542\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThobejane LT, Thango BA (2024) Partial Discharge Source Classification in Power Transformers: A Systematic Literature Review. Appl Sci 14(14):6097\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSahu A, Sharma T, Bansal R (2020) Edge Intelligence for Agriculture: Deploying Machine Learning on Microcontrollers. IEEE Access\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSundmaeker H, Verdouw CN, Wolfert S, Freire P, L (2016) Internet of Food and Farm 2020. European Commission\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWarden P, Situnayake D (2019) TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers. O\u0026rsquo;Reilly Media\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang Q, Zhou Y, Han J (2023) Internet of Things and Artificial Intelligence in Agricultural Automation. Sens Syst 11(4):210\u0026ndash;229\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoateng L, Reddy M (2024) Yield gap analysis using machine learning: A case in Ghanaian maize farms. Afr J Agron 22(1):10\u0026ndash;23. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.7890/afja.2024.220102\u003c/span\u003e\u003cspan address=\"10.7890/afja.2024.220102\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBrown J, Anderson L, Wright T (2024) Applications of machine learning in agriculture: Trends and future directions. Computers Agric 92(4):87\u0026ndash;101. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compagri.2024.107881\u003c/span\u003e\u003cspan address=\"10.1016/j.compagri.2024.107881\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSkosana S, Mlambo S, Madiope T, Thango B (2024) Evaluating Wireless Network Technologies (3G, 4G, 5G) and Their Infrastructure: A Systematic Review. Preprints 2024101331\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMkhize A, Mokhothu KD, Tshikhotho M, Thango BA (2025) Evaluating the Impact of Cloud Computing on SME Performance. Syst Rev Businesses 5(2):23. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/businesses5020023\u003c/span\u003e\u003cspan address=\"10.3390/businesses5020023\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Y, Wang L, Zhang Q (2020) Convolutional neural networks for plant disease detection: A review. Comput Electron Agric 172:105312. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compag.2020.105312\u003c/span\u003e\u003cspan address=\"10.1016/j.compag.2020.105312\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChike D, Smith R (2022) Genomics Agric 9(1):30\u0026ndash;44. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/genag.2022.010019\u003c/span\u003e\u003cspan address=\"10.1109/genag.2022.010019\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. \u0026lsquo;Predicting pest resistance traits using genomic data and machine learning\u0026rsquo;\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDlamini B, Kassa T, Zhu M (2025) Optimizing agricultural supply chains using machine learning. Int J Agricultural Logistics 6(2):100\u0026ndash;115. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.4018/ijaglog.2025.06207\u003c/span\u003e\u003cspan address=\"10.4018/ijaglog.2025.06207\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDu Plessis T, Ho J (2020) Predictive modeling of soil moisture using machine learning. Soil Syst 4(3):55. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/soilsystems4030055\u003c/span\u003e\u003cspan address=\"10.3390/soilsystems4030055\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFernandez A, Liu X, Morris D (2022) Deep learning for high-throughput plant phenotyping. Plant Methods 18:93. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s13007-022-00849-3\u003c/span\u003e\u003cspan address=\"10.1186/s13007-022-00849-3\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHassan A, O\u0026rsquo;Connell D (2019) Modeling nutrient runoff using machine learning. Environ Model Softw 123:104568. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.envsoft.2019.104568\u003c/span\u003e\u003cspan address=\"10.1016/j.envsoft.2019.104568\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJacobs S, Li D, Baloyi M (2024) Pollination strategy optimization using machine learning. Ecol Inf 73:101945. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.ecoinf.2024.101945\u003c/span\u003e\u003cspan address=\"10.1016/j.ecoinf.2024.101945\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson H, Kato Y, Mwangi M (2020) Technological implementations in precision agriculture: A systematic review. Agric Syst 178:102736. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.agsy.2020.102736\u003c/span\u003e\u003cspan address=\"10.1016/j.agsy.2020.102736\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJones K, Mutiso D (2019) Smart seed selection using machine learning in East African maize crops. J Plant Breed 16(2):75\u0026ndash;88. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/jpb.2019.160275\u003c/span\u003e\u003cspan address=\"10.1080/jpb.2019.160275\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKgomotso N, Nair A (2023) Blockchain integration for agricultural data security. J Digit Agric 4(3):188\u0026ndash;202. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/joda.2023.04305\u003c/span\u003e\u003cspan address=\"10.1007/joda.2023.04305\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee S, Makhanya V (2020) Predictive maintenance of farm equipment using ML. Agricultural Eng Today 44(1):20\u0026ndash;28. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/aet.2020.011016\u003c/span\u003e\u003cspan address=\"10.1109/aet.2020.011016\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi H, Kim J (2023) Edge computing for precision agriculture applications. IEEE Internet Things J 10(2):1051\u0026ndash;1062. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1109/jiot.2023.3219990\u003c/span\u003e\u003cspan address=\"10.1109/jiot.2023.3219990\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Y, Walker C (2018) \u0026lsquo;Applications of deep learning in agriculture: A survey\u0026rsquo;, Computers and Electronics in Agriculture, 147, pp. 70\u0026ndash;90. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.compag.2018.02.030\u003c/span\u003e\u003cspan address=\"10.1016/j.compag.2018.02.030\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMartinez G, Kumar S (2022) Satellite imagery and deep learning for crop monitoring: A review. Remote Sens Agric 9(4):190\u0026ndash;204. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48924/rsa.2022.094.190\u003c/span\u003e\u003cspan address=\"10.48924/rsa.2022.094.190\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMbatha T, Kaur S (2023) Machine learning in diagnosis of plant nutrient deficiencies. Plant Health J 11(2):95\u0026ndash;110. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.41075/phj.2023.112.095\u003c/span\u003e\u003cspan address=\"10.41075/phj.2023.112.095\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMensah R, Gopal N (2022) Machine learning models for crop rotation planning. J Precision Agric 20(3):215\u0026ndash;229. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48493/jpa.2022.203.215\u003c/span\u003e\u003cspan address=\"10.48493/jpa.2022.203.215\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoyo P, Daniels J (2025) Soil nutrient management through precision agriculture technologies. Afr J Sustainable Agric 13(1):45\u0026ndash;59. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.71320/ajsa.2025.131.045\u003c/span\u003e\u003cspan address=\"10.71320/ajsa.2025.131.045\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMusa A, Taylor B (2024) Fertilizer optimization using machine learning: A simulation-based approach. J Agricultural Optim 6(1):28\u0026ndash;39. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.39125/jao.2024.061.028\u003c/span\u003e\u003cspan address=\"10.39125/jao.2024.061.028\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNaidoo C, Briscoe J, Xulu Z (2023) Predicting soil erosion using ML techniques. Land Degrad Dev 34(2):330\u0026ndash;342. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1002/ldr.4322\u003c/span\u003e\u003cspan address=\"10.1002/ldr.4322\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNdlovu S, White R (2023) Soil nutrient prediction using ensemble ML models. South Afr J Agricultural Eng 15(1):22\u0026ndash;33. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.40291/sajae.2023.151.022\u003c/span\u003e\u003cspan address=\"10.40291/sajae.2023.151.022\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOdhiambo L, Chan A (2023) Labour management in agriculture through machine learning. Int J Smart Farming 4(2):75\u0026ndash;91. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.60481/ijsf.2023.042.075\u003c/span\u003e\u003cspan address=\"10.60481/ijsf.2023.042.075\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOkafor U, Jansen K (2018) Artificial intelligence for pest monitoring in cassava farms. Appl AI Agric 3(3):44\u0026ndash;58. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.50741/aiaa.2018.033.044\u003c/span\u003e\u003cspan address=\"10.50741/aiaa.2018.033.044\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOrtega A, Rezaei R, Sun L (2021) Machine learning in agricultural robotics: A review. Robot Auton Syst 143:103801. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.robot.2021.103801\u003c/span\u003e\u003cspan address=\"10.1016/j.robot.2021.103801\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOsei J, Patel M (2021) A comparative analysis of ML algorithms for crop yield prediction. J Data Sci Agric 9(1):14\u0026ndash;29. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.41687/jdsa.2021.091.014\u003c/span\u003e\u003cspan address=\"10.41687/jdsa.2021.091.014\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePatel D, Singh H (2019) Sensor-based precision agriculture using machine learning. Smart Farming Rev 8(2):66\u0026ndash;78. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.70123/sfr.2019.082.066\u003c/span\u003e\u003cspan address=\"10.70123/sfr.2019.082.066\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePrasad V, Shah R (2020) Nutrient mapping using UAVs and ML techniques. J Remote Sens Agric 5(1):55\u0026ndash;67. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48019/jrsa.2020.051.055\u003c/span\u003e\u003cspan address=\"10.48019/jrsa.2020.051.055\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRodriguez L, Mokgosi B (2020) Policy modelling in agriculture using machine learning. J Agricultural Econ Policy 18(2):132\u0026ndash;148. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.68217/jaep.2020.182.132\u003c/span\u003e\u003cspan address=\"10.68217/jaep.2020.182.132\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSingh R, Petersen M (2021) Forecasting crop prices using machine learning. Agricultural Econ Rev 42(1):88\u0026ndash;101. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.57902/aer.2021.421.088\u003c/span\u003e\u003cspan address=\"10.57902/aer.2021.421.088\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSingh S, Yeboah E, Chen J (2025) IoT and ML integration for real-time NPK monitoring. Sens Smart Agric 10(1):55\u0026ndash;71. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.49187/ssa.2025.101.055\u003c/span\u003e\u003cspan address=\"10.49187/ssa.2025.101.055\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSmith A, Pillay T, Zondi K (2025) Systematic review on ML applications in sustainable agriculture. J AI Agric 12(1):33\u0026ndash;50. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.40361/jaia.2025.121.033\u003c/span\u003e\u003cspan address=\"10.40361/jaia.2025.121.033\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThomas L, Evans D (2024) Smart sensor technologies in precision agriculture. Agricultural Sens Networks 7(1):10\u0026ndash;26. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.42110/asn.2024.071.010\u003c/span\u003e\u003cspan address=\"10.42110/asn.2024.071.010\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eThomas M, Ncube B (2021) Estimating GHG emissions from farming practices using ML. Clim Agric 14(3):112\u0026ndash;125. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.60877/cla.2021.143.112\u003c/span\u003e\u003cspan address=\"10.60877/cla.2021.143.112\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTurner P, Li Y (2019) Forecasting plant diseases using ML models. Crop Prot Manage 17(2):79\u0026ndash;94. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.73822/cpm.2019.172.079\u003c/span\u003e\u003cspan address=\"10.73822/cpm.2019.172.079\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVerma A, Ndlovu S, Patel R (2019) Sensor fusion for comprehensive soil nutrient analysis. Agricultural Inf 5(3):44\u0026ndash;57. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.60212/aginfo.2019.053.044\u003c/span\u003e\u003cspan address=\"10.60212/aginfo.2019.053.044\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWu L, Garcia E (2022) Machine learning and artificial intelligence in smart farming. J Intell Agric 16(1):55\u0026ndash;73. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48562/jia.2022.161.055\u003c/span\u003e\u003cspan address=\"10.48562/jia.2022.161.055\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao X, Lee K (2024) Wireless sensor networks for NPK nutrient detection. Sens Transducers 31(1):33\u0026ndash;45. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.51172/snt.2024.311.033\u003c/span\u003e\u003cspan address=\"10.51172/snt.2024.311.033\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Macronutrient sensing, Nitrate, Nitrogen, Phosphate, Phosphorus, Potassium, Calcium, Magnesium, Sulfur, pH, Microcontrollers, Embedded systems, Machine learning, Internet of Things (IoT)","lastPublishedDoi":"10.21203/rs.3.rs-6842034/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6842034/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eThe integration of machine learning (ML) into microcontroller-class hardware for macronutrient sensing has shown increasing potential for enhancing environmental and agricultural monitoring. This systematic review synthesizes current trends, methodologies, and outcomes in this emerging field. A PRISMA-guided review was conducted across Google Scholar, Web of Science, and Scopus, yielding 2,546 initial records. After rigorous screening, 39 studies were selected based on relevance to ML-based macronutrient detection using microcontrollers. Publication types, sensor targets, hardware-software configurations, and validation metrics were analyzed. Publication peaked in 2021 (n = 10) with journal articles comprising 82% of studies. Google Scholar contributed 74% of sources. Research was geographically diverse, with Australia leading (33%), followed by Finland (10%) and several Asian and African countries. Studies predominantly targeted subsurface (31%) and agricultural water (18%), with pH, nitrate, and phosphate as common analytes. Nutrient concentration detection showed bias toward trace levels: 93% of nitrate studies used very low values (0.03–3.1 ppm); 92% of phosphorus studies focused on values ≤ 0.7 ppm. Potassium sensing emphasized high ranges (85%), while calcium reporting was more balanced (74% in moderate ranges). Magnesium and sulfur were minimally represented, with most studies focusing on low or moderate values (95%). Arduino platforms dominated (59%) and were mostly tied to microcontroller use (67%). Bluetooth (64%) was the most employed communication protocol, favoring low-power, short-range deployment. Cloud integration was common via AWS (33%) and ThingSpeak (28%), with 36% using open-source or custom solutions. Development tools were led by Arduino IDE (59%), while advanced AI integration was limited (~ 5%). Validation metrics favored R² (49%), followed by accuracy (21%), RMSE, and MAE. ML models (KNN, RF, DT) were occasionally used for model validation but often lacked consistent metric reporting. Embedded ML sensing for macronutrient detection is a fast-evolving multidisciplinary field. While nitrate and phosphate detection is well studied, potassium, magnesium, and sulfur remain underexplored. Gaps in reporting standards and methodological transparency hinder reproducibility. Future research should address these limitations while advancing deployment in low-resource settings.\u003c/p\u003e","manuscriptTitle":"Machine Learning -Based Macronutrient Sensing in Embedded Systems: A Review","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-11 08:00:49","doi":"10.21203/rs.3.rs-6842034/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"64b2bc7b-79dc-4c67-8d0e-ae4054e286a0","owner":[],"postedDate":"June 11th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":49679895,"name":"Marine and Freshwater Ecology"},{"id":49679896,"name":"General Biochemistry"},{"id":49679897,"name":"Applied Biochemistry"}],"tags":[],"updatedAt":"2025-06-11T08:00:50+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-11 08:00:49","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6842034","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6842034","identity":"rs-6842034","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}