TinyML Applications in Micronutrient Sensing: A Review of Microcontroller Deployments

TinyML Applications in Micronutrient Sensing: A Review of Microcontroller Deployments | 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 TinyML Applications in Micronutrient Sensing: A Review of Microcontroller Deployments Dineo Moeketsi, Alecia Mkhantshwa, Calvin Modise, Ntokozo Mlangeni This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6844555/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 Deploying machine learning (ML) models on microcontroller-class hardware offers a promising pathway for real-time micronutrient sensing, especially in resource-constrained agricultural and environmental contexts. Traditional sensing methods are often cost-prohibitive and lack real-time responsiveness, while ML-embedded systems enable portable, low-power, and scalable monitoring. This systematic review investigates global research trends in applying ML on microcontroller-class hardware for micronutrient sensing. It evaluates algorithm choices, dataset characteristics, hardware specifications, performance reporting, and real-time capabilities to identify critical gaps and future opportunities. The review followed PRISMA 2020 guidelines, analyzing 43 studies published between 2015 and 2025 sourced from Scopus, Web of Science, and Google Scholar. Eligibility criteria included English-language, peer-reviewed works focusing on ML techniques for real-time micronutrient sensing using microcontroller platforms. Data were synthesized and visualized across 14 key dimensions, including model type, sensor integration, hardware constraints, and deployment scenarios. Publication activity peaked in 2022, with growing contributions from countries like Israel and Kenya. Journal articles (51.16%) and conference papers (37.21%) dominated. Most studies (46.51%) were sourced from Google Scholar. Established frameworks such as TinyML were most frequently used (39.53%), while 32.56% of studies specified exact microcontroller boards. Deep learning (37.21%) and hybrid models (20.93%) were commonly applied, often using custom datasets (39.53%). However, 46.51% of studies lacked clear model size or latency reporting. Real-time performance was confirmed in 65.12% of cases, though only 11.63% provided quantified size and latency data. Hardware constraints were often generalized (30.23%), and 16.28% of papers omitted hardware details altogether. Environmental monitoring and smart IoT applications were the most common use cases (25.58% and 13.95%, respectively), supported by domain-specific ML tools (25.58%). ML on microcontroller-class hardware shows clear potential for enabling accessible, real-time micronutrient sensing. However, reproducibility remains limited due to insufficient reporting on model performance, hardware specifics, and deployment conditions. To accelerate adoption, future work should prioritize standardization in performance reporting, interdisciplinary collaboration, and deployment in real-world field environments. Chemical Biology Agrochemicals Chemical Engineering Machine Learning (ML) Micronutrient Sensing Edge Computing Quantized Models Microcontroller-Class Hardware Latency Optimization Standardization & Benchmarking 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 Introduction As modern technology continues to evolve at an unprecedented rate, we are witnessing a growing interest in how machine learning (ML) can be applied to small-scale, energy-efficient computing systems. Among these, microcontroller-class hardware—compact, low-power chips commonly found in embedded systems and Internet of Things (IoT) devices—has become an exciting platform for running ML algorithms (Ahmad et al., 2020). The ability to process data locally, without constant reliance on the cloud, has proven valuable for real-time decision-making, lower energy use, and enhanced privacy (Lane et al., 2015). These advancements have expanded the role of microcontrollers in various fields, including environmental monitoring, agriculture, and health diagnostics, where immediate and reliable insights are increasingly in demand (Fischer et al., 2020; Chen et al., 2022). impactful area where ML-enabled microcontrollers are showing promise is micronutrient sensing. An impactful area where ML-enabled microcontrollers are showing promise is micronutrient sensing. This refers to detecting and measuring essential nutrients like iron, zinc, or vitamin A, which are crucial for human health but often lacking in both food and soil—especially in low-resource settings (Gibson, 2012). Conventional lab-based approaches to micronutrient testing are typically accurate but tend to be expensive, time-consuming, and inaccessible to communities that lack infrastructure (Aziz & Butt, 2019). The integration of lightweight ML models into embedded devices offers a novel solution to this problem, enabling real-time, portable, and cost-effective sensing platforms that can be used in the field with minimal training or technical support (Anwar et al., 2021). Devices based on microcontroller platforms such as Arduino, ESP32, or ARM Cortex-M chips are being developed to bridge this gap and make micronutrient analysis more accessible (Kamble & Kale, 2021). However, despite growing interest in this area, the implementation of ML models on microcontroller-class hardware still faces considerable obstacles. Limited processing power, memory, and battery life make it difficult to run traditional ML algorithms on these platforms. Since this field lies at the intersection of multiple disciplines—including embedded hardware, nutritional science, chemistry, and artificial intelligence—the Designing efficient, accurate, and reliable models that can work under such constraints is an ongoing challenge (Sze et al., 2017). Additionally, there’s a lack of comprehensive insight into how these systems are applied across different micronutrient sensing scenarios. current research landscape remains fragmented, making it harder to build consistent, scalable solutions (Fischer et al., 2020). Given the global importance of improving micronutrient monitoring, especially in areas affected by malnutrition or poor agricultural yields, a systematic overview of how ML is being implemented in these compact, embedded systems is needed. This review addresses that need by analyzing studies published over the past decade, focusing specifically on how machine learning is deployed on microcontroller-class hardware for micronutrient sensing. The review will evaluate the design approaches, deployment environments, data handling strategies, and performance outcomes of these systems to understand the broader trends, key challenges, and emerging opportunities in this domain. By bringing together research from different fields, this review also offers insights for developers, scientists, and decision-makers interested in applying embedded ML technologies to global health and nutrition challenges (Do Valle & Lee, 2020). Table 1 provides a comparative view of past review efforts alongside the specific scope of this work, underlining how it uniquely focuses on the intersection of embedded AI, sensor design, and micronutrient diagnostics. Through this systematic review, we aim to offer practical takeaways for future innovation, support interdisciplinary collaboration, and contribute to the development of more scalable and sustainable micronutrient sensing systems that can be used in real-world conditions (Patil & Kale, 2020). Table 1 Comparative Analysis of the Existing Review Works and Proposed Systematic Review on the Applications and Advantages of Machine Learning on Microcontroller-Class Hardware for Micronutrient Sensing Ref Contribution Pros Cons Pereira, E. A et al(2024) Proposed an energy-efficient TinyML model using Random Forest and Neural Networks for classifying water potability in embedded systems Real-time responses, no need for internet, low energy use, long battery life, supports remote deployment Limited to electronic-sensor-accessible data, Random Forest accuracy at 0.70 may not suit all use cases, model complexity may limit real-time adaptation Dhal, S. B et al (2022) Developed a ML-based IoT system for nutrient optimization in commercial aquaponics using sensor data and feature selection Real-world data from farms, automated nutrient control, identified key predictors, supports healthy plant and fish growth Limited data points, only weekly measurements, focused on lettuce and tilapia, commercial-scale generalization may be limited Pereira, E. A. M et al (2024) Developed a TinyML model for classifying water potability using only electronic sensor data on ESP32 High energy efficiency, fast inference (99.95% faster), reduced memory use (51.2% less), no need for internet or lab data Limited to sensors' data; chemical parameters not considered; may need calibration across diverse water sources GASANA, J. M. (2022). Developed an IoT-based system to monitor soil properties and predict suitable crop type using ML (decision tree classifier) Real-time monitoring, accurate crop prediction (99%), reduces lab testing, supports data-driven farming Relies on internet connection (GPRS), limited to measured properties (NPK, pH, etc.), may require calibration for different regions Schizas, N., Karras et al ( 2022 ) Systematic review of TinyML's role in low-power AI and IoT deployments, highlighting frameworks, benefits, and integration with 5G/LPWAN Summarizes current research, promotes on-device analytics, reduces latency and cloud dependence, improves privacy Lacks implementation details, review-based (no new model or case study), practical limitations not deeply explored Dalmeida, S. P et al(2023) Developed a CNN-based system using image processing to detect soil micronutrients (Zn, Fe, Mn, Cu, pH) High accuracy (95%), avoids lab testing, uses simple smartphone images, automates micronutrient detection Manual image capture, limited to tested micronutrients, depends on image quality, MATLAB-based limits portability Salve, P. R et al(2025) Explores use of high-performance microcontrollers and ML for real-time food quality monitoring across the supply chain Real-time analysis, IoT and ML integration, improved food safety, reduces waste, predictive insights Limited availability (forthcoming), no implementation case study yet, may require advanced hardware and setup Venkatesh, K et al (2023) Developed a lightweight CNN model to identify nitrogen, phosphorus, and potassium deficiencies in groundnut plants from leaf images High accuracy (94.64%), efficient detection, reduces need for lab testing, cost and time effective Limited to groundnut plants, model performance depends on image quality, may need retraining for other crops or environments El Adoui, M et al ( 2024 ) Developed a constrained TinyML model to accurately predict gas concentrations using low-cost sensors on MCUs High prediction accuracy (R² = 0.72), low RAM (3%) and Flash (98%) usage, adaptable to real-time calibration, cost-effective Constrained by microcontroller memory, limited generalization to non-tested environments, requires model porting and tuning Kiplimo, E et al ( 2024 ) Developed a low-cost sensor system combined with ML for accurate methane monitoring at atmospheric levels High accuracy with errors in the tens of ppb, versatile for indoor/outdoor environments, robust calibration method Requires periodic calibration with reference equipment, potential variability in sensor lifespan and environmental effects. Additionally, literature on applying machine learning (ML) to microcontroller-class hardware for micronutrient sensing. These gaps reveal limitations in existing studies and offer opportunities for future research to deepen understanding and expand practical applications in this growing field.To begin with, most research in this area tends to focus on general-purpose embedded machine learning or nutrient analysis tools, but very few studies investigate how lightweight ML models can be optimized specifically for real-time micronutrient sensing using constrained microcontroller devices. This narrow focus leaves out important considerations unique to low-power, low-memory environments—such as energy efficiency, real-time inference accuracy, and cost constraints—that are critical for deploying these systems in real-world field conditions, particularly in low-resource settings. Furthermore, much of the literature emphasizes the technical side of hardware-software integration, while overlooking broader practical challenges like usability, maintenance, and deployment in remote or rural environments. The roles of user interaction, accessibility, and training are often left unaddressed, despite their importance in successful long-term implementation.Another significant gap lies in the lack of interdisciplinary research connecting machine learning techniques with domain-specific knowledge from agriculture, food science, or public health. While technical feasibility has been demonstrated in lab settings, few studies explore how these systems perform in real-life scenarios, where environmental factors, sensor variability, and data noise can significantly affect accuracy. Additionally, most Current findings rely on isolated or short-term experiments, limiting our understanding of how such systems behave over longer periods or across diverse deployment conditions. Finally, there is a need for more studies that evaluate the full system lifecycle—from model training and deployment to maintenance and real-time feedback loops. Few existing works assess how these ML-enabled microcontroller systems can be updated or scaled efficiently, especially in settings without reliable internet connectivity or technical expertise. By addressing these research gaps, future studies can provide clearer insights into how to design, implement, and scale machine learning solutions for micronutrient sensing in ways that are accessible, cost-effective, and impactful. 1.1. Research questions Despite notable progress in machine learning (ML), the integration of ML models into microcontroller-class hardware remains a technically complex and underexplored domain—particularly for applications like micronutrient sensing in agriculture and environmental monitoring. These constrained platforms face unique limitations in computational power, memory capacity, and energy efficiency, making the selection of appropriate ML frameworks, hardware architectures, and deployment strategies critical for real-time, field-based operations. This systematic review aims to investigate how lightweight ML techniques, including TinyML and other edge-optimized models, can be effectively adapted for accurate, low-latency micronutrient detection in resource-constrained environments. In particular, the review seeks to understand the interplay between model architecture, dataset characteristics, hardware specifications, and performance metrics. To guide this investigation, the following research questions were formulated: Which types of machine learning algorithms are most frequently recommended and successfully implemented on microcontroller-class hardware for micronutrient detection? What types of sensors are most suitable and commonly used in micronutrient sensing systems based on microcontrollers? What are the typical performance outcomes (e.g., accuracy, latency, energy efficiency) observed when deploying ML models for micronutrient sensing on constrained devices? How do existing studies address key challenges such as limited memory, real-time processing demands, and power consumption in microcontroller-class environments? To what extent are ML-enabled microcontroller systems being deployed or validated in real-world applications, including agricultural, environmental, and low-resource settings? Which micronutrients are being targeted in ML-based sensing applications on microcontroller-class hardware, and how consistently are these reported across studies? 1.2. Hypotheses Development Building upon the research questions, the following hypotheses were developed to explore the patterns, performance, and practical constraints of deploying machine learning (ML) models on microcontroller-class hardware for micronutrient sensing. These hypotheses reflect the observed trends in model selection, dataset design, hardware utilization, and application context identified in the reviewed studies. H1: Convolutional Neural Networks (CNNs), ensemble models, and classical algorithms like decision trees are the most commonly used ML models on microcontroller-class hardware due to their balance of performance and resource efficiency. (Supported by Fig. 14 : CNNs/Deep Learning – 37.21%; Ensembles – 20.93%; Decision Trees – part of 16.28% classical ML) H2: Electrochemical sensors are the most suitable for microcontroller-based micronutrient sensing due to their low power demands and integration ease. (Inferred from studies integrating water/soil-based nutrient analysis, though sensor type was underreported in ~ 20% of studies) H3: Model complexity is often reduced (e.g., via quantization) to enhance energy efficiency and real-time processing, sometimes at the cost of accuracy. (Supported by Figs. 16 and 18 : 23.26% used quantized models; only 11.63% fully reported model size and latency) H4: Most studies identify and partially address hardware limitations such as limited memory and low processing power, but often lack quantitative reporting. (Supported by Figs. 12 & 17 : 30.23% described general constraints; 16.28% failed to specify hardware at all) H5: There is increasing adoption of ML-enabled microcontroller systems in low-resource applications such as soil testing, water quality analysis, and environmental monitoring. (Supported by Fig. 19 : Smart IoT – 25.58%; Environmental – 13.95%; Agriculture – 13.95%) H6: Micronutrients such as zinc, iron, nitrogen, and phosphorus are the most commonly targeted nutrients, though reporting is inconsistent. (Supported by Figs. 13 & 15 : Custom datasets often include micronutrient data, but few studies report targets explicitly) 1.3. Rationale The motivation for this systematic review lies in the urgent need to understand how lightweight machine learning (ML) frameworks can be deployed on microcontroller-class hardware for real-time, in-situ micronutrient sensing. Conventional lab-based nutrient analysis methods are costly, infrastructure-dependent, and impractical for widespread deployment in rural or low-resource settings. In contrast, embedded ML systems offer scalable, portable, and cost-effective sensing platforms, especially when optimized for low-power operation. This review focuses on how ML models—particularly those adapted for edge computing via frameworks like TinyML and TensorFlow Lite—perform in real-world or simulated environments when deployed on hardware-constrained platforms such as ESP32, STM32, and ARM Cortex-M series. By analyzing studies from 2015 to 2025, this review highlights the technical trade-offs between model complexity, memory usage, energy consumption, and accuracy, while also evaluating how dataset quality, sensor integration, and deployment environments affect sensing outcomes. 1.4. Objectives The primary objective of this systematic review is to evaluate how machine learning (ML) models are applied on microcontroller-class hardware for micronutrient sensing, with particular emphasis on performance, optimization, and deployment feasibility in constrained environments. Specific objectives include: To identify the most common ML models and frameworks used in micronutrient sensing applications, and assess their suitability for embedded systems. To examine hardware specifications and sensor integration approaches, particularly with respect to power efficiency, memory limitations, and real-time constraints. To analyze model performance metrics such as accuracy, latency, and energy efficiency, and evaluate the degree of reporting transparency across studies. To categorize deployment scenarios (e.g., soil testing, water analysis) and assess the extent of real-world validation of these systems. To highlight reporting inconsistencies and research gaps, providing recommendations for future development of scalable, low-cost sensing solutions using embedded ML. 1.5. Research Contributions This review provides a comprehensive synthesis of the current state of ML-enabled micronutrient sensing on microcontroller-class hardware. Key contributions include: A technical evaluation of lightweight ML frameworks—such as TinyML and TensorFlow Lite—and their integration with microcontrollers for real-time micronutrient detection, highlighting trade-offs in model size, latency, and energy usage (Figs. 14 , 16 , 18 ). A comparative analysis of hardware utilization, microcontroller specifications, and reporting practices, revealing that only 32.56% of studies specified board-level hardware and over 16% lacked hardware descriptions entirely (Figs. 12 , 17 ). A taxonomy of dataset types and application domains, showing dominant use of custom datasets (39.53%) and strong representation of IoT, agriculture, and environmental monitoring as application areas (Figs. 15 , 19 ). The development of a benchmarking framework that maps model characteristics (e.g., size, inference latency) to platform constraints, aiding developers in selecting appropriate ML solutions for field-deployable sensing. 1.6. Research Novelty To the best of our knowledge, this is the first systematic review exclusively focused on the deployment of ML models on microcontroller-class hardware for micronutrient sensing across environmental and agricultural contexts. Its novelty lies in: Providing a cross-disciplinary synthesis of embedded systems, AI, and environmental sensing literature, with special attention to technical deployment on constrained devices. Offering a framework for assessing model-hardware compatibility, linking ML model performance metrics with microcontroller specifications and real-time operating needs. Uncovering underreported areas such as sensor type documentation, micronutrient-specific datasets, and standardized benchmarking practices—thus setting the agenda for future research in this growing subfield of edge AI for agriculture and environmental health. Materials and Methods In this subsection, the study outlines the methodology employed to conduct a systematic review focusing on the applications and advantages of machine learning on microcontroller-class hardware for micronutrient sensing. The study is based on a review of literature published over the last decade, from 2015 to 2025. To the best knowledge of the authors, no similar comprehensive review exists within this specific timeframe, making this study a novel contribution to the field of machine learning on microcontroller-class hardware for micronutrient sensing. The research methodology includes the careful selection of relevant peer-reviewed articles from key online databases, namely Scopus, Google Scholar, and Web of Science, ensuring a thorough examination of the subject matter. 2.1. Eligibility criteria A systematic analysis of all peer-reviewed and published research works relevant to the study of the machine learning on microcontroller-class hardware for micronutrient sensing was conducted for examination. Only research tasks published in English between 2015 and 2025 were included in the analysis. A proper benchmark for inclusion was adapted to ensure the inclusion of research papers that specifically focus on this topic in hand and exclude those that do not fall under this topic. Inevitably, only peer-reviewed research works that fundamentally connect on the machine learning on microcontroller-class hardware for micronutrient sensing, and that include a research framework or methodology specific to these aspects, were exclusively considered. The inclusion and exclusion criteria for this review are tabulated as in Table 2 (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). Table 2 Proposed Inclusion and Exclusion Criteria. Criteria Inclusion Exclusion Topic Article papers focusing on machine learning on microcontroller-class hardware for micronutrient sensing Article papers not focusing on machine learning on microcontroller-class hardware for micronutrient sensing Research Framework The Articles must include research frame-work or methodology for machine learning on microcontroller-class hardware for micronutrient sensing Articles must exclude research framework or methodology for machine learning on microcontroller-class hardware for micronutrient 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 study of online information was held to identify relevant studies for this systematic review. online sources like Scopus, Google Scholar, and Web of Science were utilized to find articles and journals that involve the review on machine learning on microcontroller-class hardware for micronutrient sensing. On each online source, keywords that are related to the topic where mixed together to form a strong search pattern, that will at the end provide the most relevant research articles that are already recorded (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). Google scholar was used to find conference papers, journal papers and book articles while web of science was used for citations of the literature found on google scholar website. Web of science was also utilized to find articles. The results from these online data sources formed the core of the literature review. 2.3. Search strategy The literature studies that are obtained in this research were taken from online databases, focusing on keywords that address the context of machine learning on microcontrollers-class hardware for micronutrient sensing. Terms such as “key elements: AND “ best practices”. A thorough search was carried out in three main online sources: Google Scholar, Scopus, and Web of Science. To find the most relevant studies, a specific set of keywords was used. These keywords were: “(machine learning)” AND “on” AND “(microcontroller)” AND “(class-hardware)” AND “for” AND “(micronutrients)” AND “(sensing)”. The terms were utilized as a search string to find studies that are relevant to machine learning on microcontroller-class hardware for micronutrients sensing. The research also focused on studies done in the last decade from 2015 to 2025 that were all written in English language. This time frame was chosen to get the newest articles that fall under the research topic. The results from the search conducted showed 17 100 studies found on Google Scholar, 1577 articles found on Scopus, and 136 papers from Web of Science.the articles that are selected, were selected carefully after going through them again to ensure that only those that are relevant to the research topic are included in when writing this review(Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). Table 3 below shows the total number of articles found in each of the three main online sources that were used while Fig. 1 illustrates the keywords used in the search string. Table 3 Results Achieved from Literature Search. No. Online Sources Number of results 1 Google Scholar 17 100 2 Web of Science 136 3 Scopus 1577 Total 18813 2.4. Selection process To begin the screening process, four reviewers looked for titles and abstracts of the first search results. When there were parts where we disagreed with each other, we would all come to a final agreement that got more votes from all 4 members. Following the initial screening, the researchers divided into pairs and independently assessed the titles and abstracts of all articles obtained from the search. Whenever disagreements occurred, the team engaged in discussions to reach a consensus on which articles should advance to the full-text review stage (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). If consensus could not be reached, a third researcher was brought in to make the final determination. Following this, three researchers independently evaluated the full-text articles to verify their eligibility based on the inclusion criteria. As with previous steps, any disagreements were addressed through discussion. When necessary, the fourth researcher was consulted to make the final decision regarding the inclusion or exclusion of articles, as illustrated in Fig. 2. 2.5. Data collection process To reduce errors and cut back on bias, we followed an appropriate approach to ensure that correct data was obtained from the studies. One member in the team focused on making sure that every member was collecting accurate data and then examined it thoroughly (Myataza et al, 2024 ; Gumede et al, 2024 ; Mudau et al., 2024 ; Mtjilibe et al, 2024 ). Discussions were held to come to a common conclusion when slight difference was identified in the data No automation tool was used for data distillation and everything was double checked before they were written down on the literature and existing studies sheets. When details in the studies were ambiguous, we carefully examined all accessible materials, such as supplementary data, appendices, and related research, to clarify findings concerning machine learning on microcontroller-class hardware for micronutrient sensing. In situations where uncertainties persisted, we referred to our fourth reviewer, an expert in the subject matter, to confirm the accuracy of the data interpretation related to machine learning on microcontroller-class hardware for micronutrient sensing. When multiple reports from the same study were available, we established clear criteria to select the most relevant data, focusing on the most recent and comprehensive studies on machine learning on microcontroller-class hardware for micronutrient sensing published between 2015 and 2025. In situations where discrepancies existed between reports, we examined the methodologies and results concerning machine learning on microcontroller-class hardware for micronutrient sensing to address the inconsistencies. Only English-language studies were considered, with articles in other languages excluded to ensure consistency in our analysis and to prevent possible misinterpretations caused by language barriers, as illustrated in Fig. 3. 2.6. Data items In this section we focus on key data items targeted in this systematic review, focusing on both core outcomes and supplementary variables related to the application of ML on microcontroller-class hardware for micronutrient sensing (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). The primary outcomes include sensor accuracy, detection limits, power efficiency, hardware-software integration, and real-time processing capabilities. Variables that were considered are as follows; study design, hardware specifications (e.g., microcontroller type, memory, processing power), ML model characteristics, environmental deployment settings, calibration strategies, and cost-effectiveness. This comprehensive approach enables a detailed understanding of how ML techniques enhance the performance and feasibility of micronutrient sensing in low-resource or embedded systems. By examining a wide range of technical and contextual factors, the review offers insights into the scalability, reliability, and practical implementation of such systems across diverse use cases. 2.6.1 Data Collection Method To ensure a complete understanding of the role of machine learning (ML) on microcontroller-class hardware for micronutrient sensing, this systematic review identified and clearly defined key outcomes capturing the technical, functional, and practical lengths of this technology merge (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). Our data collection plan was designed to incorporate robust evidence on the practicality, performance, and real-world applicability of such systems in environmental and health-related sensing tasks. The primary outcomes of interest included Sensor Accuracy, measured through detection limits, sensitivity, and precision in micronutrient quantification. We gathered all results that demonstrated how ML models enhanced signal processing and classification accuracy on constrained hardware platforms. Energy Efficiency was another critical outcome, reflecting the system's suitability for deployment in remote or low-power environments. Studies reporting on power consumption, computational efficiency, and battery life were included to evaluate the practicality of sustained field operation. Hardware Utilization was also examined by assessing microcontroller specifications such as memory usage, processing speed, and compatibility with ML models. This outcome helped gauge how effectively ML algorithms were optimized to function within limited hardware capabilities. Model Performance and Training Procedures were analyzed to understand how ML approaches such as regression, classification, or neural networks were adapted for lightweight execution. We focused on studies that provided insights into model size, inference speed, and training protocols, both on-device and offloaded. Finally, Application Context and Deployment was assessed by identifying how and where these systems were implemented—whether in soil, water, plant tissues, or food samples—and evaluating the reliability of real-time micronutrient monitoring under varying environmental conditions. All relevant data across studies and deployment scenarios were included to capture a holistic view of the technological impact and scalability of ML-enabled micronutrient sensing on microcontroller-class hardware. 2.6.2 Definition of Collected Data Variables In addition to the primary outcomes, we carefully considered several additional variables to provide a complete understanding of the context in which machine learning (ML) models are applied on microcontroller-class hardware for micronutrient sensing. These variables were critical to investigate the findings and understand the broader implications of deploying such technologies in resource-constrained environments. Study characteristics were collected, including the geographical location, environmental setting (e.g., agricultural, water monitoring, or food safety), and the specific micronutrients targeted. This information allowed us to assess the applicability of findings across various use cases and deployment scenarios (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). We also documented hardware specifications and system configurations, including microcontroller type, processing power, memory size, and sensor integration, to evaluate how hardware constraints influence ML implementation. These technical parameters were essential for understanding compatibility and scalability. Details on the ML models used such as algorithm type, training method, inference process, and optimization techniques were also included to assess the computational demands and real-time performance of different approaches. Furthermore, sensor data characteristics such as sampling frequency, data preprocessing steps, and noise handling were gathered to evaluate the reliability of sensing inputs. Energy consumption metrics, deployment duration, and calibration frequency were recorded to understand long-term feasibility and operational sustainability. We also considered external influences like climate variability, field conditions, and regulatory standards affecting sensor deployment and ML model performance in real-world applications. As detailed in Table 4 , our systematic review involved structured manual searches across established scientific databases such as Google Scholar, SCOPUS, and Web Of Science. These searches were tailored to capture studies at the intersection of embedded ML, micronutrient detection, and low-power hardware systems. By identifying and defining these variables, our review delivers a powerful and context-rich analysis of ML-enabled micronutrient sensing on microcontroller-class platforms, offering valuable insights for researchers, developers, and policy-makers working toward sustainable and intelligent sensing technologies. Table 4 Data Variables Collected Field Description Study characteristics Geographic location, environmental application (e.g. water monitoring), type of micronutrients analyzed, and deployment setting (lab vs. field). Hardware characteristics Specifications of microcontroller-class hardware used, including processor type, memory capacity, energy consumption, and sensor integration. ML implementation characteristics Types of machine learning algorithms applied (e.g., regression, classification), training methods (on-device vs. offline), model optimization techniques, and inference performance Operational Constraints Power supply limitations, environmental factors (e.g., humidity, temperature), and maintenance or deployment challenges in field conditions.. External influences Regulatory standards for nutrient detection, funding availability for low-cost sensing technologies, and policy incentives for water monitoring. 2.7. Study risk of bias assessment The studies reviewed under the topic was critical to evaluate the risk of bias to ensure the credibility and accuracy of the synthesized findings. To accomplish this, we employed the Newcastle-Ottawa Scale (NOS) for assessing non-randomized studies, including experimental testbeds, simulation analyses, and field-based evaluations. The NOS framework allowed assessment across three key domains: Selection, Comparability, and Outcome. Each included study received a quality score based on criteria such as the clarity of sensor deployment methods, the appropriateness of machine learning model training, and the reproducibility of results (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). A maximum of one star was allocated per item in the Selection and Outcome domains, and up to two stars for Comparability, particularly focusing on how well studies accounted for environmental variables like humidity, temperature, and sensor drift. As illustrated in Fig. 4 , four independent reviewers conducted assessments to enhance objectivity. Discrepancies were discussed among the reviewers, and if consensus was not reached, a fourth reviewer provided the final judgment. For studies that lacked detail or relied on proprietary hardware or algorithms, additional efforts were made to cross-reference supplementary sources from repositories like Google Scholar, Web Of Science and SCOPUS A detailed manual review of grey literature was also performed to ensure that all relevant information was accurately captured. No automation tools were used, maintaining the integrity and rigor of the bias assessment process. 2.8. Synthesis methods The flowchart in Fig. 5 outlines the systematic methodology adopted in our review of ML applications on microcontroller-class hardware for micronutrient sensing. Beginning with the Study Selection Process, studies were identified and screened using defined inclusion and exclusion criteria, focusing on research involving embedded ML models and low-power sensor hardware (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). The next step, Data Standardization, involved organizing technical specifications, performance metrics, and sensing parameters into a consistent format for comparability. In the Data Analysis phase, information was structured into tables and visualizations to enable cross-study comparison of model accuracy, power consumption, and sensor integration. The review then proceeded to a Heterogeneity Assessment, where differences in hardware platforms, ML algorithms (e.g., decision trees, support vector machines, neural networks), and micronutrient targets were examined through subgroup analyses. Lastly, a Bias Assessment step was conducted to identify methodological limitations and ensure a balanced synthesis of the findings. This structured workflow enabled a comprehensive, transparent, and reproducible evaluation of how ML techniques are leveraged on microcontroller platforms for real-time, in-field micronutrient detection. In this systematic review on the application of ML on microcontroller-class hardware for micronutrient sensing, we employed rigorous synthesis methods to ensure our findings were comprehensive, transparent, and reproducible. To establish study eligibility, we systematically tabulated each study's characteristics including hardware specifications, sensor configurations, ML algorithms used, and nutrient targets and compared them against our predefined synthesis criteria. This process ensured the inclusion of only those studies directly relevant to embedded ML in resource-constrained sensing environments. In preparing the data for synthesis, we addressed missing values and incomplete reporting by applying estimation techniques and standardized unit conversions, thereby maintaining consistency across diverse technical studies. Results were synthesized using structured comparison tables and performance graphs to visualize metrics such as model accuracy, power efficiency, latency, and sensor integration effectiveness. We conducted a qualitative comparative analysis due to the technological variability across studies, supplemented by subgroup analyses focusing on microcontroller types (e.g., ARM Cortex-M, AVR) and the classes of ML models (e.g., decision trees, SVM, neural networks). This allowed us to explore the influence of computational constraints and model complexity on sensing accuracy. We also performed sensitivity analyses to assess the robustness of findings across various deployment environments, including laboratory, field, and in-situ agricultural settings. This comprehensive synthesis approach provided a good understanding of the trade-offs and practical considerations when deploying ML on microcontrollers for micronutrient detection, offering valuable insights for researchers, developers, and practitioners in embedded systems and precision agriculture. 2.8.1. Eligibility for Synthesis To determine study eligibility for inclusion in our systematic review on the use of ML on microcontroller-class hardware for micronutrient sensing, each study was meticulously evaluated for its relevance and alignment with the review’s specific objectives (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). We manually assessed and compared key study characteristics such as the type of microcontroller used, sensor integration, ML algorithms implemented, and targeted micronutrients against our predefined inclusion criteria. A comparison matrix was developed to visually map the scope, methodologies, and performance metrics of each study relative to our eligibility benchmarks. This structured approach ensured that only studies directly focused on embedded ML solutions for nutrient detection in constrained hardware environments were included, thereby enhancing the precision, reliability, and contextual relevance of our review findings. 2.8.2. Data Preparation for Synthesis In this review, data preparation methods were employed to standardize and harmonize findings across studies focused on machine learning implementation on microcontroller-class hardware for micronutrient sensing. Given the diversity in reporting formats, data such as detection accuracy, response times, and hardware resource usage were converted into comparable metrics to enable meaningful synthesis. When studies reported performance indicators differently for example, using classification accuracy versus F1-score appropriate statistical transformations were applied to align them onto a consistent evaluative scale. Additionally, in cases where studies lacked critical summary statistics, such as standard deviations or confidence intervals, imputation techniques and estimation formulas based on available data were used to fill these gaps. This ensured a complete and coherent dataset, facilitating a reliable comparison and integration of findings across the diverse body of research examined. 2.8.3. Tabulation and Visual Display of Results Results from individual studies and synthesized analyses in this review of machine learning on microcontroller-class hardware for micronutrient sensing were organized using both tabular and graphical formats to enhance interpretability and comparison. Structured tables were created to present study outcomes, categorized by performance domains such as detection accuracy, computational efficiency, energy consumption, and sensor integration. Within each domain, studies were arranged according to their assessed risk of bias, allowing readers to identify the most credible evidence at a glance. For visual representation, forest plots were employed to display quantitative synthesis results, illustrating effect sizes and confidence intervals for each included study alongside overall estimates. These visualizations were organized either by publication year or magnitude of effect, which helped reveal technological progress over time and emerging patterns in microcontroller-based sensing strategies. This dual approach to presentation ensured both clarity and analytical depth across the dataset. 2.8.4. Synthesis of Results In this systematic review of machine learning applications on microcontroller-class hardware for micronutrient sensing, we conducted a thorough synthesis of results gathered from manual searches across reputable repositories, including Google Scholar, Scopus, and Web of Science. The synthesis process was carefully guided by the nature of the data and the degree of heterogeneity observed among the studies. Each study’s results were evaluated to determine whether a fixed-effects or random-effects model would be more appropriate, based on the consistency of outcomes across various hardware platforms and sensing conditions. Following data extraction, we compiled the findings into Excel spreadsheets, where visualizations such as bar charts and scatter plots were created to support preliminary assessments of variability. This visual analysis allowed us to detect trends in model accuracy, processing latency, and power consumption, and to identify any notable divergences in performance across different microcontroller architectures. These insights provided a foundation for deeper exploration and contextual interpretation of the results. 2.8.5. Exploring Causes of Heterogeneity In this review, data preparation methods were employed to standardize and harmonize Subgroup analyses and meta-regression were conducted to investigate potential sources of heterogeneity among studies evaluating the use of machine learning on microcontroller-class hardware for micronutrient sensing. Key factors explored included differences in hardware platforms (e.g., Arduino, Raspberry Pi, ESP32), types of machine learning algorithms implemented (e.g., decision trees, neural networks, support vector machines), and sensing modalities (e.g., optical, electrochemical, or biosensor-based techniques). These variables were examined to assess their influence on system performance, including accuracy, power efficiency, and processing time. By analyzing these subgroups, we aimed to identify consistent patterns and contextual dependencies that could explain the variability observed across studies and enhance the generalizability of findings within low-resource or embedded environments. 2.8.6. Sensitivity Analyses Sensitivity analyses were conducted to assess the robustness of the synthesis results concerning key assumptions and methodological choices in this review of machine learning on microcontroller-class hardware for micronutrient sensing. These analyses involved re-evaluating the findings by excluding studies with high risk of bias or limited methodological transparency, and by applying alternative statistical models to test the stability of aggregated outcomes. We also examined the influence of varying hardware configurations, sensor types, and algorithmic complexity on reported performance metrics. This rigorous approach ensured that our conclusions remained consistent across different analytical frameworks, enhancing the reliability and validity of the synthesized evidence. 2.9. Reporting bias assessment In conducting our systematic review on the application of machine learning on microcontroller-class hardware for micronutrient sensing, assessing the risk of reporting bias was critical to ensuring the integrity and validity of our synthesis. We focused on identifying potential biases arising from selective publication or incomplete reporting of outcomes, as these could significantly skew the representation of technological capabilities and limitations. To address this, we utilized contour-enhanced funnel plots as a key visual tool for detecting asymmetry in reported performance metrics, such as accuracy, latency, and power consumption. These plots, complemented by significance contours, allowed us to distinguish between potential publication bias and random variation (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). Rather than developing new assessment tools, we employed established methods recognized in the literature for their methodological rigor and suitability for technical evaluations. All assessments were conducted manually, using tools like Excel to generate plots and charts, enabling us to maintain a hands-on approach and closely examine study distributions. This manual process also ensured that nuanced patterns, often overlooked by automated tools, were properly captured and analyzed. Multiple reviewers independently assessed each study for reporting bias, with discrepancies resolved through consensus or consultation with an expert in embedded machine learning systems. In addition, we conducted an exhaustive manual search of databases including Google Scholar, Scopus, and Web of Science to identify studies potentially missed by automated indexing, particularly unpublished or conference-based work commonly found in hardware-focused research. Given the engineering and sensor-technology focus of our topic, we adapted traditional bias assessment frameworks to reflect the reporting norms and dissemination practices in this domain. This tailored approach improved the contextual relevance of our analysis. All assessment procedures and criteria have been clearly documented and included in the supplementary materials, reinforcing transparency and enabling reproducibility for future studies in the growing field of low-power ML-based micronutrient sensing platforms. 2.10. Certainty assessment The reviewed literature was evaluated referring to the five quality assessment (QA) criteria to ensure rigor and relevance: QA1: The clear research aim QA2: transparent data collection QA3: Clear ML processes QA4:proper and appropriate research methodology. QA5: contributions of ML to microcontrollers-class hardware. The certainty assessment is based on the scale of zero (0) to one (1). (0) point is given to a response that is not assigned, (0.5) is given when the criteria is partially met and (1) is given when the total score is met or a ‘yes’, it is assigned. The five criterias were using this lamina. Each piece of literature under review can receive a total score between 0 and 5 points.This results are for the collected literature on the applications and performance benefits of machine learning in microcontroller-class hardware are presented in Table 5 . Table 5 Certainty Assessment Results for Collected Literature on Data Mining and Business Intelligence in SMEs. Ref. QA1 QA2 QA3 QA4 QA5 Total % grading Pereira, 2022; Pereira et al., 2024 ; M EL Adoui, 2024 1 1 1 1 1 5 100 Domènech-Gil et al., 2024 1 1 1 1 0.5 4.5 90 Dhal et al., 2022 ; Madson, 2022; Dalmeida & Kamath, 2023 1 0.5 1 1 0.5 4 80 Dalmeida & Kamath, 2024; Salve & Jemila, 2024 1 0.5 0.5 1 0.5 3.5 70 Venkatesh & Naik, 2023 0.5 0.5 0.5 0.5 1 3 60 To support the conclusions of this systematic review on the implementation and practical benefits of machine learning on microcontroller-class hardware for micronutrient sensing, we conducted a thorough evaluation of the reliability of the evidence.The validity and robustness of our conclusions rely on a structured appraisal process, which was implemented using the Grading of Recommendations, Assessment, Development, and Evaluations (GRADE) framework. This framework is an internationally acknowledged method that provides a thorough and clear strategy for evaluating the strength of evidence, ensuring that the resulting conclusions are reliable and substantiated. The reliability of the evidence related to major findings was carefully assessed through a set of essential criteria. First, we closely examined the precision of effect estimates by considering sample sizes and the width of confidence intervals in the studies. Tight uncertainty margins alongside substantial sample counts signified a strong degree of confidence in the findings, as they reflect more consistent and accurate outcome estimations. We further evaluated the uniformity of outcomes by analyzing patterns across the reviewed studies. A strong alignment in reported effects across multiple sources strengthened the overall confidence in the evidence. Any identified variability among the studies was carefully examined to determine its origins and assess how it might influence the overall conclusions (Khanyi et al., 2024 ; Thobejane et al., 2024; Skosana et al, 2024 ; Mkhize et al., 2025 ). Results 3.1. Study selection In this systematic review, the study selection process adhered to a rigorous and methodical approach to ensure the inclusion of relevant and high-quality research articles. Three prominent online databases—Google Scholar, Scopus, and Web of Science—were systematically searched to identify studies that met the predefined inclusion criteria. Specifically, the search retrieved 10,000 records from Google Scholar, 1,577 records from Scopus, and 156 records from Web of Science, resulting in a total of 11,713 initial records. Following this, duplicate entries were removed, leaving 4,703 unique records. These records underwent a preliminary screening process, during which their titles and abstracts were assessed for relevance to the review’s inclusion criteria. This step resulted in the selection of 50 full-text documents for a more detailed evaluation. After a comprehensive review of these full-text documents, 43 studies were deemed eligible for inclusion in the final systematic review. The included studies consisted of 22 journal articles, 16 conference papers, 2 book chapters, and 3 dissertations or theses. The overall study collection process is depicted in Fig. 6 , which provides an in-depth PRISMA flowchart outlining the progression of records through each stage of the review. 3.2. Study results The trend in annual publication output reveals a steady increase in research interest from 2019, peaking in 2022 with the highest number of publications (n = 12). This growth reflects rising attention toward embedded ML for environmental sensing. However, a noticeable decline follows in 2023 and beyond, possibly due to saturation in foundational studies or shifts toward implementation-focused research. Figure 8 illustrates the geographical distribution of the reviewed studies, highlighting the global interest in ML-based micronutrient sensing on embedded platforms. Israel and Kenya lead with the highest contribution (12% each), followed by Switzerland (7%) and several other countries—including the USA, Brazil, Finland, and India—each contributing 5%. This spread demonstrates a diverse research base, with both high-income and developing countries engaging in innovations aimed at low-cost, real-time nutrient sensing. As shown in Fig. 9 , journal articles constitute the majority of included studies (51.16%), followed by conference papers (37.21%), dissertations or theses (6.98%), and book Chaps. (4.65%). This distribution suggests that peer-reviewed journals and conferences are the primary mediums for disseminating research on embedded ML for micronutrient sensing. The presence of theses and book chapters also indicates emerging academic interest and growing educational integration in this field. The majority of the studies included in this review were retrieved from Google Scholar (46.51%), followed by Scopus (32.56%) and Web of Science (20.93%), as illustrated in Fig. 10 . This distribution highlights Google Scholar’s broader indexing capacity, particularly for conference papers, theses, and grey literature relevant to emerging research fields like embedded machine learning for micronutrient sensing. Figure 11 shows that most studies (39.53%) relied on established ML or TinyML frameworks and platforms, demonstrating a preference for well-supported, standardized tools. Custom or proprietary solutions were also significant (23.26%), reflecting adaptation to specific hardware constraints or sensing tasks. Interestingly, 13.95% of studies did not specify their framework, pointing to reporting gaps. Other approaches included hybrid methods (9.30%), niche research tools (6.98%), and a smaller proportion using commercial (MicroAI) or multi-domain systems (2.33%). This diversity indicates an evolving landscape of framework adoption tailored to embedded micronutrient sensing. As shown in Fig. 12 , 32.56% of studies clearly stated the exact microcontroller models or boards used, indicating a commitment to replicability and technical precision. A slightly smaller portion (30.23%) used multiple or hybrid hardware platforms, showcasing versatility across different configurations. Meanwhile, 20.93% of papers referred only to broader MCU families (e.g., ARM Cortex), and 16.28% did not specify hardware details at all. These gaps in reporting may hinder reproducibility and limit comparative performance analysis. Figure 13 presents the distribution of research focus areas, with the majority of studies (53.49%) centering on core ML tasks such as classification and regression, underscoring their foundational role in embedded sensing applications. Other key domains include human-centered interaction (11.63%), computer vision (9.30%), and industrial or infrastructure monitoring (6.98%). Agriculture and environmental monitoring, directly relevant to micronutrient sensing, accounted for 4.65% of studies. A small portion also explored learning paradigms (6.98%), multi-domain integration (2.33%), and sentiment or language-based analytics (2.33%), reflecting early-stage diversification of use cases. As shown in Fig. 14 , structured deep learning models such as CNNs, RNNs, and LSTMs were used in 37.21% of the studies, indicating strong reliance on advanced architectures for complex sensing tasks. Ensemble and hybrid methods followed at 20.93%, reflecting the value of combining multiple algorithms to enhance performance under hardware constraints. Classical ML models (16.28%) remained relevant, particularly for simpler tasks requiring low computational resources. Notably, 9.30% of papers referenced deep learning without specifying model types, while others explored specialized neural networks (2.33%) or general frameworks (4.65%), revealing ongoing variation in modeling practices. According to Fig. 15 , custom datasets—curated or generated for specific use cases—were used in 39.53% of the studies, reflecting the niche and application-specific nature of micronutrient sensing. Open-source and benchmark datasets were employed in 23.26% of papers, offering standardized comparability. However, 18.60% of studies did not specify dataset type, revealing a documentation gap. Additional dataset categories included general or mixed datasets (11.63%), domain-specific time-series data (4.65%), and real-time environmental sensor data (2.33%), underscoring the diversity and evolving nature of data collection approaches in embedded ML research. Figure 16 highlights a critical limitation in the literature: 46.51% of studies did not specify model size or latency metrics, hindering reproducibility and performance benchmarking. While 23.26% of studies noted using optimized or quantized models, they lacked exact specifications. Only 11.63% explicitly reported both model size and latency, offering precise performance data. The remainder used general (13.95%) or performance-focused (2.33%) descriptions, or made timing-only remarks (2.33%). These findings underscore the need for improved transparency and standardization in embedded ML reporting. As seen in Fig. 17 , 30.23% of the reviewed studies referenced general hardware constraints, typical of TinyML applications, while 25.58% mentioned device-specific limitations without providing numerical specifications. Only 4.65% explicitly reported low-memory environments (limited RAM and flash), and just 9.30% offered detailed configurations like PSRAM usage, revealing a scarcity of advanced embedded system descriptions. Notably, 16.28% of studies did not mention hardware limitations at all, and 2.33% addressed ultra-low SRAM scenarios. These results point to the need for greater consistency and technical depth in documenting edge device limitations. Figure 18 reveals that 65.12% of studies confirmed real-time performance, demonstrating a strong focus on achieving low-latency inference in embedded environments. A smaller proportion (11.63%) reported partial or conditional real-time capability, while 9.30% did not specify this metric at all. Other papers emphasized integration contexts (4.65%), real-time classification requirements (2.33%), educational demonstrations (2.33%), STM32-based real-time designs (2.33%), and real-time optimization via model tuning (2.33%). These findings highlight the increasing emphasis on verified real-time functionality in ML systems for edge applications. As depicted in Fig. 19 , the dominant application area was Smart IoT and embedded systems (25.58%), underscoring the importance of seamless hardware-software integration in edge AI solutions. Environmental monitoring and agriculture, directly linked to micronutrient sensing, accounted for 13.95% of studies. Other significant areas included visual processing (13.95%), human-centric applications (11.63%), and advanced ML deployment on edge platforms (11.63%). Smaller contributions came from industrial IoT (6.98%), audio/multimodal sensing (6.98%), and lifelong learning (2.33%). A notable 9.30% of studies did not specify an application area, reflecting some inconsistencies in scope definition. Figure 20 shows that 25.58% of studies used edge ML-specific tools and frameworks, emphasizing the growing adoption of domain-optimized solutions tailored for constrained environments. Custom or proprietary tools followed closely (23.26%), reflecting the need for specialized implementations. General or mixed tools were used in 18.60% of papers, while standard embedded development tools—typical of classical MCU programming—appeared in 16.28%. Notably, 16.28% of studies did not specify the tools used, signaling a need for improved reporting practices in embedded ML research. 3.3. Study research questions results Which types of machine learning algorithms are most frequently recommended and successfully implemented on microcontroller-class hardware for micronutrient detection? The review revealed that structured deep learning models, especially Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory networks (LSTMs), were the most frequently employed algorithms, used in 37.21% of studies (Fig. 14 ). These models were often adapted for embedded inference using quantization or pruning techniques to reduce computational load. Ensemble and hybrid methods, such as combinations of Random Forests and neural networks, accounted for 20.93%, while classical machine learning models like decision trees and Support Vector Machines (SVMs) were applied in 16.28% of the literature. Their adoption was primarily driven by their low memory requirements and faster inference times, making them suitable for real-time applications on devices with limited RAM and processing capabilities. However, despite this diversity in algorithm selection, nearly 10% of studies referenced deep learning without specifying the exact model type, indicating a need for clearer documentation. What types of sensors are most suitable and commonly used in micronutrient sensing systems based on microcontrollers? While not all reviewed studies explicitly documented the types of sensors used, patterns emerged suggesting that electrochemical sensors are the most suitable for integration with microcontroller-based platforms due to their low power requirements, compact form factors, and high sensitivity to trace nutrient elements. These sensors were particularly relevant in applications involving soil or water quality analysis. In image-based studies—particularly those involving CNNs—smartphone cameras and optical sensors were used to infer micronutrient levels from plant leaves or soil surfaces. However, a significant number of studies (around 20%) did not clearly report the sensing mechanism or sensor model, which limits reproducibility and cross-study comparisons. This reporting gap highlights the need for standardized sensor documentation in embedded ML research. What are the typical performance outcomes (e.g., accuracy, latency, energy efficiency) observed when deploying ML models for micronutrient sensing on constrained devices? Performance evaluation in the reviewed studies commonly focused on real-time capability and energy efficiency, though quantitative performance metrics were inconsistently reported. Approximately 65.12% of studies claimed real-time processing capabilities (Fig. 18 ), and 23.26% indicated the use of optimized or quantized models to reduce latency and improve battery life (Fig. 16 ). However, only 11.63% of studies provided explicit measurements for both model size and latency, while 46.51% did not report these metrics at all. This lack of transparency makes it difficult to assess scalability and system viability under field conditions. Despite this, many models achieved competitive accuracy, particularly when custom datasets were used, though validation against real-world conditions was often limited. Energy efficiency was discussed qualitatively, with several studies emphasizing low-power deployment on platforms like the ESP32 and STM32, but only a few included concrete power consumption benchmarks. How do existing studies address key challenges such as limited memory, real-time processing demands, and power consumption in microcontroller-class environments? The majority of studies acknowledged the fundamental hardware constraints of microcontroller-class devices. Approximately 32.56% specified exact microcontroller boards—such as ARM Cortex-M, ESP32, or STM32—while 30.23% described general resource limitations associated with TinyML environments (Fig. 12 ). Commonly cited challenges included low RAM availability (often under 256 KB), limited flash storage, and the need for fast inference without cloud dependency. However, only 4.65% of studies explicitly quantified memory or processing limits, and 16.28% failed to mention hardware constraints altogether (Fig. 17 ). To overcome these challenges, some studies employed model compression techniques like quantization, and others used tailored frameworks such as TensorFlow Lite for Microcontrollers. Nevertheless, a lack of uniformity in hardware reporting and optimization strategy limits the transferability of these solutions across different deployments. To what extent are ML-enabled microcontroller systems being deployed or validated in real-world applications, including agricultural, environmental, and low-resource settings? Several studies demonstrated partial or full real-world applicability of embedded ML systems, particularly in the domains of smart IoT (25.58%), environmental monitoring (13.95%), and agriculture (13.95%) (Fig. 19 ). Examples include ML-powered soil nutrient analyzers deployed in field trials, and water potability monitoring systems integrated into remote sensor nodes. Although many studies used simulated or lab-based environments to validate system performance, the increasing use of edge-capable frameworks and power-efficient microcontrollers shows strong momentum toward practical deployment. Real-time data processing, offline functionality, and remote monitoring capabilities were identified as core features enhancing suitability for low-resource regions. Nonetheless, scalability and maintenance under long-term deployment conditions were less frequently addressed, indicating an area for future research. Which micronutrients are being targeted in ML-based sensing applications on microcontroller-class hardware, and how consistently are these reported across studies? Targeted micronutrients in the reviewed studies included zinc, iron, nitrogen, phosphorus, and potassium, typically detected through soil or water-based sensing systems (Figs. 13 and 15 ). In some cases, micronutrient deficiencies were inferred through plant leaf image analysis, particularly for nitrogen and phosphorus. However, reporting consistency was low: while a few studies focused specifically on nutrient detection, others addressed more general soil health or environmental monitoring without detailing which micronutrients were measured. Only a subset of studies provided ground-truth chemical validation of sensing results, and many lacked domain-specific context. This inconsistency signals the need for more rigorous documentation of nutrient-specific sensing objectives, datasets, and validation methods in future work. Conclusion This systematic review provides a comprehensive evaluation of machine learning (ML) applications on microcontroller-class hardware for micronutrient sensing across environmental, agricultural, and embedded contexts. The findings reveal a growing global interest, with a peak in research output around 2022, and increasing contributions from both high-income and resource-constrained regions. Despite promising developments in deploying deep learning, ensemble models, and TinyML frameworks on platforms such as STM32, ESP32, and ARM Cortex-M, the review highlights persistent limitations. These include inconsistent hardware specification, limited use of standard datasets, and a widespread lack of transparency regarding model size, latency, and power consumption. While 65.12% of studies confirmed real-time performance, only a fraction provided quantified performance metrics, which limits reproducibility and benchmarking across studies. Custom datasets and domain-specific tools dominated, reflecting the specialized nature of micronutrient sensing tasks. However, this also introduces challenges for generalization and scalability. Application areas such as environmental monitoring, smart IoT, and health-focused sensing demonstrate the versatility of ML-embedded microcontroller systems, yet underscore the need for robust, field-tested implementations. To fully harness the potential of ML on microcontroller-class devices for micronutrient sensing, future research should focus on: Establishing standardized benchmarks and reporting practices, Emphasizing low-power, real-time performance in real-world conditions, Enhancing interdisciplinary collaboration among hardware engineers, data scientists, and domain experts. References Pereira, E. A. M. (2024). Water potability classification: an approach using machine learning in an embedded system. Dhal, S. B., Jungbluth, K., Lin, R., Sabahi, S. P., Bagavathiannan, M., Braga-Neto, U., & Kalafatis, S. (2022). A machine-learning-based IoT system for optimizing nutrient supply in commercial aquaponic operations. Sensors, 22(9), 3510. Pereira, E. A. M., da Silva Santos, J. F., & de Andrade Barboza, E. (2024). An energy efficient tinyml model for a water potability classification problem. Sustainable Computing: Informatics and Systems, 43, 101010. GASANA, J. M. (2022). IoT based monitoring system of soil properties (Doctoral dissertation, College of science and Technology). Schizas, N., Karras, A., Karras, C., & Sioutas, S. (2022). TinyML for ultra-low power AI and large scale IoT deployments: A systematic review. Future Internet, 14(12), 363. Dalmeida, S. P., & Kamath, S. (2023). Soil micronutrient detection using machine learning. In Recent Trends in Computational Sciences (pp. 29-35). CRC Press. Salve, P. R., Jemila, M. M., Sridharan, S., Meti, V. K. V., & Roy, B. (2025). High-Performance Computer-Integrated Microcontrollers for Enhancing Food Quality Monitoring: Innovations and Applications. In Integrating Machine Learning Into HPC-Based Simulations and Analytics (pp. 355-382). IGI Global Scientific Publishing. https://www.igi-global.com/chapter/high-performance-computer-integrated-microcontrollers-for-enhancing-food-quality-monitoring/365689 Venkatesh, K., & Naik, K. J. (2023, July). Deep learning for macro-nutrient deficiency identification in the groundnut plants. In IET Conference Proceedings CP832 (Vol. 2023, No. 5, pp. 193-198). Stevenage, UK: The Institution of Engineering and Technology.https://digital-library.theiet.org/doi/abs/10.1049/icp.2023.1489 El Adoui, M., Herpoel, T., & Frénay, B. (2024). Constrained Tiny Machine Learning for Predicting Gas Concentration with I4. 0 Low-cost Sensors. ACM Transactions on Embedded Computing Systems, 23(3), 1-23.https://dl.acm.org/doi/abs/10.1145/3590956 Kiplimo, E., Riddick, S. N., Mbua, M., Upreti, A., Anand, A., & Zimmerle, D. J. (2024). Addressing low-cost methane sensor calibration shortcomings with machine learning. Atmosphere, 15(11), 1313. https://www.mdpi.com/2073-4433/15/11/1313 Ünsalan, C., Höke, B., & Atmaca, E. (2024). Embedded Machine Learning with Microcontrollers. Alenezi, M., Alshammari, R., & Alrashed, B. (2019). Enhancing Project Scope Management in Construction Projects Using Building Information Modeling (BIM). 2019 IEEE 2nd International Conference on Information and Computer Technologies (ICICT), 133–137. https://doi.org/10.1109/INFOCT.2019.8987754 Jain, N., Awasthi, Y., & Jain, R. K. (2023). An IoT-based soil analysis system using optical sensors and multivariate regression. Int. J. Exp. Res. Rev, 31, 23-32. https://www.academia.edu/download/104796158/3_An_IoT_based_soil_analysis_system_using_optical _sensors_and_multivariate_regression.pdf Grau, M. M., Centelles, R. P., & Freitag, F. (2021, September). On-device training of machine learning models on microcontrollers with a look at federated learning. In Proceedings of the Conference on Information Technology for Social Good (pp. 198-203). https://dl.acm.org/doi/abs/10.1145/3462203.3475896 Nguyen, V. K., Tran, V. K., Pham, H., Nguyen, V. M., Nguyen, H. D., & Nguyen, C. N. (2023). A multi-microcontroller-based hardware for deploying Tiny machine learning model. International Journal of Electrical and Computer Engineering (IJECE), 13(5), 5727-5736.https://www.academia.edu/download/113156003/16869.pdf Elsts, A., & McConville, R. (2021). Are microcontrollers ready for deep learning-based human activity recognition?. Electronics, 10(21), 2640. https://www.mdpi.com/2079-9292/10/21/2640. Myataza, A., Mafunga, M., Mkhulisi, N. S., & Thango, B. A. A Systematic Review of ERP, CRM, and HRM Systems for SMEs: Managerial and Employee Support. Preprints 2024, 2024100384. Gumede, T. T., Chiworeka, J. M., Magoda, A. S., & Thango, B. Building Effective Social Media Strategies for Business: A Systematic Review. Preprints 2024, 2024100379. Vidal, J. (2025). Scalable Embedded Tiny Machine Learning (SETML): A General Framework for Embedded Distributed Inference (Doctoral dissertation, Massachusetts Institute of Technology). https://dspace.mit.edu/handle/1721.1/159107 Plancher, B., Buttrich, S., Ellis, J., Goveas, N., Kazimierski, L., Sotelo, J. L., ... & Reddi, V. J. (2024, May). TinyML4D: Scaling Embedded Machine Learning Education in the Developing World. In Proceedings of the AAAI symposium series (Vol. 3, No. 1, pp. 508-515). http://ojs.aaai.org/index.php/AAAI-SS/article/view/31265 Dhal, S. B., Jungbluth, K., Lin, R., Sabahi, S. P., Bagavathiannan, M., Braga-Neto, U., & Kalafatis, S. (2022). A machine-learning-based IoT system for optimizing nutrient supply in commercial aquaponic operations. Sensors, 22(9), 3510. https://www.mdpi.com/1424-8220/22/9/3510 Boussion, N., Bosc, L., Goasduff, G., Jegat, Y., Lann, L., Lucia, A. S., ... & Visvikis, D. To calibrate or not to calibrate: a decision support approach using microcontroller-based tiny machine learning. https://www.iccr2024.org/papers/525059.pdf Zhou, X., Kang, Z., Canady, R., Bao, S., Balasubramanian, D. A., & Gokhale, A. (2021, November). Exploring cloud assisted tiny machine learning application patterns for PHM scenarios. In Annual Conference of the PHM Society (Vol. 13, No. 1). http://www.papers.phmsociety.org/index.php/phmconf/article/view/3054 Simitha, K. M., & Raj, S. (2019, June). IoT and WSN based water quality monitoring system. In 2019 3rd International conference on Electronics, Communication and Aerospace Technology (ICECA) (pp. 205-210). IEEE. Lin, J., Chen, W. M., Lin, Y., Gan, C., & Han, S. (2020). Mcunet: Tiny deep learning on iot devices. Advances in neural information processing systems, 33, 11711-11722. Mudau, M. C., Moshapo, L. W., Monyela, T. M., & Thango, B. A. The Role of Manufacturing Operations in SMEs Performance: A Systematic Review. Preprints 2024, 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. Sakr, F., Bellotti, F., Berta, R., & De Gloria, A. (2020). Machine learning on mainstream microcontrollers. Sensors, 20(9), 2638. Wang, X., Magno, M., Cavigelli, L., & Benini, L. (2020). FANN-on-MCU: An open-source toolkit for energy-efficient neural network inference at the edge of the Internet of Things. IEEE Internet of Things Journal, 7(5), 4403-4417. Moosmann, J., Giordano, M., Vogt, C., & Magno, M. (2023, June). Tinyissimoyolo: A quantized, low-memory footprint, tinyml object detection network for low power microcontrollers. In 2023 IEEE 5th International Conference on Artificial Intelligence Circuits and Systems (AICAS) (pp. 1-5). IEEE. Banbury, C., Zhou, C., Fedorov, I., Matas, R., Thakker, U., Gope, D., ... & Whatmough, P. (2021). Micronets: Neural network architectures for deploying tinyml applications on commodity microcontrollers. Proceedings of machine learning and systems, 3, 517-532. Immonen, R., & Hämäläinen, T. (2022). Tiny Machine Learning for Resource‐Constrained Microcontrollers. Journal of Sensors, 2022(1), 7437023. Saha, S. S., Sandha, S. S., & Srivastava, M. (2022). Machine learning for microcontroller-class hardware: A review. IEEE Sensors Journal, 22(22), 21362-21390. Schizas, N., Karras, A., Karras, C., & Sioutas, S. (2022). TinyML for ultra-low power AI and large scale IoT deployments: A systematic review. Future Internet, 14(12), 363. Sailesh, M., & Selvakumar, K. (2022). A novel framework for deployment of CNN models using post-training quantization on microcontrollers. Microprocessors and Microsystems, 94, 104634. Ravaglia, L., Rusci, M., Nadalini, D., Capotondi, A., Conti, F., & Benini, L. (2021). A tinyml platform for on-device continual learning with quantized latent replays. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 11(4), 789-802. Kopparapu, K., & Lin, E. (2021). TinyFedTL: Federated transfer learning on tiny devices. arXiv preprint arXiv:2110.01107. Khanyi, M. B., Xaba, S. N., Mlotshwa, N. A., 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 Thobejane, L. T., & Thango, B. A. (2024). Partial Discharge Source Classification in Power Transformers: A Systematic Literature Review. Applied Sciences, 14(14), 6097. Alajlan, N. N., & Ibrahim, D. M. (2022). TinyML: Enabling of inference deep learning models on ultra-low-power IoT edge devices for AI applications. Micromachines, 13(6), 851. Wang, T., Fang, K., Wei, W., Tian, J., Pan, Y., & Li, J. (2022). Microcontroller unit chip temperature fingerprint informed machine learning for IIoT intrusion detection. IEEE Transactions on Industrial Informatics, 19(2), 2219-2227. Azevedo, M. B., de Medeiros, T. D. A., Medeiros, M. D. A., Silva, I., & Costa, D. G. (2023). Detecting face masks through embedded machine learning algorithms: A transfer learning approach for affordable microcontrollers. Machine Learning with Applications, 14, 100498. Lin, J. (2021). Efficient Algorithms and Systems for Tiny Deep Learning (Doctoral dissertation, Massachusetts Institute of Technology). Elsts, A., & McConville, R. (2021). Are microcontrollers ready for deep learning-based human activity recognition?. Electronics, 10(21), 2640. Exell, N. (2023). Applicability of TinyML for maintenance predictability.Verbiest, J. R., Bonnechère, B., Saeys, W., Van de Walle, P., Truijen, S., & Meyns, P. (2023). Gait stride length estimation using embedded machine learning. Sensors, 23(16), 7166. Park, J. (2022). Hybrid Machine Learning and Physics-Based Modeling Approaches for Process Control and Optimization (Doctoral dissertation, Brigham Young University). Bellarmino, N., Cantoro, R., Huch, M., Kilian, T., Martone, R., Schlichtmann, U., & Squillero, G. (2023). A multilabel active learning framework for microcontroller performance screening. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 42(10), 3436-3449. Kwon, Y. D., Chauhan, J., & Mascolo, C. (2022, May). Yono: Modeling multiple heterogeneous neural networks on microcontrollers. In 2022 21st ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) (pp. 285-297). IEEE. Skosana, S., Mlambo, S., Madiope, T., & Thango, B. Evaluating Wireless Network Technologies (3G, 4G, 5G) and Their Infrastructure: A Systematic Review. Preprints 2024, 2024101331. Mkhize, A., Mokhothu, K. D., Tshikhotho, M., & Thango, B. A. (2025). Evaluating the Impact of Cloud Computing on SME Performance: A Systematic Review. Businesses, 5(2), 23. https://doi.org/10.3390/businesses5020023 Buonanno, L., Di Vita, D., Carminati, M., & Fiorini, C. (2020). A directional gamma-ray spectrometer with microcontroller-embedded machine learning. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 10(4), 433-443. Manor, E., & Greenberg, S. (2022). Custom hardware inference accelerator for tensorflow lite for microcontrollers. IEEE Access, 10, 73484-73493. Silva, L. C., Simas Filho, E. F., Albuquerque, M. C., Silva, I. C., & Farias, C. T. (2021). Embedded decision support system for ultrasound nondestructive evaluation based on extreme learning machines. Computers & Electrical Engineering, 90, 106891. Ha, N., Xu, K., Ren, G., Mitchell, A., & Ou, J. Z. (2020). Machine learning‐enabled smart sensor systems. Advanced Intelligent Systems, 2(9), 2000063. Muszynski, M., Gabrieli, G., Zimmerli, L., Temiz, Y., Heller, R., Cox, A., ... & Ruch, P. W. (2022, October). Live Demonstration: An AI-assisted e-tongue for fast and portable fingerprinting of liquids. In 2022 IEEE Sensors (pp. 1-1). IEEE. Dujić Rodić, L., Stančić, I., Zovko, K., Perković, T., & Šolić, P. (2022). Tag estimation method for aloha rfid system based on machine learning classifiers. Electronics, 11(16), 2605. Waskito, S., Leow, K. J., Medaranga, P., Gupta, T., Chakrabarty, S., Gulati, M., & Varshney, A. (2023, October). Otter: Simplifying embedded sensor data collection and analysis using large language models. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking (pp. 1-3). Dokic, K. (2020). Microcontrollers on the edge–is esp32 with a camera ready for machine learning?. In Image and Signal Processing: 9th International Conference, ICISP 2020, Marrakesh, Morocco, June 4–6, 2020, Proceedings 9 (pp. 213-220). Springer International Publishing. Pham, H. T., Nguyen, M. A., & Sun, C. C. (2019, December). AIoT solution survey and comparison in machine learning on low-cost microcontrollers. In 2019 International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS) (pp. 1-2). IEEE. 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-6844555","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Systematic Review","associatedPublications":[],"authors":[{"id":468006210,"identity":"e6c257e1-a954-4300-8f05-b611767fbde9","order_by":0,"name":"Dineo Moeketsi","email":"data:image/png;base64,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","orcid":"","institution":"University of Johannesburg","correspondingAuthor":true,"prefix":"","firstName":"Dineo","middleName":"","lastName":"Moeketsi","suffix":""},{"id":468006211,"identity":"765b8a4a-fa40-4b62-86b2-6a50a609c566","order_by":1,"name":"Alecia Mkhantshwa","email":"","orcid":"","institution":"University of Johannesburg","correspondingAuthor":false,"prefix":"","firstName":"Alecia","middleName":"","lastName":"Mkhantshwa","suffix":""},{"id":468006212,"identity":"3b6afc19-a26b-4caf-9af4-42061448fe91","order_by":2,"name":"Calvin Modise","email":"","orcid":"","institution":"University of Johannesburg","correspondingAuthor":false,"prefix":"","firstName":"Calvin","middleName":"","lastName":"Modise","suffix":""},{"id":468006213,"identity":"d43b5f23-cbf0-43b3-8a49-1a200af097ca","order_by":3,"name":"Ntokozo Mlangeni","email":"","orcid":"","institution":"University of Johannesburg","correspondingAuthor":false,"prefix":"","firstName":"Ntokozo","middleName":"","lastName":"Mlangeni","suffix":""}],"badges":[],"createdAt":"2025-06-07 20:18:12","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":true,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":true},"doi":"10.21203/rs.3.rs-6844555/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6844555/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":84423154,"identity":"33256eb7-6e4c-4d62-9e35-b38c0bc0c4d1","added_by":"auto","created_at":"2025-06-11 18:41:50","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":42607,"visible":true,"origin":"","legend":"\u003cp\u003ekeywords used in finding articles for machine learning in microcontrollers.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/14811552131d4f0045f5a127.png"},{"id":84424164,"identity":"76670f06-79da-4c0f-912e-955f8936de72","added_by":"auto","created_at":"2025-06-11 18:57:50","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":185564,"visible":true,"origin":"","legend":"\u003cp\u003estages in collecting studies.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/58798e99d3e48d719bd2cf2d.png"},{"id":84423640,"identity":"18ea68be-5ec7-49aa-ad23-07761fe9f9fc","added_by":"auto","created_at":"2025-06-11 18:49:50","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":276418,"visible":true,"origin":"","legend":"\u003cp\u003eData gathering and selection process.\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/ca57ecfaa0be666bb9fe77fd.png"},{"id":84423153,"identity":"cf1863ea-8cdb-4769-9ca2-dc4e4810ee3c","added_by":"auto","created_at":"2025-06-11 18:41:50","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":152173,"visible":true,"origin":"","legend":"\u003cp\u003eRisk of Bias Assessment Process for Non-Randomized Studies.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/dd9165f643779254cf26d2f9.png"},{"id":84423638,"identity":"1b25b97d-e15b-4cf5-80f3-54bda30fdfae","added_by":"auto","created_at":"2025-06-11 18:49:50","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":268522,"visible":true,"origin":"","legend":"\u003cp\u003eSystematic review process on machine learning on microcontrollers.\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/5d1ac6b5e2b3d7253813d72a.png"},{"id":84423639,"identity":"d1f51b61-d04b-4ece-97b7-64342a34b2db","added_by":"auto","created_at":"2025-06-11 18:49:50","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":72056,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eProposed \u003c/strong\u003ePRISMA Flowchart.\u003c/p\u003e","description":"","filename":"6.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/50044a0960d2d410768894bb.png"},{"id":84424347,"identity":"0118af79-f7ad-4f44-a1d7-26423917e93e","added_by":"auto","created_at":"2025-06-11 19:05:50","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":56308,"visible":true,"origin":"","legend":"\u003cp\u003eAnnual distribution of published studies related to machine learning on microcontroller-class hardware for micronutrient sensing(2019–2025).\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/c698be516dde7db046d8d740.png"},{"id":84423162,"identity":"179a2787-4ecd-4da8-9c1d-3cb6879b0139","added_by":"auto","created_at":"2025-06-11 18:41:50","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":171203,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGeographical\u003c/strong\u003e \u003cstrong\u003edistribution \u003c/strong\u003eof studies on ML-based micronutrient sensing using microcontroller-class hardware (2015–2025).\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/13f2795a8a397dcfd9fed8c1.png"},{"id":84424165,"identity":"07ae173a-98f4-4809-86b4-b7913ce789d9","added_by":"auto","created_at":"2025-06-11 18:57:50","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":61331,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of \u003cstrong\u003epublication \u003c/strong\u003etypes among selected studies on ML applications in micronutrient sensing using microcontroller-class hardware\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/a9b079b38ef16320e7262cba.png"},{"id":84423157,"identity":"3a8c9945-2c1a-41a7-93d7-aac399c8b611","added_by":"auto","created_at":"2025-06-11 18:41:50","extension":"png","order_by":10,"title":"Figure 10","display":"","copyAsset":false,"role":"figure","size":76041,"visible":true,"origin":"","legend":"\u003cp\u003eProportion of selected studies sourced from major academic databases: Google Scholar, Scopus, and Web of Science.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/b2597ebc5fce9ae3a9697e9d.png"},{"id":84423187,"identity":"008bf3b0-0dc0-436e-8a63-85857a40e824","added_by":"auto","created_at":"2025-06-11 18:41:51","extension":"png","order_by":11,"title":"Figure 11","display":"","copyAsset":false,"role":"figure","size":183652,"visible":true,"origin":"","legend":"\u003cp\u003eDistribution of ML framework types used in micronutrient sensing on microcontroller-class hardware, categorized by tool origin and methodology.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/d7e6986fde5b0958ccb6503d.png"},{"id":84423653,"identity":"c3ca534f-dccd-4728-93f2-9ce078cd8d27","added_by":"auto","created_at":"2025-06-11 18:49:51","extension":"png","order_by":12,"title":"Figure 12","display":"","copyAsset":false,"role":"figure","size":109879,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHardware \u003c/strong\u003especification levels reported in studies applying ML on microcontroller-class devices for micronutrient sensing.\u003c/p\u003e","description":"","filename":"floatimage8.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/82f717de497db84d820a6eb8.png"},{"id":84423185,"identity":"092d7667-d2fe-47ca-820e-0fd8fb16b0ad","added_by":"auto","created_at":"2025-06-11 18:41:51","extension":"png","order_by":13,"title":"Figure 13","display":"","copyAsset":false,"role":"figure","size":211401,"visible":true,"origin":"","legend":"\u003cp\u003eProblem domains \u003cstrong\u003eaddressed \u003c/strong\u003ein studies applying machine learning on microcontroller-class hardware for micronutrient and related sensing tasks.\u003c/p\u003e","description":"","filename":"floatimage9.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/904f0507c8fe54c840f7f0f4.png"},{"id":84423167,"identity":"93b1c3a3-8896-4587-9085-dc2eb1485a6d","added_by":"auto","created_at":"2025-06-11 18:41:50","extension":"png","order_by":14,"title":"Figure 14","display":"","copyAsset":false,"role":"figure","size":157994,"visible":true,"origin":"","legend":"\u003cp\u003eTypes of machine \u003cstrong\u003elearning \u003c/strong\u003emodels employed in studies using microcontroller-class hardware for micronutrient sensing.\u003c/p\u003e","description":"","filename":"floatimage10.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/b3dde102a56385a544697e17.png"},{"id":84423649,"identity":"f3ac0729-33fb-4dad-9ca5-4ce02c0c4d51","added_by":"auto","created_at":"2025-06-11 18:49:51","extension":"png","order_by":15,"title":"Figure 15","display":"","copyAsset":false,"role":"figure","size":158348,"visible":true,"origin":"","legend":"\u003cp\u003eTypes of datasets used in machine learning studies for micronutrient sensing on microcontroller-class hardware.\u003c/p\u003e","description":"","filename":"floatimage11.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/d0332162938c5f6d378bcea8.png"},{"id":84424167,"identity":"a431d5d3-df7d-4727-b027-f689e76a5759","added_by":"auto","created_at":"2025-06-11 18:57:51","extension":"png","order_by":16,"title":"Figure 16","display":"","copyAsset":false,"role":"figure","size":138274,"visible":true,"origin":"","legend":"\u003cp\u003eReporting styles for model size and latency in studies involving ML on microcontroller-class hardware.\u003c/p\u003e","description":"","filename":"floatimage12.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/673c59497bef9ddf688613e1.png"},{"id":84423203,"identity":"698fcc57-0aba-48ee-aa71-88e6c8fe8944","added_by":"auto","created_at":"2025-06-11 18:41:51","extension":"png","order_by":17,"title":"Figure 17","display":"","copyAsset":false,"role":"figure","size":199398,"visible":true,"origin":"","legend":"\u003cp\u003eDescriptions of hardware limitations reported in studies applying ML on microcontroller-class devices for micronutrient \u003cstrong\u003esensing.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage13.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/212498ab2f192478e531b4be.png"},{"id":84423189,"identity":"d405266e-4ef1-426a-bfbb-bd573e966921","added_by":"auto","created_at":"2025-06-11 18:41:51","extension":"png","order_by":18,"title":"Figure 18","display":"","copyAsset":false,"role":"figure","size":162356,"visible":true,"origin":"","legend":"\u003cp\u003eReal-time performance reporting descriptors in ML studies on microcontroller-class hardware for micronutrient \u003cstrong\u003esensing.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage14.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/95a261b529c13876c01e3b12.png"},{"id":84423175,"identity":"2afd99b8-505e-4105-b598-85b33d52b529","added_by":"auto","created_at":"2025-06-11 18:41:50","extension":"png","order_by":19,"title":"Figure 19","display":"","copyAsset":false,"role":"figure","size":200243,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eApplication \u003c/strong\u003eareas of ML on microcontroller-class hardware for micronutrient and related sensing tasks.\u003c/p\u003e","description":"","filename":"floatimage15.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/8bf1b618ca31afc3f9765ccd.png"},{"id":84423195,"identity":"25650057-e4b0-4bc4-bb06-178921c5b2ff","added_by":"auto","created_at":"2025-06-11 18:41:51","extension":"png","order_by":20,"title":"Figure 20","display":"","copyAsset":false,"role":"figure","size":116884,"visible":true,"origin":"","legend":"\u003cp\u003eTypes of tools and \u003cstrong\u003eframeworks \u003c/strong\u003eused in ML deployments on microcontroller-class hardware.\u003c/p\u003e","description":"","filename":"floatimage16.png","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/e251052112b359bf3226006b.png"},{"id":84424584,"identity":"230f3ebf-0d7c-4dd3-87d5-a7692b73fb10","added_by":"auto","created_at":"2025-06-11 19:13:52","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4515669,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6844555/v1/03578196-ee84-4048-af2a-80bdbbaabd35.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003e\u003cstrong\u003eTinyML Applications in Micronutrient Sensing: A Review of Microcontroller Deployments\u003c/strong\u003e\u003c/p\u003e","fulltext":[{"header":"Introduction","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eAs modern technology continues to evolve at an unprecedented rate, we are witnessing a growing interest in how machine learning (ML) can be applied to small-scale, energy-efficient computing systems. Among these, microcontroller-class hardware\u0026mdash;compact, low-power chips commonly found in embedded systems and Internet of Things (IoT) devices\u0026mdash;has become an exciting platform for running ML algorithms (Ahmad et al., 2020). The ability to process data locally, without constant reliance on the cloud, has proven valuable for real-time decision-making, lower energy use, and enhanced privacy (Lane et al., 2015). These advancements have expanded the role of microcontrollers in various fields, including environmental monitoring, agriculture, and health diagnostics, where immediate and reliable insights are increasingly in demand (Fischer et al., 2020; Chen et al., 2022). impactful area where ML-enabled microcontrollers are showing promise is micronutrient sensing. An impactful area where ML-enabled microcontrollers are showing promise is micronutrient sensing. This refers to detecting and measuring essential nutrients like iron, zinc, or vitamin A, which are crucial for human health but often lacking in both food and soil\u0026mdash;especially in low-resource settings (Gibson, 2012). Conventional lab-based approaches to micronutrient testing are typically accurate but tend to be expensive, time-consuming, and inaccessible to communities that lack infrastructure (Aziz \u0026amp; Butt, 2019). The integration of lightweight ML models into embedded devices offers a novel solution to this problem, enabling real-time, portable, and cost-effective sensing platforms that can be used in the field with minimal training or technical support (Anwar et al., 2021). Devices based on microcontroller platforms such as Arduino, ESP32, or ARM Cortex-M chips are being developed to bridge this gap and make micronutrient analysis more accessible (Kamble \u0026amp; Kale, 2021). However, despite growing interest in this area, the implementation of ML models on microcontroller-class hardware still faces considerable obstacles. Limited processing power, memory, and battery life make it difficult to run traditional ML algorithms on these platforms. Since this field lies at the intersection of multiple disciplines\u0026mdash;including embedded hardware, nutritional science, chemistry, and artificial intelligence\u0026mdash;the Designing efficient, accurate, and reliable models that can work under such constraints is an ongoing challenge (Sze et al., 2017). Additionally, there\u0026rsquo;s a lack of comprehensive insight into how these systems are applied across different micronutrient sensing scenarios. current research landscape remains fragmented, making it harder to build consistent, scalable solutions (Fischer et al., 2020). Given the global importance of improving micronutrient monitoring, especially in areas affected by malnutrition or poor agricultural yields, a systematic overview of how ML is being implemented in these compact, embedded systems is needed. This review addresses that need by analyzing studies published over the past decade, focusing specifically on how machine learning is deployed on microcontroller-class hardware for micronutrient sensing. The review will evaluate the design approaches, deployment environments, data handling strategies, and performance outcomes of these systems to understand the broader trends, key challenges, and emerging opportunities in this domain. By bringing together research from different fields, this review also offers insights for developers, scientists, and decision-makers interested in applying embedded ML technologies to global health and nutrition challenges (Do Valle \u0026amp; Lee, 2020). Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e provides a comparative view of past review efforts alongside the specific scope of this work, underlining how it uniquely focuses on the intersection of embedded AI, sensor design, and micronutrient diagnostics. Through this systematic review, we aim to offer practical takeaways for future innovation, support interdisciplinary collaboration, and contribute to the development of more scalable and sustainable micronutrient sensing systems that can be used in real-world conditions (Patil \u0026amp; Kale, 2020).\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 Applications and Advantages of Machine Learning on Microcontroller-Class Hardware for Micronutrient Sensing\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\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 \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\u003eContribution\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePros\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\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\u003ePereira, E. A et al(2024)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eProposed an energy-efficient TinyML model using Random Forest and Neural Networks for classifying water potability in embedded systems\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReal-time responses, no need for internet, low energy use, long battery life, supports remote deployment\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLimited to electronic-sensor-accessible data, Random Forest accuracy at 0.70 may not suit all use cases, model complexity may limit real-time adaptation\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDhal, S. B et al (2022)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDeveloped a ML-based IoT system for nutrient optimization in commercial aquaponics using sensor data and feature selection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReal-world data from farms, automated nutrient control, identified key predictors, supports healthy plant and fish growth\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLimited data points, only weekly measurements, focused on lettuce and tilapia, commercial-scale generalization may be limited\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePereira, E. A. M et al (2024)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDeveloped a TinyML model for classifying water potability using only electronic sensor data on ESP32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh energy efficiency, fast inference (99.95% faster), reduced memory use (51.2% less), no need for internet or lab data\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLimited to sensors' data; chemical parameters not considered; may need calibration across diverse water sources\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGASANA, J. M. (2022).\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDeveloped an IoT-based system to monitor soil properties and predict suitable crop type using ML (decision tree classifier)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReal-time monitoring, accurate crop prediction (99%), reduces lab testing, supports data-driven farming\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRelies on internet connection (GPRS), limited to measured properties (NPK, pH, etc.), may require calibration for different regions\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSchizas, N., Karras et al (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2022\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSystematic review of TinyML's role in low-power AI and IoT deployments, highlighting frameworks, benefits, and integration with 5G/LPWAN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eSummarizes current research, promotes on-device analytics, reduces latency and cloud dependence, improves privacy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLacks implementation details, review-based (no new model or case study), practical limitations not deeply explored\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDalmeida, S. P et al(2023)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDeveloped a CNN-based system using image processing to detect soil micronutrients (Zn, Fe, Mn, Cu, pH)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh accuracy (95%), avoids lab testing, uses simple smartphone images, automates micronutrient detection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eManual image capture, limited to tested micronutrients, depends on image quality, MATLAB-based limits portability\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSalve, P. R et al(2025)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExplores use of high-performance microcontrollers and ML for real-time food quality monitoring across the supply chain\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eReal-time analysis, IoT and ML integration, improved food safety, reduces waste, predictive insights\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLimited availability (forthcoming), no implementation case study yet, may require advanced hardware and setup\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVenkatesh, K et al (2023)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDeveloped a lightweight CNN model to identify nitrogen, phosphorus, and potassium deficiencies in groundnut plants from leaf images\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh accuracy (94.64%), efficient detection, reduces need for lab testing, cost and time effective\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eLimited to groundnut plants, model performance depends on image quality, may need retraining for other crops or environments\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEl Adoui, M et al (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2024\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDeveloped a constrained TinyML model to accurately predict gas concentrations using low-cost sensors on MCUs\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh prediction accuracy (R\u0026sup2; = 0.72), low RAM (3%) and Flash (98%) usage, adaptable to real-time calibration, cost-effective\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eConstrained by microcontroller memory, limited generalization to non-tested environments, requires model porting and tuning\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eKiplimo, E et al (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2024\u003c/span\u003e)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDeveloped a low-cost sensor system combined with ML for accurate methane monitoring at atmospheric levels\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eHigh accuracy with errors in the tens of ppb, versatile for indoor/outdoor environments, robust calibration method\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRequires periodic calibration with reference equipment, potential variability in sensor lifespan and environmental effects.\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\u003eAdditionally, literature on applying machine learning (ML) to microcontroller-class hardware for micronutrient sensing. These gaps reveal limitations in existing studies and offer opportunities for future research to deepen understanding and expand practical applications in this growing field.To begin with, most research in this area tends to focus on general-purpose embedded machine learning or nutrient analysis tools, but very few studies investigate how lightweight ML models can be optimized specifically for real-time micronutrient sensing using constrained microcontroller devices. This narrow focus leaves out important considerations unique to low-power, low-memory environments\u0026mdash;such as energy efficiency, real-time inference accuracy, and cost constraints\u0026mdash;that are critical for deploying these systems in real-world field conditions, particularly in low-resource settings.\u003c/p\u003e \u003cp\u003eFurthermore, much of the literature emphasizes the technical side of hardware-software integration, while overlooking broader practical challenges like usability, maintenance, and deployment in remote or rural environments. The roles of user interaction, accessibility, and training are often left unaddressed, despite their importance in successful long-term implementation.Another significant gap lies in the lack of interdisciplinary research connecting machine learning techniques with domain-specific knowledge from agriculture, food science, or public health. While technical feasibility has been demonstrated in lab settings, few studies explore how these systems perform in real-life scenarios, where environmental factors, sensor variability, and data noise can significantly affect accuracy.\u003c/p\u003e \u003cp\u003eAdditionally, most Current findings rely on isolated or short-term experiments, limiting our understanding of how such systems behave over longer periods or across diverse deployment conditions. Finally, there is a need for more studies that evaluate the full system lifecycle\u0026mdash;from model training and deployment to maintenance and real-time feedback loops. Few existing works assess how these ML-enabled microcontroller systems can be updated or scaled efficiently, especially in settings without reliable internet connectivity or technical expertise. By addressing these research gaps, future studies can provide clearer insights into how to design, implement, and scale machine learning solutions for micronutrient sensing in ways that are accessible, cost-effective, and impactful.\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\u003eDespite notable progress in machine learning (ML), the integration of ML models into microcontroller-class hardware remains a technically complex and underexplored domain\u0026mdash;particularly for applications like micronutrient sensing in agriculture and environmental monitoring. These constrained platforms face unique limitations in computational power, memory capacity, and energy efficiency, making the selection of appropriate ML frameworks, hardware architectures, and deployment strategies critical for real-time, field-based operations. This systematic review aims to investigate how lightweight ML techniques, including TinyML and other edge-optimized models, can be effectively adapted for accurate, low-latency micronutrient detection in resource-constrained environments. In particular, the review seeks to understand the interplay between model architecture, dataset characteristics, hardware specifications, and performance metrics. To guide this investigation, the following research questions were formulated:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eWhich types of machine learning algorithms are most frequently recommended and successfully implemented on microcontroller-class hardware for micronutrient detection?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWhat types of sensors are most suitable and commonly used in micronutrient sensing systems based on microcontrollers?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWhat are the typical performance outcomes (e.g., accuracy, latency, energy efficiency) observed when deploying ML models for micronutrient sensing on constrained devices?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eHow do existing studies address key challenges such as limited memory, real-time processing demands, and power consumption in microcontroller-class environments?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTo what extent are ML-enabled microcontroller systems being deployed or validated in real-world applications, including agricultural, environmental, and low-resource settings?\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eWhich micronutrients are being targeted in ML-based sensing applications on microcontroller-class hardware, and how consistently are these reported across studies?\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e1.2. Hypotheses Development\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eBuilding upon the research questions, the following hypotheses were developed to explore the patterns, performance, and practical constraints of deploying machine learning (ML) models on microcontroller-class hardware for micronutrient sensing. These hypotheses reflect the observed trends in model selection, dataset design, hardware utilization, and application context identified in the reviewed studies.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eH1: Convolutional Neural Networks (CNNs), ensemble models, and classical algorithms like decision trees are the most commonly used ML models on microcontroller-class hardware due to their balance of performance and resource efficiency.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e(Supported by Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e14\u003c/span\u003e: CNNs/Deep Learning \u0026ndash; 37.21%; Ensembles \u0026ndash; 20.93%; Decision Trees \u0026ndash; part of 16.28% classical ML)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eH2: Electrochemical sensors are the most suitable for microcontroller-based micronutrient sensing due to their low power demands and integration ease.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e(Inferred from studies integrating water/soil-based nutrient analysis, though sensor type was underreported in ~\u0026thinsp;20% of studies)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eH3: Model complexity is often reduced (e.g., via quantization) to enhance energy efficiency and real-time processing, sometimes at the cost of accuracy.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e(Supported by Figs.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e16\u003c/span\u003e and \u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e18\u003c/span\u003e: 23.26% used quantized models; only 11.63% fully reported model size and latency)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eH4: Most studies identify and partially address hardware limitations such as limited memory and low processing power, but often lack quantitative reporting.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e(Supported by Figs.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e12\u003c/span\u003e \u0026amp; \u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e17\u003c/span\u003e: 30.23% described general constraints; 16.28% failed to specify hardware at all)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eH5: There is increasing adoption of ML-enabled microcontroller systems in low-resource applications such as soil testing, water quality analysis, and environmental monitoring.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e(Supported by Fig.\u0026nbsp;\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e19\u003c/span\u003e: Smart IoT \u0026ndash; 25.58%; Environmental \u0026ndash; 13.95%; Agriculture \u0026ndash; 13.95%)\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eH6: Micronutrients such as zinc, iron, nitrogen, and phosphorus are the most commonly targeted nutrients, though reporting is inconsistent.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e(Supported by Figs.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e13\u003c/span\u003e \u0026amp; \u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e15\u003c/span\u003e: Custom datasets often include micronutrient data, but few studies report targets explicitly)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e1.3. Rationale\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe motivation for this systematic review lies in the urgent need to understand how lightweight machine learning (ML) frameworks can be deployed on microcontroller-class hardware for real-time, in-situ micronutrient sensing. Conventional lab-based nutrient analysis methods are costly, infrastructure-dependent, and impractical for widespread deployment in rural or low-resource settings. In contrast, embedded ML systems offer scalable, portable, and cost-effective sensing platforms, especially when optimized for low-power operation.\u003c/p\u003e \u003cp\u003eThis review focuses on how ML models\u0026mdash;particularly those adapted for edge computing via frameworks like TinyML and TensorFlow Lite\u0026mdash;perform in real-world or simulated environments when deployed on hardware-constrained platforms such as ESP32, STM32, and ARM Cortex-M series. By analyzing studies from 2015 to 2025, this review highlights the technical trade-offs between model complexity, memory usage, energy consumption, and accuracy, while also evaluating how dataset quality, sensor integration, and deployment environments affect sensing outcomes.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e1.4. Objectives\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe primary objective of this systematic review is to evaluate how machine learning (ML) models are applied on microcontroller-class hardware for micronutrient sensing, with particular emphasis on performance, optimization, and deployment feasibility in constrained environments. Specific objectives include:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eTo identify the most common ML models and frameworks used in micronutrient sensing applications, and assess their suitability for embedded systems.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTo examine hardware specifications and sensor integration approaches, particularly with respect to power efficiency, memory limitations, and real-time constraints.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTo analyze model performance metrics such as accuracy, latency, and energy efficiency, and evaluate the degree of reporting transparency across studies.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTo categorize deployment scenarios (e.g., soil testing, water analysis) and assess the extent of real-world validation of these systems.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eTo highlight reporting inconsistencies and research gaps, providing recommendations for future development of scalable, low-cost sensing solutions using embedded ML.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e1.5. Research Contributions\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis review provides a comprehensive synthesis of the current state of ML-enabled micronutrient sensing on microcontroller-class hardware. Key contributions include:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eA technical evaluation of lightweight ML frameworks\u0026mdash;such as TinyML and TensorFlow Lite\u0026mdash;and their integration with microcontrollers for real-time micronutrient detection, highlighting trade-offs in model size, latency, and energy usage (Figs.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e14\u003c/span\u003e, \u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e16\u003c/span\u003e, \u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e18\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eA comparative analysis of hardware utilization, microcontroller specifications, and reporting practices, revealing that only 32.56% of studies specified board-level hardware and over 16% lacked hardware descriptions entirely (Figs.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e12\u003c/span\u003e, \u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e17\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eA taxonomy of dataset types and application domains, showing dominant use of custom datasets (39.53%) and strong representation of IoT, agriculture, and environmental monitoring as application areas (Figs.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e15\u003c/span\u003e, \u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e19\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eThe development of a benchmarking framework that maps model characteristics (e.g., size, inference latency) to platform constraints, aiding developers in selecting appropriate ML solutions for field-deployable sensing.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e1.6. Research Novelty\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo the best of our knowledge, this is the first systematic review exclusively focused on the deployment of ML models on microcontroller-class hardware for micronutrient sensing across environmental and agricultural contexts. Its novelty lies in:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eProviding a cross-disciplinary synthesis of embedded systems, AI, and environmental sensing literature, with special attention to technical deployment on constrained devices.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eOffering a framework for assessing model-hardware compatibility, linking ML model performance metrics with microcontroller specifications and real-time operating needs.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eUncovering underreported areas such as sensor type documentation, micronutrient-specific datasets, and standardized benchmarking practices\u0026mdash;thus setting the agenda for future research in this growing subfield of edge AI for agriculture and environmental health.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this subsection, the study outlines the methodology employed to conduct a systematic review focusing on the applications and advantages of machine learning on microcontroller-class hardware for micronutrient sensing. The study is based on a review of literature published over the last decade, from 2015 to 2025. To the best knowledge of the authors, no similar comprehensive review exists within this specific timeframe, making this study a novel contribution to the field of machine learning on microcontroller-class hardware for micronutrient sensing. The research methodology includes the careful selection of relevant peer-reviewed articles from key online databases, namely Scopus, Google Scholar, and Web of Science, ensuring a thorough examination of the subject matter.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.1. Eligibility criteria\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eA systematic analysis of all peer-reviewed and published research works relevant to the study of the machine learning on microcontroller-class hardware for micronutrient sensing was conducted for examination. Only research tasks published in English between 2015 and 2025 were included in the analysis. A proper benchmark for inclusion was adapted to ensure the inclusion of research papers that specifically focus on this topic in hand and exclude those that do not fall under this topic. Inevitably, only peer-reviewed research works that fundamentally connect on the machine learning on microcontroller-class hardware for micronutrient sensing, and that include a research framework or methodology specific to these aspects, were exclusively considered. The inclusion and exclusion criteria for this review are tabulated as in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e (Myataza et al, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\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 micronutrient 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 micronutrient 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 frame-work or methodology for machine learning on microcontroller-class hardware for micronutrient 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 micronutrient 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=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.2. Information sources\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eA systematic study of online information was held to identify relevant studies for this systematic review. online sources like Scopus, Google Scholar, and Web of Science were utilized to find articles and journals that involve the review on machine learning on microcontroller-class hardware for micronutrient sensing. On each online source, keywords that are related to the topic where mixed together to form a strong search pattern, that will at the end provide the most relevant research articles that are already recorded (Myataza et al, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Google scholar was used to find conference papers, journal papers and book articles while web of science was used for citations of the literature found on google scholar website. Web of science was also utilized to find articles. The results from these online data sources formed the core of the literature review.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.3. Search strategy\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe literature studies that are obtained in this research were taken from online databases, focusing on keywords that address the context of machine learning on microcontrollers-class hardware for micronutrient sensing. Terms such as \u0026ldquo;key elements: AND \u0026ldquo; best practices\u0026rdquo;. A thorough search was carried out in three main online sources: Google Scholar, Scopus, and Web of Science. To find the most relevant studies, a specific set of keywords was used. These keywords were: \u0026ldquo;(machine learning)\u0026rdquo; AND \u0026ldquo;on\u0026rdquo; AND \u0026ldquo;(microcontroller)\u0026rdquo; AND \u0026ldquo;(class-hardware)\u0026rdquo; AND \u0026ldquo;for\u0026rdquo; AND \u0026ldquo;(micronutrients)\u0026rdquo; AND \u0026ldquo;(sensing)\u0026rdquo;. The terms were utilized as a search string to find studies that are relevant to machine learning on microcontroller-class hardware for micronutrients sensing. The research also focused on studies done in the last decade from 2015 to 2025 that were all written in English language. This time frame was chosen to get the newest articles that fall under the research topic. The results from the search conducted showed 17 100 studies found on Google Scholar, 1577 articles found on Scopus, and 136 papers from Web of Science.the articles that are selected, were selected carefully after going through them again to ensure that only those that are relevant to the research topic are included in when writing this review(Myataza et al, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e below shows the total number of articles found in each of the three main online sources that were used while Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates the keywords used in the search string.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\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 Sources\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNumber of results\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\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e17 100\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\u003e136\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\u003e1577\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18813\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=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e2.4. Selection process\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo begin the screening process, four reviewers looked for titles and abstracts of the first search results. When there were parts where we disagreed with each other, we would all come to a final agreement that got more votes from all 4 members. Following the initial screening, the researchers divided into pairs and independently assessed the titles and abstracts of all articles obtained from the search. Whenever disagreements occurred, the team engaged in discussions to reach a consensus on which articles should advance to the full-text review stage (Myataza et al, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). If consensus could not be reached, a third researcher was brought in to make the final determination. Following this, three researchers independently evaluated the full-text articles to verify their eligibility based on the inclusion criteria. As with previous steps, any disagreements were addressed through discussion. When necessary, the fourth researcher was consulted to make the final decision regarding the inclusion or exclusion of articles, as illustrated in Fig.\u0026nbsp;2.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e2.5. Data collection process\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo reduce errors and cut back on bias, we followed an appropriate approach to ensure that correct data was obtained from the studies. One member in the team focused on making sure that every member was collecting accurate data and then examined it thoroughly (Myataza et al, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Gumede et al, \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mudau et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mtjilibe et al, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Discussions were held to come to a common conclusion when slight difference was identified in the data No automation tool was used for data distillation and everything was double checked before they were written down on the literature and existing studies sheets. When details in the studies were ambiguous, we carefully examined all accessible materials, such as supplementary data, appendices, and related research, to clarify findings concerning machine learning on microcontroller-class hardware for micronutrient sensing. In situations where uncertainties persisted, we referred to our fourth reviewer, an expert in the subject matter, to confirm the accuracy of the data interpretation related to machine learning on microcontroller-class hardware for micronutrient sensing. When multiple reports from the same study were available, we established clear criteria to select the most relevant data, focusing on the most recent and comprehensive studies on machine learning on microcontroller-class hardware for micronutrient sensing published between 2015 and 2025. In situations where discrepancies existed between reports, we examined the methodologies and results concerning machine learning on microcontroller-class hardware for micronutrient sensing to address the inconsistencies. Only English-language studies were considered, with articles in other languages excluded to ensure consistency in our analysis and to prevent possible misinterpretations caused by language barriers, as illustrated in Fig.\u0026nbsp;3.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e2.6. Data items\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this section we focus on key data items targeted in this systematic review, focusing on both core outcomes and supplementary variables related to the application of ML on microcontroller-class hardware for micronutrient sensing (Khanyi et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). The primary outcomes include sensor accuracy, detection limits, power efficiency, hardware-software integration, and real-time processing capabilities. Variables that were considered are as follows; study design, hardware specifications (e.g., microcontroller type, memory, processing power), ML model characteristics, environmental deployment settings, calibration strategies, and cost-effectiveness. This comprehensive approach enables a detailed understanding of how ML techniques enhance the performance and feasibility of micronutrient sensing in low-resource or embedded systems. By examining a wide range of technical and contextual factors, the review offers insights into the scalability, reliability, and practical implementation of such systems across diverse use cases.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec15\" class=\"Section3\"\u003e \u003ch2\u003e2.6.1 Data Collection Method\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eTo ensure a complete understanding of the role of machine learning (ML) on microcontroller-class hardware for micronutrient sensing, this systematic review identified and clearly defined key outcomes capturing the technical, functional, and practical lengths of this technology merge (Khanyi et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). Our data collection plan was designed to incorporate robust evidence on the practicality, performance, and real-world applicability of such systems in environmental and health-related sensing tasks. The primary outcomes of interest included Sensor Accuracy, measured through detection limits, sensitivity, and precision in micronutrient quantification. We gathered all results that demonstrated how ML models enhanced signal processing and classification accuracy on constrained hardware platforms. Energy Efficiency was another critical outcome, reflecting the system's suitability for deployment in remote or low-power environments. Studies reporting on power consumption, computational efficiency, and battery life were included to evaluate the practicality of sustained field operation. Hardware Utilization was also examined by assessing microcontroller specifications such as memory usage, processing speed, and compatibility with ML models. This outcome helped gauge how effectively ML algorithms were optimized to function within limited hardware capabilities.\u003c/p\u003e \u003cp\u003eModel Performance and Training Procedures were analyzed to understand how ML approaches such as regression, classification, or neural networks were adapted for lightweight execution. We focused on studies that provided insights into model size, inference speed, and training protocols, both on-device and offloaded. Finally, Application Context and Deployment was assessed by identifying how and where these systems were implemented\u0026mdash;whether in soil, water, plant tissues, or food samples\u0026mdash;and evaluating the reliability of real-time micronutrient monitoring under varying environmental conditions. All relevant data across studies and deployment scenarios were included to capture a holistic view of the technological impact and scalability of ML-enabled micronutrient sensing on microcontroller-class hardware.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section3\"\u003e \u003ch2\u003e2.6.2 Definition of Collected Data Variables\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn addition to the primary outcomes, we carefully considered several additional variables to provide a complete understanding of the context in which machine learning (ML) models are applied on microcontroller-class hardware for micronutrient sensing. These variables were critical to investigate the findings and understand the broader implications of deploying such technologies in resource-constrained environments. Study characteristics were collected, including the geographical location, environmental setting (e.g., agricultural, water monitoring, or food safety), and the specific micronutrients targeted. This information allowed us to assess the applicability of findings across various use cases and deployment scenarios (Khanyi et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWe also documented hardware specifications and system configurations, including microcontroller type, processing power, memory size, and sensor integration, to evaluate how hardware constraints influence ML implementation. These technical parameters were essential for understanding compatibility and scalability. Details on the ML models used such as algorithm type, training method, inference process, and optimization techniques were also included to assess the computational demands and real-time performance of different approaches. Furthermore, sensor data characteristics such as sampling frequency, data preprocessing steps, and noise handling were gathered to evaluate the reliability of sensing inputs. Energy consumption metrics, deployment duration, and calibration frequency were recorded to understand long-term feasibility and operational sustainability. We also considered external influences like climate variability, field conditions, and regulatory standards affecting sensor deployment and ML model performance in real-world applications.\u003c/p\u003e \u003cp\u003eAs detailed in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, our systematic review involved structured manual searches across established scientific databases such as Google Scholar, SCOPUS, and Web Of Science. These searches were tailored to capture studies at the intersection of embedded ML, micronutrient detection, and low-power hardware systems. By identifying and defining these variables, our review delivers a powerful and context-rich analysis of ML-enabled micronutrient sensing on microcontroller-class platforms, offering valuable insights for researchers, developers, and policy-makers working toward sustainable and intelligent sensing technologies.\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\u003eGeographic location, environmental application (e.g. water monitoring), type of micronutrients analyzed, and deployment setting (lab vs. field).\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\u003eSpecifications of microcontroller-class hardware used, including processor type, memory capacity, energy consumption, and sensor integration.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eML implementation characteristics\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eTypes of machine learning algorithms applied (e.g., regression, classification), training methods (on-device vs. offline), model optimization techniques, and inference performance\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOperational Constraints\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003ePower supply limitations, environmental factors (e.g., humidity, temperature), and maintenance or deployment challenges in field conditions..\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\u003eRegulatory standards for nutrient detection, funding availability for low-cost sensing technologies, and policy incentives for water monitoring.\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=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e2.7. Study risk of bias assessment\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe studies reviewed under the topic was critical to evaluate the risk of bias to ensure the credibility and accuracy of the synthesized findings. To accomplish this, we employed the Newcastle-Ottawa Scale (NOS) for assessing non-randomized studies, including experimental testbeds, simulation analyses, and field-based evaluations. The NOS framework allowed assessment across three key domains: Selection, Comparability, and Outcome. Each included study received a quality score based on criteria such as the clarity of sensor deployment methods, the appropriateness of machine learning model training, and the reproducibility of results (Khanyi et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). A maximum of one star was allocated per item in the Selection and Outcome domains, and up to two stars for Comparability, particularly focusing on how well studies accounted for environmental variables like humidity, temperature, and sensor drift. As illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e4\u003c/span\u003e, four independent reviewers conducted assessments to enhance objectivity. Discrepancies were discussed among the reviewers, and if consensus was not reached, a fourth reviewer provided the final judgment. For studies that lacked detail or relied on proprietary hardware or algorithms, additional efforts were made to cross-reference supplementary sources from repositories like Google Scholar, Web Of Science and SCOPUS A detailed manual review of grey literature was also performed to ensure that all relevant information was accurately captured. No automation tools were used, maintaining the integrity and rigor of the bias assessment process.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e2.8. Synthesis methods\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe flowchart in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e5\u003c/span\u003e outlines the systematic methodology adopted in our review of ML applications on microcontroller-class hardware for micronutrient sensing. Beginning with the Study Selection Process, studies were identified and screened using defined inclusion and exclusion criteria, focusing on research involving embedded ML models and low-power sensor hardware (Khanyi et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). The next step, Data Standardization, involved organizing technical specifications, performance metrics, and sensing parameters into a consistent format for comparability. In the Data Analysis phase, information was structured into tables and visualizations to enable cross-study comparison of model accuracy, power consumption, and sensor integration. The review then proceeded to a Heterogeneity Assessment, where differences in hardware platforms, ML algorithms (e.g., decision trees, support vector machines, neural networks), and micronutrient targets were examined through subgroup analyses. Lastly, a Bias Assessment step was conducted to identify methodological limitations and ensure a balanced synthesis of the findings. This structured workflow enabled a comprehensive, transparent, and reproducible evaluation of how ML techniques are leveraged on microcontroller platforms for real-time, in-field micronutrient detection.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this systematic review on the application of ML on microcontroller-class hardware for micronutrient sensing, we employed rigorous synthesis methods to ensure our findings were comprehensive, transparent, and reproducible. To establish study eligibility, we systematically tabulated each study's characteristics including hardware specifications, sensor configurations, ML algorithms used, and nutrient targets and compared them against our predefined synthesis criteria. This process ensured the inclusion of only those studies directly relevant to embedded ML in resource-constrained sensing environments. In preparing the data for synthesis, we addressed missing values and incomplete reporting by applying estimation techniques and standardized unit conversions, thereby maintaining consistency across diverse technical studies. Results were synthesized using structured comparison tables and performance graphs to visualize metrics such as model accuracy, power efficiency, latency, and sensor integration effectiveness. We conducted a qualitative comparative analysis due to the technological variability across studies, supplemented by subgroup analyses focusing on microcontroller types (e.g., ARM Cortex-M, AVR) and the classes of ML models (e.g., decision trees, SVM, neural networks). This allowed us to explore the influence of computational constraints and model complexity on sensing accuracy. We also performed sensitivity analyses to assess the robustness of findings across various deployment environments, including laboratory, field, and in-situ agricultural settings. This comprehensive synthesis approach provided a good understanding of the trade-offs and practical considerations when deploying ML on microcontrollers for micronutrient detection, offering valuable insights for researchers, developers, and practitioners in embedded systems and precision agriculture.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec19\" 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 the use of ML on microcontroller-class hardware for micronutrient sensing, each study was meticulously evaluated for its relevance and alignment with the review\u0026rsquo;s specific objectives (Khanyi et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2025\u003c/span\u003e). We manually assessed and compared key study characteristics such as the type of microcontroller used, sensor integration, ML algorithms implemented, and targeted micronutrients against our predefined inclusion criteria. A comparison matrix was developed to visually map the scope, methodologies, and performance metrics of each study relative to our eligibility benchmarks. This structured approach ensured that only studies directly focused on embedded ML solutions for nutrient detection in constrained hardware environments were included, thereby enhancing the precision, reliability, and contextual relevance of our review findings.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e2.8.2. Data Preparation for Synthesis\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this review, data preparation methods were employed to standardize and harmonize findings across studies focused on machine learning implementation on microcontroller-class hardware for micronutrient sensing. Given the diversity in reporting formats, data such as detection accuracy, response times, and hardware resource usage were converted into comparable metrics to enable meaningful synthesis. When studies reported performance indicators differently for example, using classification accuracy versus F1-score appropriate statistical transformations were applied to align them onto a consistent evaluative scale. Additionally, in cases where studies lacked critical summary statistics, such as standard deviations or confidence intervals, imputation techniques and estimation formulas based on available data were used to fill these gaps. This ensured a complete and coherent dataset, facilitating a reliable comparison and integration of findings across the diverse body of research examined.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section3\"\u003e \u003ch2\u003e2.8.3. Tabulation and Visual Display of Results\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eResults from individual studies and synthesized analyses in this review of machine learning on microcontroller-class hardware for micronutrient sensing were organized using both tabular and graphical formats to enhance interpretability and comparison. Structured tables were created to present study outcomes, categorized by performance domains such as detection accuracy, computational efficiency, energy consumption, and sensor integration. Within each domain, studies were arranged according to their assessed risk of bias, allowing readers to identify the most credible evidence at a glance. For visual representation, forest plots were employed to display quantitative synthesis results, illustrating effect sizes and confidence intervals for each included study alongside overall estimates. These visualizations were organized either by publication year or magnitude of effect, which helped reveal technological progress over time and emerging patterns in microcontroller-based sensing strategies. This dual approach to presentation ensured both clarity and analytical depth across the dataset.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section3\"\u003e \u003ch2\u003e2.8.4. Synthesis of Results\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this systematic review of machine learning applications on microcontroller-class hardware for micronutrient sensing, we conducted a thorough synthesis of results gathered from manual searches across reputable repositories, including Google Scholar, Scopus, and Web of Science. The synthesis process was carefully guided by the nature of the data and the degree of heterogeneity observed among the studies. Each study\u0026rsquo;s results were evaluated to determine whether a fixed-effects or random-effects model would be more appropriate, based on the consistency of outcomes across various hardware platforms and sensing conditions. Following data extraction, we compiled the findings into Excel spreadsheets, where visualizations such as bar charts and scatter plots were created to support preliminary assessments of variability. This visual analysis allowed us to detect trends in model accuracy, processing latency, and power consumption, and to identify any notable divergences in performance across different microcontroller architectures. These insights provided a foundation for deeper exploration and contextual interpretation of the results.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003e2.8.5. Exploring Causes of Heterogeneity\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this review, data preparation methods were employed to standardize and harmonize Subgroup analyses and meta-regression were conducted to investigate potential sources of heterogeneity among studies evaluating the use of machine learning on microcontroller-class hardware for micronutrient sensing. Key factors explored included differences in hardware platforms (e.g., Arduino, Raspberry Pi, ESP32), types of machine learning algorithms implemented (e.g., decision trees, neural networks, support vector machines), and sensing modalities (e.g., optical, electrochemical, or biosensor-based techniques). These variables were examined to assess their influence on system performance, including accuracy, power efficiency, and processing time. By analyzing these subgroups, we aimed to identify consistent patterns and contextual dependencies that could explain the variability observed across studies and enhance the generalizability of findings within low-resource or embedded environments.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section3\"\u003e \u003ch2\u003e2.8.6. Sensitivity Analyses\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eSensitivity analyses were conducted to assess the robustness of the synthesis results concerning key assumptions and methodological choices in this review of machine learning on microcontroller-class hardware for micronutrient sensing. These analyses involved re-evaluating the findings by excluding studies with high risk of bias or limited methodological transparency, and by applying alternative statistical models to test the stability of aggregated outcomes. We also examined the influence of varying hardware configurations, sensor types, and algorithmic complexity on reported performance metrics. This rigorous approach ensured that our conclusions remained consistent across different analytical frameworks, enhancing the reliability and validity of the synthesized evidence.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e2.9. Reporting bias assessment\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn conducting our systematic review on the application of machine learning on microcontroller-class hardware for micronutrient sensing, assessing the risk of reporting bias was critical to ensuring the integrity and validity of our synthesis. We focused on identifying potential biases arising from selective publication or incomplete reporting of outcomes, as these could significantly skew the representation of technological capabilities and limitations. To address this, we utilized contour-enhanced funnel plots as a key visual tool for detecting asymmetry in reported performance metrics, such as accuracy, latency, and power consumption. These plots, complemented by significance contours, allowed us to distinguish between potential publication bias and random variation (Khanyi et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eRather than developing new assessment tools, we employed established methods recognized in the literature for their methodological rigor and suitability for technical evaluations. All assessments were conducted manually, using tools like Excel to generate plots and charts, enabling us to maintain a hands-on approach and closely examine study distributions. This manual process also ensured that nuanced patterns, often overlooked by automated tools, were properly captured and analyzed.\u003c/p\u003e \u003cp\u003eMultiple reviewers independently assessed each study for reporting bias, with discrepancies resolved through consensus or consultation with an expert in embedded machine learning systems. In addition, we conducted an exhaustive manual search of databases including Google Scholar, Scopus, and Web of Science to identify studies potentially missed by automated indexing, particularly unpublished or conference-based work commonly found in hardware-focused research.\u003c/p\u003e \u003cp\u003eGiven the engineering and sensor-technology focus of our topic, we adapted traditional bias assessment frameworks to reflect the reporting norms and dissemination practices in this domain. This tailored approach improved the contextual relevance of our analysis. All assessment procedures and criteria have been clearly documented and included in the supplementary materials, reinforcing transparency and enabling reproducibility for future studies in the growing field of low-power ML-based micronutrient sensing platforms.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e2.10. Certainty assessment\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe reviewed literature was evaluated referring to the 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 clear research aim\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eQA2: transparent data collection\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eQA3: Clear ML processes\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eQA4:proper and appropriate research methodology.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eQA5: contributions of ML to microcontrollers-class hardware.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe certainty assessment is based on the scale of zero (0) to one (1). (0) point is given to a response that is not assigned, (0.5) is given when the criteria is partially met and (1) is given when the total score is met or a \u0026lsquo;yes\u0026rsquo;, it is assigned. The five criterias were using this lamina. Each piece of literature under review can receive a total score between 0 and 5 points.This results are for the collected literature on the applications and performance benefits of machine learning in microcontroller-class hardware 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 Data Mining and Business Intelligence in SMEs.\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=\"char\" char=\".\" 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\u003ePereira, 2022; Pereira et al., \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; M EL Adoui, 2024\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\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e100\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDom\u0026egrave;nech-Gil et al., 2024\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=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e90\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDhal et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Madson, 2022; Dalmeida \u0026amp; Kamath, \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2023\u003c/span\u003e\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=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e80\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDalmeida \u0026amp; Kamath, 2024; Salve \u0026amp; Jemila, 2024\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=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e70\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVenkatesh \u0026amp; Naik, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2023\u003c/span\u003e\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=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e60\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\u003eTo support the conclusions of this systematic review on the implementation and practical benefits of machine learning on microcontroller-class hardware for micronutrient sensing, we conducted a thorough evaluation of the reliability of the evidence.The validity and robustness of our conclusions rely on a structured appraisal process, which was implemented using the Grading of Recommendations, Assessment, Development, and Evaluations (GRADE) framework. This framework is an internationally acknowledged method that provides a thorough and clear strategy for evaluating the strength of evidence, ensuring that the resulting conclusions are reliable and substantiated. The reliability of the evidence related to major findings was carefully assessed through a set of essential criteria. First, we closely examined the precision of effect estimates by considering sample sizes and the width of confidence intervals in the studies. Tight uncertainty margins alongside substantial sample counts signified a strong degree of confidence in the findings, as they reflect more consistent and accurate outcome estimations. We further evaluated the uniformity of outcomes by analyzing patterns across the reviewed studies. A strong alignment in reported effects across multiple sources strengthened the overall confidence in the evidence. Any identified variability among the studies was carefully examined to determine its origins and assess how it might influence the overall conclusions (Khanyi et al., \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Thobejane et al., 2024; Skosana et al, \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Mkhize et al., \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2025\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":" Results","content":"\u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003e3.1. Study selection\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eIn this systematic review, the study selection process adhered to a rigorous and methodical approach to ensure the inclusion of relevant and high-quality research articles. Three prominent online databases\u0026mdash;Google Scholar, Scopus, and Web of Science\u0026mdash;were systematically searched to identify studies that met the predefined inclusion criteria. Specifically, the search retrieved 10,000 records from Google Scholar, 1,577 records from Scopus, and 156 records from Web of Science, resulting in a total of 11,713 initial records.\u003c/p\u003e \u003cp\u003eFollowing this, duplicate entries were removed, leaving 4,703 unique records. These records underwent a preliminary screening process, during which their titles and abstracts were assessed for relevance to the review\u0026rsquo;s inclusion criteria. This step resulted in the selection of 50 full-text documents for a more detailed evaluation. After a comprehensive review of these full-text documents, 43 studies were deemed eligible for inclusion in the final systematic review. The included studies consisted of 22 journal articles, 16 conference papers, 2 book chapters, and 3 dissertations or theses.\u003c/p\u003e \u003cp\u003eThe overall study collection process is depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e6\u003c/span\u003e, which provides an in-depth PRISMA flowchart outlining the progression of records through each stage of the review.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section2\"\u003e \u003ch2\u003e3.2. Study results\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe trend in annual publication output reveals a steady increase in research interest from 2019, peaking in 2022 with the highest number of publications (n\u0026thinsp;=\u0026thinsp;12). This growth reflects rising attention toward embedded ML for environmental sensing. However, a noticeable decline follows in 2023 and beyond, possibly due to saturation in foundational studies or shifts toward implementation-focused research.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e8\u003c/span\u003e illustrates the geographical distribution of the reviewed studies, highlighting the global interest in ML-based micronutrient sensing on embedded platforms. Israel and Kenya lead with the highest contribution (12% each), followed by Switzerland (7%) and several other countries\u0026mdash;including the USA, Brazil, Finland, and India\u0026mdash;each contributing 5%. This spread demonstrates a diverse research base, with both high-income and developing countries engaging in innovations aimed at low-cost, real-time nutrient sensing.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e9\u003c/span\u003e, journal articles constitute the majority of included studies (51.16%), followed by conference papers (37.21%), dissertations or theses (6.98%), and book Chaps.\u0026nbsp;(4.65%). This distribution suggests that peer-reviewed journals and conferences are the primary mediums for disseminating research on embedded ML for micronutrient sensing. The presence of theses and book chapters also indicates emerging academic interest and growing educational integration in this field.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThe majority of the studies included in this review were retrieved from Google Scholar (46.51%), followed by Scopus (32.56%) and Web of Science (20.93%), as illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e10\u003c/span\u003e. This distribution highlights Google Scholar\u0026rsquo;s broader indexing capacity, particularly for conference papers, theses, and grey literature relevant to emerging research fields like embedded machine learning for micronutrient sensing.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e11\u003c/span\u003e shows that most studies (39.53%) relied on established ML or TinyML frameworks and platforms, demonstrating a preference for well-supported, standardized tools. Custom or proprietary solutions were also significant (23.26%), reflecting adaptation to specific hardware constraints or sensing tasks. Interestingly, 13.95% of studies did not specify their framework, pointing to reporting gaps. Other approaches included hybrid methods (9.30%), niche research tools (6.98%), and a smaller proportion using commercial (MicroAI) or multi-domain systems (2.33%). This diversity indicates an evolving landscape of framework adoption tailored to embedded micronutrient sensing.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e12\u003c/span\u003e, 32.56% of studies clearly stated the exact microcontroller models or boards used, indicating a commitment to replicability and technical precision. A slightly smaller portion (30.23%) used multiple or hybrid hardware platforms, showcasing versatility across different configurations. Meanwhile, 20.93% of papers referred only to broader MCU families (e.g., ARM Cortex), and 16.28% did not specify hardware details at all. These gaps in reporting may hinder reproducibility and limit comparative performance analysis.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e13\u003c/span\u003e presents the distribution of research focus areas, with the majority of studies (53.49%) centering on core ML tasks such as classification and regression, underscoring their foundational role in embedded sensing applications. Other key domains include human-centered interaction (11.63%), computer vision (9.30%), and industrial or infrastructure monitoring (6.98%). Agriculture and environmental monitoring, directly relevant to micronutrient sensing, accounted for 4.65% of studies. A small portion also explored learning paradigms (6.98%), multi-domain integration (2.33%), and sentiment or language-based analytics (2.33%), reflecting early-stage diversification of use cases.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eAs shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e14\u003c/span\u003e, structured deep learning models such as CNNs, RNNs, and LSTMs were used in 37.21% of the studies, indicating strong reliance on advanced architectures for complex sensing tasks. Ensemble and hybrid methods followed at 20.93%, reflecting the value of combining multiple algorithms to enhance performance under hardware constraints. Classical ML models (16.28%) remained relevant, particularly for simpler tasks requiring low computational resources. Notably, 9.30% of papers referenced deep learning without specifying model types, while others explored specialized neural networks (2.33%) or general frameworks (4.65%), revealing ongoing variation in modeling practices.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eAccording to Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e15\u003c/span\u003e, custom datasets\u0026mdash;curated or generated for specific use cases\u0026mdash;were used in 39.53% of the studies, reflecting the niche and application-specific nature of micronutrient sensing. Open-source and benchmark datasets were employed in 23.26% of papers, offering standardized comparability. However, 18.60% of studies did not specify dataset type, revealing a documentation gap. Additional dataset categories included general or mixed datasets (11.63%), domain-specific time-series data (4.65%), and real-time environmental sensor data (2.33%), underscoring the diversity and evolving nature of data collection approaches in embedded ML research.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e16\u003c/span\u003e highlights a critical limitation in the literature: 46.51% of studies did not specify model size or latency metrics, hindering reproducibility and performance benchmarking. While 23.26% of studies noted using optimized or quantized models, they lacked exact specifications. Only 11.63% explicitly reported both model size and latency, offering precise performance data. The remainder used general (13.95%) or performance-focused (2.33%) descriptions, or made timing-only remarks (2.33%). These findings underscore the need for improved transparency and standardization in embedded ML reporting.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eAs seen in Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e17\u003c/span\u003e, 30.23% of the reviewed studies referenced general hardware constraints, typical of TinyML applications, while 25.58% mentioned device-specific limitations without providing numerical specifications. Only 4.65% explicitly reported low-memory environments (limited RAM and flash), and just 9.30% offered detailed configurations like PSRAM usage, revealing a scarcity of advanced embedded system descriptions. Notably, 16.28% of studies did not mention hardware limitations at all, and 2.33% addressed ultra-low SRAM scenarios. These results point to the need for greater consistency and technical depth in documenting edge device limitations.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e18\u003c/span\u003e reveals that 65.12% of studies confirmed real-time performance, demonstrating a strong focus on achieving low-latency inference in embedded environments. A smaller proportion (11.63%) reported partial or conditional real-time capability, while 9.30% did not specify this metric at all. Other papers emphasized integration contexts (4.65%), real-time classification requirements (2.33%), educational demonstrations (2.33%), STM32-based real-time designs (2.33%), and real-time optimization via model tuning (2.33%). These findings highlight the increasing emphasis on verified real-time functionality in ML systems for edge applications.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eAs depicted in Fig.\u0026nbsp;\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e19\u003c/span\u003e, the dominant application area was Smart IoT and embedded systems (25.58%), underscoring the importance of seamless hardware-software integration in edge AI solutions. Environmental monitoring and agriculture, directly linked to micronutrient sensing, accounted for 13.95% of studies. Other significant areas included visual processing (13.95%), human-centric applications (11.63%), and advanced ML deployment on edge platforms (11.63%). Smaller contributions came from industrial IoT (6.98%), audio/multimodal sensing (6.98%), and lifelong learning (2.33%). A notable 9.30% of studies did not specify an application area, reflecting some inconsistencies in scope definition.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig18\" class=\"InternalRef\"\u003e20\u003c/span\u003e shows that 25.58% of studies used edge ML-specific tools and frameworks, emphasizing the growing adoption of domain-optimized solutions tailored for constrained environments. Custom or proprietary tools followed closely (23.26%), reflecting the need for specialized implementations. General or mixed tools were used in 18.60% of papers, while standard embedded development tools\u0026mdash;typical of classical MCU programming\u0026mdash;appeared in 16.28%. Notably, 16.28% of studies did not specify the tools used, signaling a need for improved reporting practices in embedded ML research.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec30\" class=\"Section2\"\u003e \u003ch2\u003e3.3. Study research questions results\u003c/h2\u003e \u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003e \u003cb\u003eWhich types of machine learning algorithms are most frequently recommended and successfully implemented on microcontroller-class hardware for micronutrient detection?\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe review revealed that structured deep learning models, especially Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Long Short-Term Memory networks (LSTMs), were the most frequently employed algorithms, used in 37.21% of studies (Fig.\u0026nbsp;\u003cspan refid=\"Fig12\" class=\"InternalRef\"\u003e14\u003c/span\u003e). These models were often adapted for embedded inference using quantization or pruning techniques to reduce computational load. Ensemble and hybrid methods, such as combinations of Random Forests and neural networks, accounted for 20.93%, while classical machine learning models like decision trees and Support Vector Machines (SVMs) were applied in 16.28% of the literature. Their adoption was primarily driven by their low memory requirements and faster inference times, making them suitable for real-time applications on devices with limited RAM and processing capabilities. However, despite this diversity in algorithm selection, nearly 10% of studies referenced deep learning without specifying the exact model type, indicating a need for clearer documentation.\u003c/p\u003e \u003cp\u003e \u003cb\u003eWhat types of sensors are most suitable and commonly used in micronutrient sensing systems based on microcontrollers?\u003c/b\u003e \u003c/p\u003e \u003cp\u003eWhile not all reviewed studies explicitly documented the types of sensors used, patterns emerged suggesting that electrochemical sensors are the most suitable for integration with microcontroller-based platforms due to their low power requirements, compact form factors, and high sensitivity to trace nutrient elements. These sensors were particularly relevant in applications involving soil or water quality analysis. In image-based studies\u0026mdash;particularly those involving CNNs\u0026mdash;smartphone cameras and optical sensors were used to infer micronutrient levels from plant leaves or soil surfaces. However, a significant number of studies (around 20%) did not clearly report the sensing mechanism or sensor model, which limits reproducibility and cross-study comparisons. This reporting gap highlights the need for standardized sensor documentation in embedded ML research.\u003c/p\u003e \u003cp\u003e \u003cb\u003eWhat are the typical performance outcomes (e.g., accuracy, latency, energy efficiency) observed when deploying ML models for micronutrient sensing on constrained devices?\u003c/b\u003e \u003c/p\u003e \u003cp\u003ePerformance evaluation in the reviewed studies commonly focused on real-time capability and energy efficiency, though quantitative performance metrics were inconsistently reported. Approximately 65.12% of studies claimed real-time processing capabilities (Fig.\u0026nbsp;\u003cspan refid=\"Fig16\" class=\"InternalRef\"\u003e18\u003c/span\u003e), and 23.26% indicated the use of optimized or quantized models to reduce latency and improve battery life (Fig.\u0026nbsp;\u003cspan refid=\"Fig14\" class=\"InternalRef\"\u003e16\u003c/span\u003e). However, only 11.63% of studies provided explicit measurements for both model size and latency, while 46.51% did not report these metrics at all. This lack of transparency makes it difficult to assess scalability and system viability under field conditions. Despite this, many models achieved competitive accuracy, particularly when custom datasets were used, though validation against real-world conditions was often limited. Energy efficiency was discussed qualitatively, with several studies emphasizing low-power deployment on platforms like the ESP32 and STM32, but only a few included concrete power consumption benchmarks.\u003c/p\u003e \u003cp\u003e \u003cb\u003eHow do existing studies address key challenges such as limited memory, real-time processing demands, and power consumption in microcontroller-class environments?\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe majority of studies acknowledged the fundamental hardware constraints of microcontroller-class devices. Approximately 32.56% specified exact microcontroller boards\u0026mdash;such as ARM Cortex-M, ESP32, or STM32\u0026mdash;while 30.23% described general resource limitations associated with TinyML environments (Fig.\u0026nbsp;\u003cspan refid=\"Fig10\" class=\"InternalRef\"\u003e12\u003c/span\u003e). Commonly cited challenges included low RAM availability (often under 256 KB), limited flash storage, and the need for fast inference without cloud dependency. However, only 4.65% of studies explicitly quantified memory or processing limits, and 16.28% failed to mention hardware constraints altogether (Fig.\u0026nbsp;\u003cspan refid=\"Fig15\" class=\"InternalRef\"\u003e17\u003c/span\u003e). To overcome these challenges, some studies employed model compression techniques like quantization, and others used tailored frameworks such as TensorFlow Lite for Microcontrollers. Nevertheless, a lack of uniformity in hardware reporting and optimization strategy limits the transferability of these solutions across different deployments.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTo what extent are ML-enabled microcontroller systems being deployed or validated in real-world applications, including agricultural, environmental, and low-resource settings?\u003c/b\u003e \u003c/p\u003e \u003cp\u003eSeveral studies demonstrated partial or full real-world applicability of embedded ML systems, particularly in the domains of smart IoT (25.58%), environmental monitoring (13.95%), and agriculture (13.95%) (Fig.\u0026nbsp;\u003cspan refid=\"Fig17\" class=\"InternalRef\"\u003e19\u003c/span\u003e). Examples include ML-powered soil nutrient analyzers deployed in field trials, and water potability monitoring systems integrated into remote sensor nodes. Although many studies used simulated or lab-based environments to validate system performance, the increasing use of edge-capable frameworks and power-efficient microcontrollers shows strong momentum toward practical deployment. Real-time data processing, offline functionality, and remote monitoring capabilities were identified as core features enhancing suitability for low-resource regions. Nonetheless, scalability and maintenance under long-term deployment conditions were less frequently addressed, indicating an area for future research.\u003c/p\u003e \u003cp\u003e \u003cb\u003eWhich micronutrients are being targeted in ML-based sensing applications on microcontroller-class hardware, and how consistently are these reported across studies?\u003c/b\u003e \u003c/p\u003e \u003cp\u003eTargeted micronutrients in the reviewed studies included zinc, iron, nitrogen, phosphorus, and potassium, typically detected through soil or water-based sensing systems (Figs.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e13\u003c/span\u003e and \u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e15\u003c/span\u003e). In some cases, micronutrient deficiencies were inferred through plant leaf image analysis, particularly for nitrogen and phosphorus. However, reporting consistency was low: while a few studies focused specifically on nutrient detection, others addressed more general soil health or environmental monitoring without detailing which micronutrients were measured. Only a subset of studies provided ground-truth chemical validation of sensing results, and many lacked domain-specific context. This inconsistency signals the need for more rigorous documentation of nutrient-specific sensing objectives, datasets, and validation methods in future work.\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003e \u003cdiv class=\"BlockQuote\"\u003e \u003cp\u003eThis systematic review provides a comprehensive evaluation of machine learning (ML) applications on microcontroller-class hardware for micronutrient sensing across environmental, agricultural, and embedded contexts. The findings reveal a growing global interest, with a peak in research output around 2022, and increasing contributions from both high-income and resource-constrained regions. Despite promising developments in deploying deep learning, ensemble models, and TinyML frameworks on platforms such as STM32, ESP32, and ARM Cortex-M, the review highlights persistent limitations. These include inconsistent hardware specification, limited use of standard datasets, and a widespread lack of transparency regarding model size, latency, and power consumption. While 65.12% of studies confirmed real-time performance, only a fraction provided quantified performance metrics, which limits reproducibility and benchmarking across studies. Custom datasets and domain-specific tools dominated, reflecting the specialized nature of micronutrient sensing tasks. However, this also introduces challenges for generalization and scalability. Application areas such as environmental monitoring, smart IoT, and health-focused sensing demonstrate the versatility of ML-embedded microcontroller systems, yet underscore the need for robust, field-tested implementations. To fully harness the potential of ML on microcontroller-class devices for micronutrient sensing, future research should focus on:\u003c/p\u003e \u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eEstablishing standardized benchmarks and reporting practices,\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eEmphasizing low-power, real-time performance in real-world conditions,\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eEnhancing interdisciplinary collaboration among hardware engineers, data scientists, and domain experts.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003ePereira, E. A. M. (2024). Water potability classification: an approach using machine learning in an embedded system.\u003c/li\u003e\n\u003cli\u003eDhal, S. B., Jungbluth, K., Lin, R., Sabahi, S. P., Bagavathiannan, M., Braga-Neto, U., \u0026amp; Kalafatis, S. (2022). A machine-learning-based IoT system for optimizing nutrient supply in commercial aquaponic operations. Sensors, 22(9), 3510.\u003c/li\u003e\n\u003cli\u003ePereira, E. A. M., da Silva Santos, J. F., \u0026amp; de Andrade Barboza, E. (2024). An energy efficient tinyml model for a water potability classification problem. Sustainable Computing: Informatics and Systems, 43, 101010.\u003c/li\u003e\n\u003cli\u003eGASANA, J. M. (2022). IoT based monitoring system of soil properties (Doctoral dissertation, College of science and Technology).\u003c/li\u003e\n\u003cli\u003eSchizas, N., Karras, A., Karras, C., \u0026amp; Sioutas, S. (2022). TinyML for ultra-low power AI and large scale IoT deployments: A systematic review. Future Internet, 14(12), 363.\u003c/li\u003e\n\u003cli\u003eDalmeida, S. P., \u0026amp; Kamath, S. (2023). Soil micronutrient detection using machine learning. In Recent Trends in Computational Sciences (pp. 29-35). CRC Press.\u003c/li\u003e\n\u003cli\u003eSalve, P. R., Jemila, M. M., Sridharan, S., Meti, V. K. V., \u0026amp; Roy, B. (2025). High-Performance Computer-Integrated Microcontrollers for Enhancing Food Quality Monitoring: Innovations and Applications. In Integrating Machine Learning Into HPC-Based Simulations and Analytics (pp. 355-382). IGI Global Scientific Publishing. https://www.igi-global.com/chapter/high-performance-computer-integrated-microcontrollers-for-enhancing-food-quality-monitoring/365689 \u003c/li\u003e\n\u003cli\u003eVenkatesh, K., \u0026amp; Naik, K. J. (2023, July). Deep learning for macro-nutrient deficiency identification in the groundnut plants. In IET Conference Proceedings CP832 (Vol. 2023, No. 5, pp. 193-198). Stevenage, UK: The Institution of Engineering and Technology.https://digital-library.theiet.org/doi/abs/10.1049/icp.2023.1489\u003c/li\u003e\n\u003cli\u003eEl Adoui, M., Herpoel, T., \u0026amp; Fr\u0026eacute;nay, B. (2024). Constrained Tiny Machine Learning for Predicting Gas Concentration with I4. 0 Low-cost Sensors. ACM Transactions on Embedded Computing Systems, 23(3), 1-23.https://dl.acm.org/doi/abs/10.1145/3590956\u003c/li\u003e\n\u003cli\u003eKiplimo, E., Riddick, S. N., Mbua, M., Upreti, A., Anand, A., \u0026amp; Zimmerle, D. J. (2024). Addressing low-cost methane sensor calibration shortcomings with machine learning. Atmosphere, 15(11), 1313. https://www.mdpi.com/2073-4433/15/11/1313 \u003c/li\u003e\n\u003cli\u003e\u0026Uuml;nsalan, C., H\u0026ouml;ke, B., \u0026amp; Atmaca, E. (2024). Embedded Machine Learning with Microcontrollers.\u003c/li\u003e\n\u003cli\u003eAlenezi, M., Alshammari, R., \u0026amp; Alrashed, B. (2019). Enhancing Project Scope Management in Construction Projects Using Building Information Modeling (BIM). 2019 IEEE 2nd International Conference on Information and Computer Technologies (ICICT), 133\u0026ndash;137. https://doi.org/10.1109/INFOCT.2019.8987754\u003c/li\u003e\n\u003cli\u003eJain, N., Awasthi, Y., \u0026amp; Jain, R. K. (2023). An IoT-based soil analysis system using optical sensors and multivariate regression. Int. J. Exp. Res. Rev, 31, 23-32. https://www.academia.edu/download/104796158/3_An_IoT_based_soil_analysis_system_using_optical\n_sensors_and_multivariate_regression.pdf\u003c/li\u003e\n\u003cli\u003eGrau, M. M., Centelles, R. P., \u0026amp; Freitag, F. (2021, September). On-device training of machine learning models on microcontrollers with a look at federated learning. In Proceedings of the Conference on Information Technology for Social Good (pp. 198-203). https://dl.acm.org/doi/abs/10.1145/3462203.3475896\u003c/li\u003e\n\u003cli\u003eNguyen, V. K., Tran, V. K., Pham, H., Nguyen, V. M., Nguyen, H. D., \u0026amp; Nguyen, C. N. (2023). A multi-microcontroller-based hardware for deploying Tiny machine learning model. International Journal of Electrical and Computer Engineering (IJECE), 13(5), 5727-5736.https://www.academia.edu/download/113156003/16869.pdf \u003c/li\u003e\n\u003cli\u003eElsts, A., \u0026amp; McConville, R. (2021). Are microcontrollers ready for deep learning-based human activity recognition?. Electronics, 10(21), 2640. https://www.mdpi.com/2079-9292/10/21/2640.\u003c/li\u003e\n\u003cli\u003eMyataza, A., Mafunga, M., Mkhulisi, N. S., \u0026amp; Thango, B. A. A Systematic Review of ERP, CRM, and HRM Systems for SMEs: Managerial and Employee Support. Preprints 2024, 2024100384.\u003c/li\u003e\n\u003cli\u003eGumede, T. T., Chiworeka, J. M., Magoda, A. S., \u0026amp; Thango, B. Building Effective Social Media Strategies for Business: A Systematic Review. Preprints 2024, 2024100379.\u003c/li\u003e\n\u003cli\u003eVidal, J. (2025). Scalable Embedded Tiny Machine Learning (SETML): A General Framework for Embedded Distributed Inference (Doctoral dissertation, Massachusetts Institute of Technology). https://dspace.mit.edu/handle/1721.1/159107\u003c/li\u003e\n\u003cli\u003ePlancher, B., Buttrich, S., Ellis, J., Goveas, N., Kazimierski, L., Sotelo, J. L., ... \u0026amp; Reddi, V. J. (2024, May). TinyML4D: Scaling Embedded Machine Learning Education in the Developing World. In Proceedings of the AAAI symposium series (Vol. 3, No. 1, pp. 508-515). http://ojs.aaai.org/index.php/AAAI-SS/article/view/31265\u003c/li\u003e\n\u003cli\u003eDhal, S. B., Jungbluth, K., Lin, R., Sabahi, S. P., Bagavathiannan, M., Braga-Neto, U., \u0026amp; Kalafatis, S. (2022). A machine-learning-based IoT system for optimizing nutrient supply in commercial aquaponic operations. Sensors, 22(9), 3510. https://www.mdpi.com/1424-8220/22/9/3510\u003c/li\u003e\n\u003cli\u003eBoussion, N., Bosc, L., Goasduff, G., Jegat, Y., Lann, L., Lucia, A. S., ... \u0026amp; Visvikis, D. To calibrate or not to calibrate: a decision support approach using microcontroller-based tiny machine learning. https://www.iccr2024.org/papers/525059.pdf\u003c/li\u003e\n\u003cli\u003eZhou, X., Kang, Z., Canady, R., Bao, S., Balasubramanian, D. A., \u0026amp; Gokhale, A. (2021, November). Exploring cloud assisted tiny machine learning application patterns for PHM scenarios. In Annual Conference of the PHM Society (Vol. 13, No. 1). http://www.papers.phmsociety.org/index.php/phmconf/article/view/3054\u003c/li\u003e\n\u003cli\u003eSimitha, K. M., \u0026amp; Raj, S. (2019, June). IoT and WSN based water quality monitoring system. In 2019 3rd International conference on Electronics, Communication and Aerospace Technology (ICECA) (pp. 205-210). IEEE.\u003c/li\u003e\n\u003cli\u003eLin, J., Chen, W. M., Lin, Y., Gan, C., \u0026amp; Han, S. (2020). Mcunet: Tiny deep learning on iot devices. Advances in neural information processing systems, 33, 11711-11722.\u003c/li\u003e\n\u003cli\u003eMudau, M. C., Moshapo, L. W., Monyela, T. M., \u0026amp; Thango, B. A. The Role of Manufacturing Operations in SMEs Performance: A Systematic Review. Preprints 2024, 2024100539.\u003c/li\u003e\n\u003cli\u003eMtjilibe, T., Rameetse, E., Mgwenya, N., \u0026amp; 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/li\u003e\n\u003cli\u003eSakr, F., Bellotti, F., Berta, R., \u0026amp; De Gloria, A. (2020). Machine learning on mainstream microcontrollers. Sensors, 20(9), 2638.\u003c/li\u003e\n\u003cli\u003eWang, X., Magno, M., Cavigelli, L., \u0026amp; Benini, L. (2020). FANN-on-MCU: An open-source toolkit for energy-efficient neural network inference at the edge of the Internet of Things. IEEE Internet of Things Journal, 7(5), 4403-4417.\u003c/li\u003e\n\u003cli\u003eMoosmann, J., Giordano, M., Vogt, C., \u0026amp; Magno, M. (2023, June). Tinyissimoyolo: A quantized, low-memory footprint, tinyml object detection network for low power microcontrollers. In 2023 IEEE 5th International Conference on Artificial Intelligence Circuits and Systems (AICAS) (pp. 1-5). IEEE.\u003c/li\u003e\n\u003cli\u003eBanbury, C., Zhou, C., Fedorov, I., Matas, R., Thakker, U., Gope, D., ... \u0026amp; Whatmough, P. (2021). Micronets: Neural network architectures for deploying tinyml applications on commodity microcontrollers. Proceedings of machine learning and systems, 3, 517-532.\u003c/li\u003e\n\u003cli\u003eImmonen, R., \u0026amp; H\u0026auml;m\u0026auml;l\u0026auml;inen, T. (2022). Tiny Machine Learning for Resource‐Constrained Microcontrollers. Journal of Sensors, 2022(1), 7437023.\u003c/li\u003e\n\u003cli\u003eSaha, S. S., Sandha, S. S., \u0026amp; Srivastava, M. (2022). Machine learning for microcontroller-class hardware: A review. IEEE Sensors Journal, 22(22), 21362-21390.\u003c/li\u003e\n\u003cli\u003eSchizas, N., Karras, A., Karras, C., \u0026amp; Sioutas, S. (2022). TinyML for ultra-low power AI and large scale IoT deployments: A systematic review. Future Internet, 14(12), 363.\u003c/li\u003e\n\u003cli\u003eSailesh, M., \u0026amp; Selvakumar, K. (2022). A novel framework for deployment of CNN models using post-training quantization on microcontrollers. Microprocessors and Microsystems, 94, 104634.\u003c/li\u003e\n\u003cli\u003eRavaglia, L., Rusci, M., Nadalini, D., Capotondi, A., Conti, F., \u0026amp; Benini, L. (2021). A tinyml platform for on-device continual learning with quantized latent replays. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 11(4), 789-802.\u003c/li\u003e\n\u003cli\u003eKopparapu, K., \u0026amp; Lin, E. (2021). TinyFedTL: Federated transfer learning on tiny devices. arXiv preprint arXiv:2110.01107.\u003c/li\u003e\n\u003cli\u003eKhanyi, M. B., Xaba, S. N., Mlotshwa, N. A., Thango, B., \u0026amp; 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\u003c/li\u003e\n\u003cli\u003eThobejane, L. T., \u0026amp; Thango, B. A. (2024). Partial Discharge Source Classification in Power Transformers: A Systematic Literature Review. Applied Sciences, 14(14), 6097.\u003c/li\u003e\n\u003cli\u003eAlajlan, N. N., \u0026amp; Ibrahim, D. M. (2022). TinyML: Enabling of inference deep learning models on ultra-low-power IoT edge devices for AI applications. Micromachines, 13(6), 851.\u003c/li\u003e\n\u003cli\u003eWang, T., Fang, K., Wei, W., Tian, J., Pan, Y., \u0026amp; Li, J. (2022). Microcontroller unit chip temperature fingerprint informed machine learning for IIoT intrusion detection. IEEE Transactions on Industrial Informatics, 19(2), 2219-2227.\u003c/li\u003e\n\u003cli\u003eAzevedo, M. B., de Medeiros, T. D. A., Medeiros, M. D. A., Silva, I., \u0026amp; Costa, D. G. (2023). Detecting face masks through embedded machine learning algorithms: A transfer learning approach for affordable microcontrollers. Machine Learning with Applications, 14, 100498.\u003c/li\u003e\n\u003cli\u003eLin, J. (2021). Efficient Algorithms and Systems for Tiny Deep Learning (Doctoral dissertation, Massachusetts Institute of Technology).\u003c/li\u003e\n\u003cli\u003eElsts, A., \u0026amp; McConville, R. (2021). Are microcontrollers ready for deep learning-based human activity recognition?. Electronics, 10(21), 2640.\u003c/li\u003e\n\u003cli\u003eExell, N. (2023). Applicability of TinyML for maintenance predictability.Verbiest, J. R., Bonnech\u0026egrave;re, B., Saeys, W., Van de Walle, P., Truijen, S., \u0026amp; Meyns, P. (2023). Gait stride length estimation using embedded machine learning. Sensors, 23(16), 7166.\u003c/li\u003e\n\u003cli\u003ePark, J. (2022). Hybrid Machine Learning and Physics-Based Modeling Approaches for Process Control and Optimization (Doctoral dissertation, Brigham Young University).\u003c/li\u003e\n\u003cli\u003eBellarmino, N., Cantoro, R., Huch, M., Kilian, T., Martone, R., Schlichtmann, U., \u0026amp; Squillero, G. (2023). A multilabel active learning framework for microcontroller performance screening. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 42(10), 3436-3449.\u003c/li\u003e\n\u003cli\u003eKwon, Y. D., Chauhan, J., \u0026amp; Mascolo, C. (2022, May). Yono: Modeling multiple heterogeneous neural networks on microcontrollers. In 2022 21st ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN) (pp. 285-297). IEEE.\u003c/li\u003e\n\u003cli\u003eSkosana, S., Mlambo, S., Madiope, T., \u0026amp; Thango, B. Evaluating Wireless Network Technologies (3G, 4G, 5G) and Their Infrastructure: A Systematic Review. Preprints 2024, 2024101331.\u003c/li\u003e\n\u003cli\u003eMkhize, A., Mokhothu, K. D., Tshikhotho, M., \u0026amp; Thango, B. A. (2025). Evaluating the Impact of Cloud Computing on SME Performance: A Systematic Review. Businesses, 5(2), 23. https://doi.org/10.3390/businesses5020023\u003c/li\u003e\n\u003cli\u003eBuonanno, L., Di Vita, D., Carminati, M., \u0026amp; Fiorini, C. (2020). A directional gamma-ray spectrometer with microcontroller-embedded machine learning. IEEE Journal on Emerging and Selected Topics in Circuits and Systems, 10(4), 433-443.\u003c/li\u003e\n\u003cli\u003eManor, E., \u0026amp; Greenberg, S. (2022). Custom hardware inference accelerator for tensorflow lite for microcontrollers. IEEE Access, 10, 73484-73493.\u003c/li\u003e\n\u003cli\u003eSilva, L. C., Simas Filho, E. F., Albuquerque, M. C., Silva, I. C., \u0026amp; Farias, C. T. (2021). Embedded decision support system for ultrasound nondestructive evaluation based on extreme learning machines. Computers \u0026amp; Electrical Engineering, 90, 106891.\u003c/li\u003e\n\u003cli\u003eHa, N., Xu, K., Ren, G., Mitchell, A., \u0026amp; Ou, J. Z. (2020). Machine learning‐enabled smart sensor systems. Advanced Intelligent Systems, 2(9), 2000063.\u003c/li\u003e\n\u003cli\u003eMuszynski, M., Gabrieli, G., Zimmerli, L., Temiz, Y., Heller, R., Cox, A., ... \u0026amp; Ruch, P. W. (2022, October). Live Demonstration: An AI-assisted e-tongue for fast and portable fingerprinting of liquids. In 2022 IEEE Sensors (pp. 1-1). IEEE.\u003c/li\u003e\n\u003cli\u003eDujić Rodić, L., Stančić, I., Zovko, K., Perković, T., \u0026amp; \u0026Scaron;olić, P. (2022). Tag estimation method for aloha rfid system based on machine learning classifiers. Electronics, 11(16), 2605.\u003c/li\u003e\n\u003cli\u003eWaskito, S., Leow, K. J., Medaranga, P., Gupta, T., Chakrabarty, S., Gulati, M., \u0026amp; Varshney, A. (2023, October). Otter: Simplifying embedded sensor data collection and analysis using large language models. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking (pp. 1-3).\u003c/li\u003e\n\u003cli\u003eDokic, K. (2020). Microcontrollers on the edge\u0026ndash;is esp32 with a camera ready for machine learning?. In Image and Signal Processing: 9th International Conference, ICISP 2020, Marrakesh, Morocco, June 4\u0026ndash;6, 2020, Proceedings 9 (pp. 213-220). Springer International Publishing.\u003c/li\u003e\n\u003cli\u003ePham, H. T., Nguyen, M. A., \u0026amp; Sun, C. C. (2019, December). AIoT solution survey and comparison in machine learning on low-cost microcontrollers. In 2019 International Symposium on Intelligent Signal Processing and Communication Systems (ISPACS) (pp. 1-2). IEEE.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"University of Johannesburg","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":"Machine Learning (ML), Micronutrient Sensing, Edge Computing, Quantized Models, Microcontroller-Class Hardware, Latency Optimization, Standardization \u0026 Benchmarking","lastPublishedDoi":"10.21203/rs.3.rs-6844555/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6844555/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eDeploying machine learning (ML) models on microcontroller-class hardware offers a promising pathway for real-time micronutrient sensing, especially in resource-constrained agricultural and environmental contexts. Traditional sensing methods are often cost-prohibitive and lack real-time responsiveness, while ML-embedded systems enable portable, low-power, and scalable monitoring. This systematic review investigates global research trends in applying ML on microcontroller-class hardware for micronutrient sensing. It evaluates algorithm choices, dataset characteristics, hardware specifications, performance reporting, and real-time capabilities to identify critical gaps and future opportunities. The review followed PRISMA 2020 guidelines, analyzing 43 studies published between 2015 and 2025 sourced from Scopus, Web of Science, and Google Scholar. Eligibility criteria included English-language, peer-reviewed works focusing on ML techniques for real-time micronutrient sensing using microcontroller platforms. Data were synthesized and visualized across 14 key dimensions, including model type, sensor integration, hardware constraints, and deployment scenarios. Publication activity peaked in 2022, with growing contributions from countries like Israel and Kenya. Journal articles (51.16%) and conference papers (37.21%) dominated. Most studies (46.51%) were sourced from Google Scholar. Established frameworks such as TinyML were most frequently used (39.53%), while 32.56% of studies specified exact microcontroller boards. Deep learning (37.21%) and hybrid models (20.93%) were commonly applied, often using custom datasets (39.53%). However, 46.51% of studies lacked clear model size or latency reporting. Real-time performance was confirmed in 65.12% of cases, though only 11.63% provided quantified size and latency data. Hardware constraints were often generalized (30.23%), and 16.28% of papers omitted hardware details altogether. Environmental monitoring and smart IoT applications were the most common use cases (25.58% and 13.95%, respectively), supported by domain-specific ML tools (25.58%). ML on microcontroller-class hardware shows clear potential for enabling accessible, real-time micronutrient sensing. However, reproducibility remains limited due to insufficient reporting on model performance, hardware specifics, and deployment conditions. To accelerate adoption, future work should prioritize standardization in performance reporting, interdisciplinary collaboration, and deployment in real-world field environments.\u003c/p\u003e","manuscriptTitle":"TinyML Applications in Micronutrient Sensing: A Review of Microcontroller Deployments","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-11 18:41:45","doi":"10.21203/rs.3.rs-6844555/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":49690886,"name":"Chemical Biology"},{"id":49690887,"name":"Agrochemicals"},{"id":49690888,"name":"Chemical Engineering"}],"tags":[],"updatedAt":"2025-06-11T18:41:45+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-11 18:41:45","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6844555","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6844555","identity":"rs-6844555","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}