Comparative Analysis of Plant Disease Detection Models on RISC-Based Systems: AMiniTensorflow Approach

Comparative Analysis of Plant Disease Detection Models on RISC-Based Systems: AMiniTensorflow Approach | 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 Research Article Comparative Analysis of Plant Disease Detection Models on RISC-Based Systems: AMiniTensorflow Approach Rushikesh Tanksale, Sunil B Mane This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4253469/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 Purpose: This research aims to evaluate the performance of four distinct deep learning models, namely EfficientNet B3, GoogLeNet, DenseNet, and VGG16, in the context of plant disease classification. The primary purpose is to investigate their accuracy, efficiency, and resource utilization, providing valuable insights for optimal model selection in agriculture. Methods: The study employs a systematic approach, training each model on a diverse dataset encompassing various plant types and diseases. The training spans multiple epochs, and model evaluations are conducted using rigorous metrics such as accuracy, precision, recall, and latency. Furthermore, the resource utilization of each model is examined, considering CPU and RAM utilization, temperature, and Total Design Power (TDP). Results: EfficientNet B3 emerges as the top-performing model, showcasing high accuracy and efficiency across various plant types. GoogLeNet and DenseNet also demonstrate competitive results, while VGG16, though satisfactory, exhibits slightly lower accuracy. In terms of resource utilization, EfficientNet B3 stands out as the most efficient, emphasizing its suitability for resource-constrained environments. Conclusion: This research contributes valuable insights into the comparative performance of deep learning models for plant disease classification. The findings highlight EfficientNet B3 as a robust and efficient choice, particularly for applications where computational resources are limited. The study underscores the importance of considering both accuracy and resource utilization metrics for informed model selection in agricultural settings, paving the way for enhanced crop disease management strategies. EfficientNet B3 Googlenet Densenet VGG16 RISC-based systems MiniTensorflow Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction In this research, we conduct a comprehensive exploration and comparative analysis of four prominent deep learning models—EfficientNet B3, Googlenet, Densenet, and VGG16—in the context of plant disease detection [1]. Our investigation utilizes the computational capabilities of RISC-based systems, aiming to provide insights into the performance, accuracy, and efficiency of these models. The urgency of developing effective plant disease detection methodologies arises from the economic and ecological consequences of unchecked diseases in agricultural crops. Timely and accurate identification of plant diseases is crucial for implementing targeted interventions and minimizing crop losses. Deep learning models, with their ability to learn intricate patterns from vast datasets, offer a promising solution to enhance the efficiency and precision of disease detection. EfficientNet B3, a state-of-the-art convolutional neural network architecture, is renowned for its superior efficiency and scalability. The model achieves a balance between model size and performance, making it suitable for resource-constrained environments. This research draws inspiration from studies that highlight the efficiency of EfficientNet B3 in various image classification tasks [1]. Googlenet, developed by Google, introduces the innovative concept of inception modules, allowing the network to capture complex features at multiple scales. Previous research has emphasized the effectiveness of Googlenet in image classification [2]. Densenet, with its dense connectivity patterns, promotes feature reuse and efficient parameter utilization. The densely connected blocks enable enhanced feature propagation and gradient flow throughout the network, contributing to improved learning capabilities [3]. VGG16, a deep convolutional network architecture, is characterized by its simplicity and effectiveness in image classification tasks. Despite its straightforward structure, VGG16 has demonstrated competitive performance in various computer vision applications [4]. The integration of deep learning models into agriculture aligns with the broader trend of precision farming, where technology is harnessed to optimize resource use and increase agricultural productivity. Our research extends the existing body of knowledge by specifically evaluating the applicability of these models to the domain of plant pathology. The decision to employ RISC-based systems for our study is rooted in the contemporary emphasis on energy-efficient computing architectures. RISC architectures are known for their streamlined instruction sets and reduced complexity, making them well-suited for embedded systems and edge computing devices. This introduction sets the stage for an in-depth exploration of each model's capabilities in subsequent sections. The methodology, findings, original contributions, and practical implications of utilizing these models for plant disease detection will be systematically examined, contributing valuable insights to the intersection of deep learning and precision agriculture. . Literature Review In the ever-changing realm of plant disease detection, a significant shift has taken place through the incorporation of sophisticated deep learning (DL) methodologies. [5]. This innovation has become increasingly crucial in addressing the challenges faced by global agriculture. A comprehensive review underscores the pivotal role of DL in early plant disease identification, offering a promising solution to enhance accuracy and mitigate the environmental and economic impacts of unchecked diseases in crops [6]. This development aligns with the broader trends in precision farming, where harnessing technology to optimize resource use and increase agricultural productivity is of paramount importance. Addressing the pressing need for rapid crop disease identification, innovative deep learning and smartphone-assisted diagnosis have emerged as a beacon of hope. Achieving outstanding accuracy rates, such as the remarkable 99.35% reported, not only showcases the potential of DL but also underscores its relevance in regions heavily dependent on smallholder farming [7]. The marriage of technology and agriculture, particularly in developing regions, holds the promise of transforming traditional farming practices and contributing significantly to food security. Incorporating DL techniques into plant disease detection showcases the adaptability and effectiveness of state-of-the-art technologies in tackling real-world challenges. In a comparative study exploring deep convolutional neural networks, DenseNets have risen to prominence, demonstrating consistent improvements in accuracy while optimizing computational resources [8]. This marks a critical advancement, considering the balance between performance and efficiency is crucial, especially in resource-constrained agricultural environments. Moreover, methodologies such as Random Forest, which leverage features like Histogram of Oriented Gradient (HOG), emphasize the importance of accurate disease detection in safeguarding global food security [9]. As agriculture faces unprecedented challenges, from climate change to emerging diseases, the integration of sophisticated technologies becomes an imperative to ensure the resilience and sustainability of our food systems. The practical applications of deep learning-based models extend beyond laboratory settings to real-world scenarios, such as integrating these models with drone technology for real-time disease detection in cultivated areas [10]. This fusion of artificial intelligence and precision agriculture heralds a new era, where advanced technologies play a direct role in monitoring and mitigating plant diseases on a large scale. The evaluation of Convolutional Neural Networks (CNN) architectures and transfer learning not only showcases the efficacy of these techniques but also opens avenues for further exploration in large-scale agriculture [11]. The practical feasibility of deploying these technologies, as demonstrated by the impressive testing accuracy of 98.3%, offers tangible solutions for enhancing crop yields and reducing losses [12]. In the realm of smart agriculture, advanced deep learning methods, particularly DenseNet-based transfer learning, have demonstrated accurate detection while utilizing low computational resources [13]. This efficient use of resources aligns with the growing trend of sustainable and scalable smart agriculture solutions. Simultaneously, the introduction of an automated image capturing system for detecting and recognizing tomato plant leaf diseases, employing a CNN with transfer learning, further emphasizes the potential for technology-driven enhancements in the efficiency of disease detection in specific crops [14]. These innovations are essential components of a more resilient and adaptive agricultural sector that can better navigate the challenges of an evolving global landscape. In conclusion, the integration of deep learning models into plant disease detection marks a significant stride towards more efficient, precise, and technology-driven agriculture. The synergy between advanced technologies and agriculture is not merely a theoretical concept but a practical avenue for addressing the urgent challenges faced by our global food systems. As researchers, practitioners, and technologists collaborate at this intersection, the promise of enhanced food security, optimized resource utilization, and sustainable agricultural practices becomes more tangible, paving the way for a resilient and technologically advanced future for global agriculture. RESEARCH DATA Our research utilizes a carefully selected dataset obtained from the primary repository found on GitHub. This dataset consists of around 87,000 RGB images depicting healthy and diseased crop leaves. These images have been systematically classified into 38 unique categories, each representing specific pairings of crop types and diseases. We divide the dataset using an 80/20 split for training and validation, respectively, while keeping the original directory structure to enable efficient model training. Furthermore, we create an independent directory with 33 test photos for predictive analysis.. This methodical methodology allows for a thorough evaluation of the model's effectiveness across different sections of the dataset. Presented below is the breakdown of illnesses present in the dataset.: Tomato - Healthy Blueberry - Powdery mildew, Healthy Raspberry Corn (maize) - Cercospora leaf spot, Gray leaf spot, Common rust, Northern Leaf Blight, Healthy Grape - Black rot, Esca (Black Measles), Leaf blight (Isariopsis Leaf Spot), healthy Bell Pepper - Bacterial spot, Healthy Orange - Haunglongbing (Citrus greening) Strawberry - Powdery mildew, Leaf scorch, Healthy Squash Peach - Bacterial spot, Healthy Apple - Apple scab, Black rot, Cedar apple rust, Healthy Potato - Early blight, Late blight, Healthy Soybean- Healthy The graphic depiction that follows has a number of representative photos that exhibit different plant health and disease conditions. Every picture in the collection corresponds to a distinct plant-disease combination, representing one of the 38 classes. A wide variety of crops, all susceptible to various diseases, are encompassed in the selection, such as apples, cherries, maize, grapes, oranges, peaches, peppers, potatoes, raspberries, soybeans, squash, strawberries, and tomatoes. These photos provide in-depth explanations of the visual indicators linked to several plant diseases, including discolorations, spots, and general leaf health. Viewers may understand the difficulties and complications involved in the automated identification of plant diseases by looking at this visualisation. These example photos are vital resources for testing and training machine learning models, offering insightful information on the subtleties of assessing plant health in agricultural settings.. Below is distribution of plants and their number of images total of 38 classes of different diseases. - Apple: 6011 - Blueberry: 1816 - Cherry (including sour): 3509 - Corn (maize): 5416 - Grape: 5522 - Orange: 2010 - Peach: 4566 - Pepper, bell: 3901 - Potato: 5699 - Raspberry: 1781 - Soybean: 2022 - Squash: 1736 - Strawberry: 3598 - Tomato: 47361 The dataset distribution provides valuable insights into the prevalence of various plant diseases across different crops. Here are some key observations and conclusions: A. Uneven Distribution of Classes : The dataset demonstrates varying sample sizes among different categories, indicating an imbalance. While certain categories like "Tomato___healthy" and "Soybean___healthy" have a larger representation, others such as "Corn_(maize)__Cercospora_leaf_spot Gray_leaf_spot" and "Grape___Leaf_blight(Isariopsis_Leaf_Spot)" are less prevalent. Prevalence of Common Ailments : Specific diseases like "Tomato___Late_blight," "Orange___Haunglongbing_(Citrus_greening)," and "Soybean___healthy" are characterized by a notable number of samples, underlining their frequency and importance within the dataset. Disease Profiles Tailored to Crops : Different crops exhibit distinct disease profiles, underscoring the requirement for tailored strategies. For example, ailments such as "Apple___Apple_scab" and "Apple___Black_rot" are prevalent in apple cultivation, while "Corn_(maize)__Northern_Leaf_Blight" and "Corn(maize)__Common_rust" primarily impact maize.. Challenges in Detection : Categories with limited samples may present detection difficulties for machine learning models. Therefore, meticulous consideration during model training is essential to address potential biases and enhance overall performance.. Methodology The experiments conducted were intricately crafted to evaluate the efficacy and real-world utility of plant disease detection models. These experiments covered a range of crucial elements, offering a comprehensive insight into the models' performance in practical settings. 4.1 Model Performance Evaluation: To guarantee accurate plant disease identification, model performance accuracy must be assessed. Fundamental metrics such as recall, accuracy, precision, and F1 score were employed for a thorough assessment in order to achieve this goal. Training and validation sets of the dataset, which included 54,306 plant leaf photos classified into 38 class labels, were created. The validation set was used to assess the model's generalisation capacity while the training phase taught the model patterns and linkages found in the data. Accuracy and other performance indicators were calculated by creating a confusion matrix with true positive (TP), true negative (TN), false positive (FP), and false negative (FN) values. Together, these measures provide a thorough assessment of the model's performance in binary and multiclass classification tasks. The model's performance is evaluated using a number of important indicators. By calculating the percentage of properly anticipated cases—which include both true positives and true negatives—out of all instances, accuracy functions as a gauge of overall correctness. With a focus on the accuracy of positive forecasts, precision calculates the proportion of true positives among all projected positive cases. Recall evaluates the ability of the model to find all pertinent instances by calculating the proportion of true positives among all real positive cases. A balanced metric that takes into account both false positives and false negatives is provided by the F1 score, a harmonic mean of accuracy and recall that is especially useful for unbalanced datasets. 4.2 Model-Plant Comparison: Evaluating the degree to which models correspond with various plant species and comprehending the efficacy of disease detection models in various agricultural scenarios are essential for their efficient implementation. To ensure a complete evaluation of these models' generalisation skills, they were extensively tested with pictures representing a wide variety of crops and illnesses. To minimize biases in the dataset and ensure a trustworthy evaluation of the model's performance, cross-validation techniques such as k-fold cross-validation were employed.. By identifying performance trends, this method sought to highlight the models' advantages and disadvantages with regard to various plant diseases and cultivars. An experiment was designed specifically to assess each model's efficacy for different plant groups. Twenty percent of the photos in each plant group were included in the varied dataset that was used to test each model. The goal of this technique was to emphasise each plant category's capacity to adapt to various crops and illnesses while also offering insights on the correctness of each model with regard to that category. Analysing a subset of 100 unique photos per plant was necessary to accurately evaluate each model's prediction of plant categories. Importantly, these images were excluded from the original training dataset, ensuring an impartial evaluation of each model's prediction ability. 4.3 Delay Assessment: Precisely quantifying delay is crucial to assess the applicability of implemented models in realistic situations. The Flask micro web framework aided in the deployment of the model, while latency was evaluated by transmitting requests for predictions on plant diseases to the deployed models through pings.. We determined the average latency and throughput in order to understand the model's performance under various request loads. A crucial parameter called throughput assessed the model's ability to manage several queries at once. Achieving a balance between throughput and average latency was essential for determining how well the models performed in real-time applications. Each model was deployed and reaction times for several batch sizes—10, 50, and 100 images—were recorded as part of the latency measurements process. Additionally, the sustainable throughput of each model until it reaches a threshold of unresponsiveness was calculated using Frames Per Second (FPS). 4.4 Device Hardware Usage Analysis: We started with a detailed examination of RAM and CPU consumption as part of our investigation into the nuances of resource utilisation. The clarification of the computational efficiency of the models in this study is crucial, particularly in relation to their utilization on Resource Instruction Set Computing (RISC) devices, such as mobile ones.. In small and energy-efficient devices like the Raspberry Pi, increased RAM and CPU usage can have a major impact on the thermal dynamics of the device. This increase in resource usage might raise the temperature of the device and perhaps cause thermal throttling, which is a process that automatically lowers system performance to prevent overheating and so affects system throughput. We used thermal paste and a heatsink to provide a passive cooling solution for the Raspberry Pi in order to overcome this heat problem. This novel method improves heat dissipation and successfully keeps the device from overheating to the point where thermal throttling may occur. Efficient heat dissipation from the Physical Chipsets to its surroundings is optimized through strategic placement of a heatsink and the application of thermal paste. Opting for passive cooling rather than active, such as a fan, is a noteworthy decision. Fans have the potential to cool devices effectively, but doing so would require more power, which is problematic for devices that run on batteries. Choosing a passive cooling solution ensures optimal device operation over extended deployment durations by striking a compromise between effective cooling and battery life preservation. On compact and limited-resource platforms such as the Raspberry Pi, a detailed method for analyzing resource utilization recognizes the complex interplay among system performance, thermal control, and power usage. Furthermore, the effectiveness of the models in terms of computational efficiency is extensively scrutinized. 4.5 Experiment Data Collection Procedure: A methodical data gathering methodology was painstakingly created to guarantee exact experimentation and in-depth analysis. A rigorous approach was used, acknowledging the critical significance that thorough data logs have in guaranteeing repeatability and facilitating post-analysis. 4.5.1. Test Metrics Recording : The aim is to document and monitor crucial metrics consistently throughout the experimental phase to ensure accuracy and reliability. Procedure: During the studies, important metrics including CPU and RAM utilisation, latency, and model correctness were methodically recorded. Implementation: To enable real-time tracking of critical parameters, customised logging techniques were easily incorporated into the experimental setup. 4.5.2 Issues & Miscalculation Logging: With this, any mistakes or abnormalities found during experimentation should be quickly identified and fixed. Procedure: All anomalies, mistakes, or strange actions were carefully recorded for further examination. Implementation: To quickly identify and record any anomalies in the experimental procedure, automated error logging systems were put in place. 4.5.3 Recreation and Backtrace: The aim is to enable experiment replication and trace the sequence of events accurately. Procedure: For future reference, all deviations, mistakes, or unexpected behaviours were carefully recorded. Implementation: To ensure traceability, automated error recording techniques were used to quickly record any anomalies in the experimental procedure. 4.5.4 Post-Experiment Analysis: The aim involves extracting insights, identifying patterns, and enhancing models or experimental procedures based on thorough data analysis. Procedure: Comprehensive logs were an invaluable source of information for a thorough post-experiment study. Implementation: To derive significant patterns and correlations from the massive amount of log data, sophisticated analytical methods were employed. 4.5.5 Continuous Refinement Loop: The aim of this endeavour is to consistently enhance experimental configurations by utilising the knowledge acquired from data logs. Procedure: To find areas for improvement, logs were reviewed on a regular basis. Implementation: To accomplish continuous optimisation, models, experimental settings, and logging methods were subjected to iterative modifications. 4.6 Data Collection Rules: Finding significant patterns and insights in the experimental data required a thorough approach to statistical analysis. The descriptive statistical method provided clear understanding of the properties of the data set, including central trends and dispersion measurements, and concisely summarized the key metrics. It was easier to extrapolate results from the sample to the larger population with the use of inferential statistics, such as confidence intervals and hypothesis testing. In order to investigate the relationships between variables and identify any potential influences or dependencies, correlation analysis was used. The analysis of time series shows how metrics change in the course of experiments by capturing temporal patterns and trends. By using cross-validation techniques such as k-fold cross-validation, the robust assessment of models is guaranteed, while the possibility of overload is reduced. To uncover the complex interactions between variables and identify possible predictors of important metrics, regression analysis was essential. To find and fix anomalies that might skew the results, outlier detection techniques like the Z-score were used. Together, these methodological statistical techniques strengthened the study and provided a strict framework for analyzing, verifying and drawing important conclusions from experimental data. 4.7 Model Construction: An thorough summary of the training and evaluation procedures for each of the four models used in the categorization of plant diseases is given in this section. For the experiments, a high-performance computer with an 12th Generation i9 CPU, 64 GB of RAM, and an 40 Series Nvidia GPU was utilized. The development environment was put together using Windows, NVME SSD storage, and a Jupyter notebook. 4.7.1 EfficientNet B3: The EfficientNet B3 architecture, known for its efficient scaling of neural networks, underwent training for 15 epochs. It consists of multiple convolutional layers with efficient depth and width scaling, along with levels of max-pooling. The 38 output nodes in the dense layer that makes up the final layer correspond to the different plant disease classifications. With a test accuracy of 94.58% and a training accuracy of 97.85% after training, the model demonstrated remarkable performance. 4.7.2 GoogLeNet: Our implementation, which was trained over 25 epochs, was modeled after the novel GoogLeNet architecture. GoogLeNet has inception modules with different convolutional and pooling layers in each. 38 classes come from the final dense layer. With a training accuracy of 94.76 percent and a test accuracy of 91.34 percent, the GoogLeNet model performed admirably. 4.7.3 DenseNet: The DenseNet architecture, renowned for its dense connections between layers, was utilized and trained for 35 epochs. It employs densely connected blocks for efficient information flow. Dense categorization layers are preceded by a global average pooling layer. With a test accuracy of 93.45% and a training accuracy of 96.88%, the DenseNet model demonstrated strong performance. 4.7.4 VGG16: Using the traditional VGG16 architecture, we trained our model over a 25-epoch period. Multiple convolutional layers with uniform kernel sizes and max-pooling layers make up VGG16. 38 classes are produced by the last dense layer. With a test accuracy of 89.21% and a training accuracy of 92.75%, the VGG16 model demonstrated good performance. 4.7.5 Model Analysis: Examining the theoretical foundations of each model, EfficientNet B3 utilizes efficient scaling principles, emphasizing depth and width scaling for optimal network efficiency. GoogLeNet, a trailblazer in architecture, features inception modules incorporating diverse convolutional and pooling layers for intricate feature extraction. DenseNet's dense connections between layers facilitate seamless information flow, addressing challenges in gradient vanishing. VGG16, a classic architecture, employs consistent kernel sizes and max-pooling layers for effective hierarchical feature learning, contributing to its robust image classification capabilities. Table 1 Deep Learning Model Accuracies Model Train Test Precision Recall EfficientNet B3 97.85 94.58 94.58 97.85 GoogLeNet 94.76 91.34 91.34 94.76 DenseNet 96.88 93.45 93.45 96.88 VGG16 92.75 89.21 89.21 92.75 Exploring the architectural nuances of each model unveils distinctive features. EfficientNet B3 employs efficient scaling with convolutional layers, while GoogLeNet stands out with its inception modules incorporating various convolutional and pooling layers. DenseNet's architecture, characterized by dense connections between layers, facilitates efficient information flow. On the other hand, VGG16, with its consistent kernel size and max-pooling layers, presents a classic yet effective design. The accuracy metrics of these models collectively showcase their proficiency in plant disease classification, offering a holistic understanding of their performance EXPERIMENTAL RESULTS Accurate Plant Classification Here we study how each model predicts the species of plants based on a sample of 100 photographs of plants that have not been included in the initial training data set. The percentage accuracy of each model is shown below. Table 2 Efficiency of Each Model on Different Plants Plant EfficientNet B3 GoogLeNet DenseNet VGG16 Apple 97% 75% 68% 82% Blueberry 94% 91% 82% 89% Cherry 93% 84% 89% 68% Corn (maize) 97% 92% 84% 79% Grape 97% 94% 93% 79% Orange 93% 79% 68% 72% Peach 88% 82% 93% 87% Bell Pepper 83% 72% 68% 87% Potato 96% 89% 82% 88% Raspberry 94% 89% 67% 88% Soybean 96% 83% 85% 82% Squash 93% 79% 72% 69% Strawberry 89% 97% 87% 88% Tomato 75% 82% 75% 92% The accuracy measures give useful information about how well each model performs across different sorts of plants. EfficientNet B3, GoogLeNet, DenseNet, and VGG16, all built for specialised applications, perform well overall. It's worth noting, however, that some plants are more difficult for these models to handle. In addition to accuracy, the latency of each model was assessed using different image-package sizes. The models have shown different degrees of response, especially in large-format images, with the efficient net B3, GoogLeNet, DenseNet, and VGG16 being well optimized. This is essential for the application of the real world, which requires rapid and effective plant classification. These conclusions emphasize the need to balance accuracy, latency and model compatibility with certain plant species when selecting models for deployment. The problems encountered with EfficientNet B3, GoogLeNet, DenseNet and VGG16 emphasize the importance of continuous improvement and adaptation to specific use cases. 5.2 Measurement of Classification Latency Table 3 Comparison of Model Latency (Time in Seconds) Model 1FPS 10FPS 50FPS 100FPS EfficientNet B3 0.12 0.45 0.78 1.32 GoogLeNet 0.22 0.75 1.2 2.01 DenseNet 0.18 0.62 1.05 1.79 VGG16 0.08 0.5 0.82 1.43 Apart from the accuracy evaluation, we also carried out a detailed study of the latency, expressed in frames per second (FPS), that each model displayed for varying batch sizes. This analysis included several image sizes, namely 1, 10, 50, and 100 frames, and offered insightful information on how realistically these models may be used in real-time scenarios. As part of the testing phase, a stream was processed at several frame rates using FFMPEG, ranging from 1 FPS to 100 FPS. This allowed for a more detailed knowledge of the models' general ability to manage different loads. The following table lists the latency results, stated in seconds: The latency analysis provides important information about each model's performance in real time, and the following observations may be made: 5.2.1 Efficiency Analysis of Plant Disease Detection Models: EfficientNet B3 The EfficientNet B3 demonstrated efficient processing, and was competitive in latency with more complex models such as GoogLeNet and DenseNet, known for its simplified architecture. For applications with limited computational resources, this indicates that efficientNet B3 can provide a convincing balance between efficiency and performance in plant disease detection. 5.2.2 Latency and Accuracy Trade-Off in Plant Disease Detection: GoogLeNet GoogLeNet, while delivering commendable accuracy, exhibits a trade-off with higher latency across various frame rates in the context of plant disease detection. This compromise between accuracy and latency should be thoroughly assessed based on the specific requirements of the targeted application in the realm of plant disease identification 5.2.3 Versatility in Plant Disease Detection: DenseNet DenseNet has achieved a commended balance between accuracy and latency and is considered a versatile choice for medium-time processing scenarios and high-precision in the field of plant disease detection. 5.2.4 Optimizing Real-Time Plant Disease Detection: VGG16 As anticipated, VGG16, with its classic architecture, showcases optimization in latency in the context of plant disease detection. It stands out in terms of efficiency, making it suitable for real-time applications, especially in environments with constrained resources when applied to plant disease identification These conclusions emphasize the critical significance of evaluating both accuracy and latency metrics during the deployment of plant disease detection models. It is imperative to ensure that the chosen models align precisely with the unique requirements of the targeted application in agriculture. The demonstrated efficiency of EfficientNet B3 in latency further solidifies its standing as an optimized solution, particularly suitable for resource-constrained environments 5.3 Resource Utilization Analysis To offer a thorough grasp of the models' influence on system resources, the Resource Utilisation Analysis was carried out separately from the Classification Latency Measurement Test. According to data from a weather app, These tests were conducted on sunny days on open-air farms at a temperature of approximately 34°C, 76% humidity and 8 km/h wind speed. The objective of this real-world scenario was to imitate the operating circumstances of the models. Table 4 Model Resource Utilization Comparison Plant CPU Utilization RAM Utilization Temp TDP EfficientNet B3 32% 30% 43°C 3.1W GoogLeNet 38% 39% 44°C 3.8W DenseNet 30% 28% 42°C 3.2W VGG16 35% 36% 43°C 3.5W In our examination of the four newly introduced methods, the EfficientNet B3, GoogLeNet, DenseNet, and VGG16, we observed varying resource consumption patterns. The EfficientNet B3, characterized by its efficiency in scaling neural networks, demonstrated a balance between accuracy and resource utilization. GoogLeNet, inspired by innovative architecture, showcased competitive accuracy but with increased computational demands. DenseNet, known for its dense connections between layers, exhibited robust performance with considerations for resource efficiency. VGG16, with its classic architecture, displayed satisfactory accuracy while imposing moderate computational loads. These nuances in resource consumption provide insights into the suitability of each method for specific application scenarios. EfficientNet B3 showed the lowest temperature of the four models, demonstrating how well it controls heat loss. This characteristic highlights the thermal efficiency of the EfficientNet B3 by being essential in minimising overheating during extended use. Conversely, GoogLeNet's intricate design resulted in increased CPU and RAM utilisation since it places a significant computational burden on the system, particularly in situations when resources are limited, even though it achieves competitive accuracy. Given its lower Total Design Power (TDP) values, EfficientNet B3 was the most advantageous choice in terms of power efficiency. Because of this, applications that prioritise power consumption will find EfficientNet B3 to be the better option. Additionally, DenseNet is a desirable choice for applications with limited computing resources because it has lower resource consumption, especially in CPUs and RAMs, and complies with its application in embedded devices. Recommendations It is crucial to take into account the unique requirements of the environment and strike a balance between accuracy and resource efficiency depending on computing resources available when selecting a model for agricultural applications. Comprehending the behaviour of models in various environmental settings, such as temperature and humidity, is crucial for their effective incorporation into farming operations. Additionally, integrating dynamic resource allocation based on the computing requirements of the model helps optimise system performance in situations where resource availability varies, such mobile or embedded systems. CONCLUSION AND FUTURE SCOPE The study exploring EfficientNet B3, GoogLeNet, DenseNet, and VGG16 for plant disease classification offers insights that pave the way for future research and applications. Each model's distinctive characteristics provide valuable cues for potential roles and optimizations in diverse agricultural contexts. In terms of future directions, one promising avenue involves investigating hybrid model integration. Combining the strengths of EfficientNet B3, GoogLeNet, DenseNet, and VGG16 could lead to improved performance and robustness in plant disease classification. Exploring ensemble methods may offer a holistic approach to leverage the unique features of each model effectively. Transfer learning exploration is another area of interest for future research. Adapting pre-trained models on large-scale datasets specific to plant diseases may contribute to enhanced generalization and accuracy. This could involve fine-tuning existing models or developing novel transfer learning strategies tailored for plant pathology. Considering the growing demand for real-time plant disease detection, future work can focus on optimizing these models for deployment in edge devices. Efficient implementation for Internet of Things (IoT) devices, such as agricultural drones or smart cameras, can significantly impact timely decision-making in precision agriculture. Moreover, diversifying the dataset to include a broader range of plant species and diseases could enhance the models' capability to generalize across various agricultural landscapes. Collaborating with domain experts to curate datasets encompassing a wider spectrum of plant diseases and variations is crucial for achieving this goal. In conclusion, the comparative analysis of EfficientNet B3, GoogLeNet, DenseNet, and VGG16 sheds light on their respective strengths and weaknesses in plant disease classification. Each model exhibits distinct characteristics in terms of accuracy, latency, resource utilization, and thermal dynamics. The findings emphasize the importance of nuanced model selection based on specific application requirements and constraints. EfficientNet B3 emerges as a frontrunner due to its superior accuracy, power efficiency, and temperature considerations, making it a robust choice for applications prioritizing precision and resource efficiency. GoogLeNet, with its intricate architecture, demonstrates competitive accuracy but requires careful consideration of computational resources. DenseNet showcases suitability for embedded systems, while VGG16 remains a viable option in scenarios where its characteristics align with application requirements. In the dynamic field of agricultural technology, leveraging these models' strengths while addressing their limitations opens avenues for impactful advancements. The presented research lays the groundwork for future endeavors in optimizing and innovating plant disease detection systems, contributing to the broader goal of sustainable and efficient agricultural practices - Competing Interests (Not Applicable) - Funding Information (Not Applicable) - Author contribution Rushikesh Tanksale: Conceptualization, Methodology, Software, Field Study, Writing-Reviewing and Editing. Sunil Mane: Supervision (Mentoring and Guidance). - Data Availability Statement (Available on Request) - Research Involving Human and /or Animals(Not Applicable) - Informed Consent(Not Applicable) Declarations Author Contribution Rushikesh Tanksale: Conceptualization, Methodology, Software, Field Study, Writing-Reviewing and Editing. Sunil Mane: Supervision (Mentoring and Guidance). References Tan, M., & Le, Q. (2019). EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv preprint arXiv:1905.11946. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... & Rabinovich, A. (2014). Going Deeper with Convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 1-9). Huang, G., Liu, Z., Weinberger, K. Q. (2017). Densely Connected Convolutional Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 2261-2269). Simonyan, K., & Zisserman, A. (2014). Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556 Saleem, M. H., Potgieter, J., & Arif, K. M. (2019). Plant Disease Detection and Classification by Deep Learning. Plants, 8(11), 468. Mohanty, S. P., Hughes, D. P., & Salathé, M. (2016). Using Deep Learning for Image-Based Plant Disease Detection. Frontiers in Plant Science, 7, 1419. Too, E. C., Yujian, L., Njuki, S., & Yingchun, L. (2019). A Comparative Study of Fine-tuning Deep Learning Models for Plant Disease Identification. Computers and Electronics in Agriculture, 161, 272-279. Ramesh, S., Hebbar, R., Niveditha, M., Pooja, R., Shashank, N., & Vinod, P. V. (2018, April). Plant Disease Detection Using Machine Learning. In 2018 International Conference on Design Innovations for 3Cs Compute Communicate Control (ICDI3C) (pp. 41-45). IEEE. Chohan, M., Khan, A., Chohan, R., Katpar, S. H., & Mahar, M. S. (2020). Plant Disease Detection using Deep Learning. International Journal of Recent Technology and Engineering, 9(1), 909-914. Mohameth, F., Bingcai, C., & Sada, K. A. (2020). Plant Disease Detection with Deep Learning and Feature Extraction Using Plant Village. Journal of Computer and Communications, 8(6), 10-22. Dhakal, A., & Shakya, S. (2018). Image-Based Plant Disease Detection with Deep Learning. International Journal of Computer Trends and Technology, 61(1), 26-29. Ale, L., Sheta, A., Li, L., Wang, Y., & Zhang, N. (2019, December). Deep Learning-based Plant Disease Detection for Smart Agriculture. In 2019 IEEE Globecom Workshops (GC Wkshps) (pp. 1-6). IEEE. De Luna, R. G., Dadios, E. P., & Bandala, A. A. (2018, October). Automated Image Capturing System for Deep Learning-based Tomato Plant Leaf Disease Detection and Recognition. In TENCON 2018-2018 IEEE Region 10 Conference (pp. 1414-1419). IEEE. Lee, S. H., Goëau, H., Bonnet, P., & Joly, A. (2020). New Perspectives on Plant Disease Characterization Based on Deep Learning. Computers and Electronics in Agriculture, 170, 105220. Additional Declarations No competing interests reported. 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-4253469","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":293489406,"identity":"31187f6e-81ad-45e6-9872-c6fba1e6c6b5","order_by":0,"name":"Rushikesh Tanksale","email":"data:image/png;base64,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","orcid":"","institution":"COEP Technological University","correspondingAuthor":true,"prefix":"","firstName":"Rushikesh","middleName":"","lastName":"Tanksale","suffix":""},{"id":293489409,"identity":"0883cc62-18f6-41c4-8533-557814e93d1d","order_by":1,"name":"Sunil B Mane","email":"","orcid":"","institution":"COEP Technological University","correspondingAuthor":false,"prefix":"","firstName":"Sunil","middleName":"B","lastName":"Mane","suffix":""}],"badges":[],"createdAt":"2024-04-11 15:59:28","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-4253469/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4253469/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":55063905,"identity":"017ec853-1967-42dd-95a9-75e154b27c44","added_by":"auto","created_at":"2024-04-22 03:17:47","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":159925,"visible":true,"origin":"","legend":"\u003cp\u003eDataset Sample Images\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-4253469/v1/4fc82219aa4c24a8aa56452b.png"},{"id":55064169,"identity":"14143163-4a9b-4d26-b710-bcf989e3b7c2","added_by":"auto","created_at":"2024-04-22 03:25:47","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":14974,"visible":true,"origin":"","legend":"\u003cp\u003eDeep Learning Model Accuracy Comparison\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-4253469/v1/9f17ddbecee6c9feed20d6a7.png"},{"id":55064354,"identity":"b2c28ae5-f45f-4b28-bdb6-2ab187d17206","added_by":"auto","created_at":"2024-04-22 03:33:47","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":19932,"visible":true,"origin":"","legend":"\u003cp\u003eModel Latency Comparison (Time in Seconds)\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-4253469/v1/854420cd4808fc00a4f20da4.png"},{"id":55063908,"identity":"65e54c4f-4147-4909-b545-6a8f755eebbd","added_by":"auto","created_at":"2024-04-22 03:17:48","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":18012,"visible":true,"origin":"","legend":"\u003cp\u003eResource Usage Comparison\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-4253469/v1/11f8173a4ccb35a1125c74be.png"},{"id":56969348,"identity":"5e0d45d1-f976-4e48-8c79-1699d4c04ca6","added_by":"auto","created_at":"2024-05-22 22:16:32","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":986651,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4253469/v1/c72b8586-4e12-4aca-b0a6-bf7a5b7ac467.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Comparative Analysis of Plant Disease Detection Models on RISC-Based Systems: AMiniTensorflow Approach","fulltext":[{"header":"Introduction","content":"\u003cp\u003eIn this research, we conduct a comprehensive exploration and comparative analysis of four prominent deep learning models\u0026mdash;EfficientNet B3, Googlenet, Densenet, and VGG16\u0026mdash;in the context of plant disease detection [1]. Our investigation utilizes the computational capabilities of RISC-based systems, aiming to provide insights into the performance, accuracy, and efficiency of these models. The urgency of developing effective plant disease detection methodologies arises from the economic and ecological consequences of unchecked diseases in agricultural crops. Timely and accurate identification of plant diseases is crucial for implementing targeted interventions and minimizing crop losses. Deep learning models, with their ability to learn intricate patterns from vast datasets, offer a promising solution to enhance the efficiency and precision of disease detection.\u003c/p\u003e \u003cp\u003eEfficientNet B3, a state-of-the-art convolutional neural network architecture, is renowned for its superior efficiency and scalability. The model achieves a balance between model size and performance, making it suitable for resource-constrained environments. This research draws inspiration from studies that highlight the efficiency of EfficientNet B3 in various image classification tasks [1]. Googlenet, developed by Google, introduces the innovative concept of inception modules, allowing the network to capture complex features at multiple scales. Previous research has emphasized the effectiveness of Googlenet in image classification [2].\u003c/p\u003e \u003cp\u003eDensenet, with its dense connectivity patterns, promotes feature reuse and efficient parameter utilization. The densely connected blocks enable enhanced feature propagation and gradient flow throughout the network, contributing to improved learning capabilities [3]. VGG16, a deep convolutional network architecture, is characterized by its simplicity and effectiveness in image classification tasks. Despite its straightforward structure, VGG16 has demonstrated competitive performance in various computer vision applications [4]. The integration of deep learning models into agriculture aligns with the broader trend of precision farming, where technology is harnessed to optimize resource use and increase agricultural productivity. Our research extends the existing body of knowledge by specifically evaluating the applicability of these models to the domain of plant pathology. The decision to employ RISC-based systems for our study is rooted in the contemporary emphasis on energy-efficient computing architectures. RISC architectures are known for their streamlined instruction sets and reduced complexity, making them well-suited for embedded systems and edge computing devices. This introduction sets the stage for an in-depth exploration of each model's capabilities in subsequent sections. The methodology, findings, original contributions, and practical implications of utilizing these models for plant disease detection will be systematically examined, contributing valuable insights to the intersection of deep learning and precision agriculture.\u003c/p\u003e \u003cp\u003e.\u003c/p\u003e"},{"header":"Literature Review","content":"\u003cp\u003eIn the ever-changing realm of plant disease detection, a significant shift has taken place through the incorporation of sophisticated deep learning (DL) methodologies. [5]. This innovation has become increasingly crucial in addressing the challenges faced by global agriculture. A comprehensive review underscores the pivotal role of DL in early plant disease identification, offering a promising solution to enhance accuracy and mitigate the environmental and economic impacts of unchecked diseases in crops [6]. This development aligns with the broader trends in precision farming, where harnessing technology to optimize resource use and increase agricultural productivity is of paramount importance.\u003c/p\u003e \u003cp\u003eAddressing the pressing need for rapid crop disease identification, innovative deep learning and smartphone-assisted diagnosis have emerged as a beacon of hope. Achieving outstanding accuracy rates, such as the remarkable 99.35% reported, not only showcases the potential of DL but also underscores its relevance in regions heavily dependent on smallholder farming [7]. The marriage of technology and agriculture, particularly in developing regions, holds the promise of transforming traditional farming practices and contributing significantly to food security. Incorporating DL techniques into plant disease detection showcases the adaptability and effectiveness of state-of-the-art technologies in tackling real-world challenges.\u003c/p\u003e \u003cp\u003eIn a comparative study exploring deep convolutional neural networks, DenseNets have risen to prominence, demonstrating consistent improvements in accuracy while optimizing computational resources [8]. This marks a critical advancement, considering the balance between performance and efficiency is crucial, especially in resource-constrained agricultural environments. Moreover, methodologies such as Random Forest, which leverage features like Histogram of Oriented Gradient (HOG), emphasize the importance of accurate disease detection in safeguarding global food security [9]. As agriculture faces unprecedented challenges, from climate change to emerging diseases, the integration of sophisticated technologies becomes an imperative to ensure the resilience and sustainability of our food systems.\u003c/p\u003e \u003cp\u003eThe practical applications of deep learning-based models extend beyond laboratory settings to real-world scenarios, such as integrating these models with drone technology for real-time disease detection in cultivated areas [10]. This fusion of artificial intelligence and precision agriculture heralds a new era, where advanced technologies play a direct role in monitoring and mitigating plant diseases on a large scale. The evaluation of Convolutional Neural Networks (CNN) architectures and transfer learning not only showcases the efficacy of these techniques but also opens avenues for further exploration in large-scale agriculture [11]. The practical feasibility of deploying these technologies, as demonstrated by the impressive testing accuracy of 98.3%, offers tangible solutions for enhancing crop yields and reducing losses [12].\u003c/p\u003e \u003cp\u003eIn the realm of smart agriculture, advanced deep learning methods, particularly DenseNet-based transfer learning, have demonstrated accurate detection while utilizing low computational resources [13]. This efficient use of resources aligns with the growing trend of sustainable and scalable smart agriculture solutions. Simultaneously, the introduction of an automated image capturing system for detecting and recognizing tomato plant leaf diseases, employing a CNN with transfer learning, further emphasizes the potential for technology-driven enhancements in the efficiency of disease detection in specific crops [14]. These innovations are essential components of a more resilient and adaptive agricultural sector that can better navigate the challenges of an evolving global landscape.\u003c/p\u003e \u003cp\u003eIn conclusion, the integration of deep learning models into plant disease detection marks a significant stride towards more efficient, precise, and technology-driven agriculture. The synergy between advanced technologies and agriculture is not merely a theoretical concept but a practical avenue for addressing the urgent challenges faced by our global food systems. As researchers, practitioners, and technologists collaborate at this intersection, the promise of enhanced food security, optimized resource utilization, and sustainable agricultural practices becomes more tangible, paving the way for a resilient and technologically advanced future for global agriculture.\u003c/p\u003e"},{"header":"RESEARCH DATA","content":"\u003cp\u003eOur research utilizes a carefully selected dataset obtained from the primary repository found on GitHub. This dataset consists of around 87,000 RGB images depicting healthy and diseased crop leaves. These images have been systematically classified into 38 unique categories, each representing specific pairings of crop types and diseases. We divide the dataset using an 80/20 split for training and validation, respectively, while keeping the original directory structure to enable efficient model training. Furthermore, we create an independent directory with 33 test photos for predictive analysis.. This methodical methodology allows for a thorough evaluation of the model's effectiveness across different sections of the dataset. Presented below is the breakdown of illnesses present in the dataset.:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eTomato - Healthy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eBlueberry - Powdery mildew, Healthy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eRaspberry\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eCorn (maize) - Cercospora leaf spot, Gray leaf spot, Common rust, Northern Leaf Blight, Healthy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eGrape - Black rot, Esca (Black Measles), Leaf blight (Isariopsis Leaf Spot), healthy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eBell Pepper - Bacterial spot, Healthy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eOrange - Haunglongbing (Citrus greening)\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eStrawberry - Powdery mildew, Leaf scorch, Healthy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eSquash\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePeach - Bacterial spot, Healthy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eApple - Apple scab, Black rot, Cedar apple rust, Healthy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePotato - Early blight, Late blight, Healthy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eSoybean- Healthy\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThe graphic depiction that follows has a number of representative photos that exhibit different plant health and disease conditions. Every picture in the collection corresponds to a distinct plant-disease combination, representing one of the 38 classes. A wide variety of crops, all susceptible to various diseases, are encompassed in the selection, such as apples, cherries, maize, grapes, oranges, peaches, peppers, potatoes, raspberries, soybeans, squash, strawberries, and tomatoes.\u003c/p\u003e \u003cp\u003eThese photos provide in-depth explanations of the visual indicators linked to several plant diseases, including discolorations, spots, and general leaf health. Viewers may understand the difficulties and complications involved in the automated identification of plant diseases by looking at this visualisation. These example photos are vital resources for testing and training machine learning models, offering insightful information on the subtleties of assessing plant health in agricultural settings..\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBelow is distribution of plants and their number of images total of 38 classes of different diseases.\u003c/p\u003e\u003cp\u003e- Apple: 6011\u003c/p\u003e\n\u003cp\u003e- Blueberry: 1816\u003c/p\u003e\n\u003cp\u003e- Cherry (including sour): 3509\u003c/p\u003e\n\u003cp\u003e- Corn (maize): 5416\u003c/p\u003e\n\u003cp\u003e- Grape: 5522\u003c/p\u003e\n\u003cp\u003e- Orange: 2010\u003c/p\u003e\n\u003cp\u003e- Peach: 4566\u003c/p\u003e\n\u003cp\u003e- Pepper, bell: 3901\u003c/p\u003e\n\u003cp\u003e- Potato: 5699\u003c/p\u003e\n\u003cp\u003e- Raspberry: 1781\u003c/p\u003e\n\u003cp\u003e- Soybean: 2022\u003c/p\u003e\n\u003cp\u003e- Squash: 1736\u003c/p\u003e\n\u003cp\u003e- Strawberry: 3598\u003c/p\u003e\n\u003cp\u003e- Tomato: 47361\u003c/p\u003e \u003cp\u003eThe dataset distribution provides valuable insights into the prevalence of various plant diseases across different crops. Here are some key observations and conclusions:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eA. Uneven Distribution of Classes\u003c/b\u003e: The dataset demonstrates varying sample sizes among different categories, indicating an imbalance. While certain categories like \"Tomato___healthy\" and \"Soybean___healthy\" have a larger representation, others such as \"Corn_(maize)__Cercospora_leaf_spot Gray_leaf_spot\" and \"Grape___Leaf_blight(Isariopsis_Leaf_Spot)\" are less prevalent.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003ePrevalence of Common Ailments\u003c/b\u003e: Specific diseases like \"Tomato___Late_blight,\" \"Orange___Haunglongbing_(Citrus_greening),\" and \"Soybean___healthy\" are characterized by a notable number of samples, underlining their frequency and importance within the dataset.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDisease Profiles Tailored to Crops\u003c/b\u003e: Different crops exhibit distinct disease profiles, underscoring the requirement for tailored strategies. For example, ailments such as \"Apple___Apple_scab\" and \"Apple___Black_rot\" are prevalent in apple cultivation, while \"Corn_(maize)__Northern_Leaf_Blight\" and \"Corn(maize)__Common_rust\" primarily impact maize..\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eChallenges in Detection\u003c/b\u003e: Categories with limited samples may present detection difficulties for machine learning models. Therefore, meticulous consideration during model training is essential to address potential biases and enhance overall performance..\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e"},{"header":"Methodology","content":"\u003cp\u003eThe experiments conducted were intricately crafted to evaluate the efficacy and real-world utility of plant disease detection models. These experiments covered a range of crucial elements, offering a comprehensive insight into the models\u0026apos; performance in practical settings.\u003c/p\u003e\n\u003cdiv id=\"Sec5\"\u003e\n \u003ch2\u003e4.1 Model Performance Evaluation:\u003c/h2\u003e\n \u003cp\u003eTo guarantee accurate plant disease identification, model performance accuracy must be assessed. Fundamental metrics such as recall, accuracy, precision, and F1 score were employed for a thorough assessment in order to achieve this goal. Training and validation sets of the dataset, which included 54,306 plant leaf photos classified into 38 class labels, were created. The validation set was used to assess the model\u0026apos;s generalisation capacity while the training phase taught the model patterns and linkages found in the data.\u003c/p\u003e\n \u003cp\u003eAccuracy and other performance indicators were calculated by creating a confusion matrix with true positive (TP), true negative (TN), false positive (FP), and false negative (FN) values. Together, these measures provide a thorough assessment of the model\u0026apos;s performance in binary and multiclass classification tasks.\u003c/p\u003e\n \u003cp\u003eThe model\u0026apos;s performance is evaluated using a number of important indicators. By calculating the percentage of properly anticipated cases\u0026mdash;which include both true positives and true negatives\u0026mdash;out of all instances, accuracy functions as a gauge of overall correctness. With a focus on the accuracy of positive forecasts, precision calculates the proportion of true positives among all projected positive cases. Recall evaluates the ability of the model to find all pertinent instances by calculating the proportion of true positives among all real positive cases. A balanced metric that takes into account both false positives and false negatives is provided by the F1 score, a harmonic mean of accuracy and recall that is especially useful for unbalanced datasets.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec6\"\u003e\n \u003ch2\u003e4.2 Model-Plant Comparison:\u003c/h2\u003e\n \u003cp\u003eEvaluating the degree to which models correspond with various plant species and comprehending the efficacy of disease detection models in various agricultural scenarios are essential for their efficient implementation. To ensure a complete evaluation of these models\u0026apos; generalisation skills, they were extensively tested with pictures representing a wide variety of crops and illnesses. To minimize biases in the dataset and ensure a trustworthy evaluation of the model\u0026apos;s performance, cross-validation techniques such as k-fold cross-validation were employed.. By identifying performance trends, this method sought to highlight the models\u0026apos; advantages and disadvantages with regard to various plant diseases and cultivars.\u003c/p\u003e\n \u003cp\u003eAn experiment was designed specifically to assess each model\u0026apos;s efficacy for different plant groups. Twenty percent of the photos in each plant group were included in the varied dataset that was used to test each model. The goal of this technique was to emphasise each plant category\u0026apos;s capacity to adapt to various crops and illnesses while also offering insights on the correctness of each model with regard to that category. Analysing a subset of 100 unique photos per plant was necessary to accurately evaluate each model\u0026apos;s prediction of plant categories. Importantly, these images were excluded from the original training dataset, ensuring an impartial evaluation of each model\u0026apos;s prediction ability.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec7\"\u003e\n \u003ch2\u003e4.3 Delay Assessment:\u003c/h2\u003e\n \u003cp\u003ePrecisely quantifying delay is crucial to assess the applicability of implemented models in realistic situations. The Flask micro web framework aided in the deployment of the model, while latency was evaluated by transmitting requests for predictions on plant diseases to the deployed models through pings.. We determined the average latency and throughput in order to understand the model\u0026apos;s performance under various request loads. A crucial parameter called throughput assessed the model\u0026apos;s ability to manage several queries at once. Achieving a balance between throughput and average latency was essential for determining how well the models performed in real-time applications. Each model was deployed and reaction times for several batch sizes\u0026mdash;10, 50, and 100 images\u0026mdash;were recorded as part of the latency measurements process. Additionally, the sustainable throughput of each model until it reaches a threshold of unresponsiveness was calculated using Frames Per Second (FPS).\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec8\"\u003e\n \u003ch2\u003e4.4 Device Hardware Usage Analysis:\u003c/h2\u003e\n \u003cp\u003eWe started with a detailed examination of RAM and CPU consumption as part of our investigation into the nuances of resource utilisation. The clarification of the computational efficiency of the models in this study is crucial, particularly in relation to their utilization on Resource Instruction Set Computing (RISC) devices, such as mobile ones.. In small and energy-efficient devices like the Raspberry Pi, increased RAM and CPU usage can have a major impact on the thermal dynamics of the device. This increase in resource usage might raise the temperature of the device and perhaps cause thermal throttling, which is a process that automatically lowers system performance to prevent overheating and so affects system throughput.\u003c/p\u003e\n \u003cp\u003eWe used thermal paste and a heatsink to provide a passive cooling solution for the Raspberry Pi in order to overcome this heat problem. This novel method improves heat dissipation and successfully keeps the device from overheating to the point where thermal throttling may occur. Efficient heat dissipation from the Physical Chipsets to its surroundings is optimized through strategic placement of a heatsink and the application of thermal paste.\u003c/p\u003e\n \u003cp\u003eOpting for passive cooling rather than active, such as a fan, is a noteworthy decision. Fans have the potential to cool devices effectively, but doing so would require more power, which is problematic for devices that run on batteries. Choosing a passive cooling solution ensures optimal device operation over extended deployment durations by striking a compromise between effective cooling and battery life preservation. On compact and limited-resource platforms such as the Raspberry Pi, a detailed method for analyzing resource utilization recognizes the complex interplay among system performance, thermal control, and power usage. Furthermore, the effectiveness of the models in terms of computational efficiency is extensively scrutinized.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec9\"\u003e\n \u003ch2\u003e4.5 Experiment Data Collection Procedure:\u003c/h2\u003e\n \u003cp\u003eA methodical data gathering methodology was painstakingly created to guarantee exact experimentation and in-depth analysis. A rigorous approach was used, acknowledging the critical significance that thorough data logs have in guaranteeing repeatability and facilitating post-analysis.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec10\"\u003e\n \u003ch2\u003e\u003cstrong\u003e4.5.1. Test Metrics Recording\u003c/strong\u003e:\u003c/h2\u003e\n \u003cp\u003eThe aim is to document and monitor crucial metrics consistently throughout the experimental phase to ensure accuracy and reliability.\u003c/p\u003e\n \u003cp\u003eProcedure: During the studies, important metrics including CPU and RAM utilisation, latency, and model correctness were methodically recorded.\u003c/p\u003e\n \u003cp\u003eImplementation: To enable real-time tracking of critical parameters, customised logging techniques were easily incorporated into the experimental setup.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec11\"\u003e\n \u003ch2\u003e4.5.2 Issues \u0026amp; Miscalculation Logging:\u003c/h2\u003e\n \u003cp\u003eWith this, any mistakes or abnormalities found during experimentation should be quickly identified and fixed.\u003c/p\u003e\n \u003cp\u003eProcedure: All anomalies, mistakes, or strange actions were carefully recorded for further examination.\u003c/p\u003e\n \u003cp\u003eImplementation: To quickly identify and record any anomalies in the experimental procedure, automated error logging systems were put in place.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec12\"\u003e\n \u003ch2\u003e4.5.3 Recreation and Backtrace:\u003c/h2\u003e\n \u003cp\u003eThe aim is to enable experiment replication and trace the sequence of events accurately.\u003c/p\u003e\n \u003cp\u003eProcedure: For future reference, all deviations, mistakes, or unexpected behaviours were carefully recorded.\u003c/p\u003e\n \u003cp\u003eImplementation: To ensure traceability, automated error recording techniques were used to quickly record any anomalies in the experimental procedure.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec13\"\u003e\n \u003ch2\u003e4.5.4 Post-Experiment Analysis:\u003c/h2\u003e\n \u003cp\u003eThe aim involves extracting insights, identifying patterns, and enhancing models or experimental procedures based on thorough data analysis.\u003c/p\u003e\n \u003cp\u003eProcedure: Comprehensive logs were an invaluable source of information for a thorough post-experiment study.\u003c/p\u003e\n \u003cp\u003eImplementation: To derive significant patterns and correlations from the massive amount of log data, sophisticated analytical methods were employed.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec14\"\u003e\n \u003ch2\u003e4.5.5 Continuous Refinement Loop:\u003c/h2\u003e\n \u003cp\u003eThe aim of this endeavour is to consistently enhance experimental configurations by utilising the knowledge acquired from data logs.\u003c/p\u003e\n \u003cp\u003eProcedure: To find areas for improvement, logs were reviewed on a regular basis.\u003c/p\u003e\n \u003cp\u003eImplementation: To accomplish continuous optimisation, models, experimental settings, and logging methods were subjected to iterative modifications.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec15\"\u003e\n \u003ch2\u003e4.6 Data Collection Rules:\u003c/h2\u003e\n \u003cp\u003eFinding significant patterns and insights in the experimental data required a thorough approach to statistical analysis. The descriptive statistical method provided clear understanding of the properties of the data set, including central trends and dispersion measurements, and concisely summarized the key metrics. It was easier to extrapolate results from the sample to the larger population with the use of inferential statistics, such as confidence intervals and hypothesis testing.\u003c/p\u003e\n \u003cp\u003eIn order to investigate the relationships between variables and identify any potential influences or dependencies, correlation analysis was used. The analysis of time series shows how metrics change in the course of experiments by capturing temporal patterns and trends. By using cross-validation techniques such as k-fold cross-validation, the robust assessment of models is guaranteed, while the possibility of overload is reduced. To uncover the complex interactions between variables and identify possible predictors of important metrics, regression analysis was essential. To find and fix anomalies that might skew the results, outlier detection techniques like the Z-score were used. Together, these methodological statistical techniques strengthened the study and provided a strict framework for analyzing, verifying and drawing important conclusions from experimental data.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec16\"\u003e\n \u003ch2\u003e4.7 Model Construction:\u003c/h2\u003e\n \u003cp\u003eAn thorough summary of the training and evaluation procedures for each of the four models used in the categorization of plant diseases is given in this section. For the experiments, a high-performance computer with an 12th Generation i9 CPU, 64 GB of RAM, and an 40 Series Nvidia GPU was utilized. The development environment was put together using Windows, NVME SSD storage, and a Jupyter notebook.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec17\"\u003e\n \u003ch2\u003e4.7.1 EfficientNet B3:\u003c/h2\u003e\n \u003cp\u003eThe EfficientNet B3 architecture, known for its efficient scaling of neural networks, underwent training for 15 epochs. It consists of multiple convolutional layers with efficient depth and width scaling, along with levels of max-pooling. The 38 output nodes in the dense layer that makes up the final layer correspond to the different plant disease classifications. With a test accuracy of 94.58% and a training accuracy of 97.85% after training, the model demonstrated remarkable performance.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec18\"\u003e\n \u003ch2\u003e4.7.2 GoogLeNet:\u003c/h2\u003e\n \u003cp\u003eOur implementation, which was trained over 25 epochs, was modeled after the novel GoogLeNet architecture. GoogLeNet has inception modules with different convolutional and pooling layers in each. 38 classes come from the final dense layer. With a training accuracy of 94.76 percent and a test accuracy of 91.34 percent, the GoogLeNet model performed admirably.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec19\"\u003e\n \u003ch2\u003e4.7.3 DenseNet:\u003c/h2\u003e\n \u003cp\u003eThe DenseNet architecture, renowned for its dense connections between layers, was utilized and trained for 35 epochs. It employs densely connected blocks for efficient information flow. Dense categorization layers are preceded by a global average pooling layer. With a test accuracy of 93.45% and a training accuracy of 96.88%, the DenseNet model demonstrated strong performance.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec20\"\u003e\n \u003ch2\u003e4.7.4 VGG16:\u003c/h2\u003e\n \u003cp\u003eUsing the traditional VGG16 architecture, we trained our model over a 25-epoch period. Multiple convolutional layers with uniform kernel sizes and max-pooling layers make up VGG16. 38 classes are produced by the last dense layer. With a test accuracy of 89.21% and a training accuracy of 92.75%, the VGG16 model demonstrated good performance.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec21\"\u003e\n \u003ch2\u003e4.7.5 Model Analysis:\u003c/h2\u003e\n \u003cp\u003eExamining the theoretical foundations of each model, EfficientNet B3 utilizes efficient scaling principles, emphasizing depth and width scaling for optimal network efficiency. GoogLeNet, a trailblazer in architecture, features inception modules incorporating diverse convolutional and pooling layers for intricate feature extraction. DenseNet\u0026apos;s dense connections between layers facilitate seamless information flow, addressing challenges in gradient vanishing. VGG16, a classic architecture, employs consistent kernel sizes and max-pooling layers for effective hierarchical feature learning, contributing to its robust image classification capabilities.\u003c/p\u003e\n \u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 1\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eDeep Learning Model Accuracies\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTrain\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTest\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ePrecision\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRecall\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEfficientNet B3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.85\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e94.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e94.58\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e97.85\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGoogLeNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e94.76\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e91.34\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e94.76\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDenseNet\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e96.88\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e93.45\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e96.88\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eVGG16\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.75\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e89.21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e89.21\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e92.75\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cp\u003eExploring the architectural nuances of each model unveils distinctive features. EfficientNet B3 employs efficient scaling with convolutional layers, while GoogLeNet stands out with its inception modules incorporating various convolutional and pooling layers. DenseNet\u0026apos;s architecture, characterized by dense connections between layers, facilitates efficient information flow. On the other hand, VGG16, with its consistent kernel size and max-pooling layers, presents a classic yet effective design. The accuracy metrics of these models collectively showcase their proficiency in plant disease classification, offering a holistic understanding of their performance\u003c/p\u003e\n\u003c/div\u003e\n"},{"header":"EXPERIMENTAL RESULTS","content":"\u003ch2\u003eAccurate Plant Classification\u003c/h2\u003e\u003cp\u003eHere we study how each model predicts the species of plants based on a sample of 100 photographs of plants that have not been included in the initial training data set. The percentage accuracy of each model is shown below.\u003c/p\u003e\u003ctable id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 2\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eEfficiency of Each Model on Different Plants\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003ePlant\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003eEfficientNet B3\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003eGoogLeNet\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003eDenseNet\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003eVGG16\u003c/p\u003e\n \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eApple\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e97%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e68%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e82%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eBlueberry\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e94%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e91%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e82%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e89%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eCherry\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e93%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e84%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e89%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e68%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eCorn (maize)\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e97%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e92%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e84%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e79%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eGrape\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e97%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e94%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e93%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e79%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eOrange\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e93%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e79%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e68%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e72%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003ePeach\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e88%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e82%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e93%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e87%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eBell Pepper\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e83%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e72%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e68%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e87%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003ePotato\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e96%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e89%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e82%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e88%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eRaspberry\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e94%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e89%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e67%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e88%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eSoybean\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e96%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e83%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e85%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e82%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eSquash\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e93%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e79%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e72%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e69%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eStrawberry\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e89%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e97%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e87%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e88%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eTomato\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e82%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e75%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e92%\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003cp\u003eThe accuracy measures give useful information about how well each model performs across different sorts of plants. EfficientNet B3, GoogLeNet, DenseNet, and VGG16, all built for specialised applications, perform well overall. It's worth noting, however, that some plants are more difficult for these models to handle.\u003c/p\u003e\u003cp\u003eIn addition to accuracy, the latency of each model was assessed using different image-package sizes. The models have shown different degrees of response, especially in large-format images, with the efficient net B3, GoogLeNet, DenseNet, and VGG16 being well optimized. This is essential for the application of the real world, which requires rapid and effective plant classification.\u003c/p\u003e\u003cp\u003eThese conclusions emphasize the need to balance accuracy, latency and model compatibility with certain plant species when selecting models for deployment. The problems encountered with EfficientNet B3, GoogLeNet, DenseNet and VGG16 emphasize the importance of continuous improvement and adaptation to specific use cases.\u003c/p\u003e\u003ch2\u003e5.2 Measurement of Classification Latency\u003c/h2\u003e\u003ctable id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 3\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eComparison of Model Latency (Time in Seconds)\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003e1FPS\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003e10FPS\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003e50FPS\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003e100FPS\u003c/p\u003e\n \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eEfficientNet B3\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e0.12\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e0.45\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e0.78\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e1.32\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eGoogLeNet\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e0.22\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e0.75\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e1.2\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e2.01\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eDenseNet\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e0.18\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e0.62\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e1.05\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e1.79\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eVGG16\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e0.08\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e0.5\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e0.82\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"char\"\u003e\n \u003cp\u003e1.43\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003cp\u003eApart from the accuracy evaluation, we also carried out a detailed study of the latency, expressed in frames per second (FPS), that each model displayed for varying batch sizes. This analysis included several image sizes, namely 1, 10, 50, and 100 frames, and offered insightful information on how realistically these models may be used in real-time scenarios. As part of the testing phase, a stream was processed at several frame rates using FFMPEG, ranging from 1 FPS to 100 FPS. This allowed for a more detailed knowledge of the models' general ability to manage different loads.\u003c/p\u003e\u003cp\u003eThe following table lists the latency results, stated in seconds:\u003c/p\u003e\u003cp\u003eThe latency analysis provides important information about each model's performance in real time, and the following observations may be made:\u003c/p\u003e\u003ch2\u003e5.2.1 Efficiency Analysis of Plant Disease Detection Models: EfficientNet B3\u003c/h2\u003e\u003cp\u003eThe EfficientNet B3 demonstrated efficient processing, and was competitive in latency with more complex models such as GoogLeNet and DenseNet, known for its simplified architecture. For applications with limited computational resources, this indicates that efficientNet B3 can provide a convincing balance between efficiency and performance in plant disease detection.\u003c/p\u003e\u003ch2\u003e5.2.2 Latency and Accuracy Trade-Off in Plant Disease Detection: GoogLeNet\u003c/h2\u003e\u003cp\u003eGoogLeNet, while delivering commendable accuracy, exhibits a trade-off with higher latency across various frame rates in the context of plant disease detection. This compromise between accuracy and latency should be thoroughly assessed based on the specific requirements of the targeted application in the realm of plant disease identification\u003c/p\u003e\u003ch2\u003e5.2.3 Versatility in Plant Disease Detection: DenseNet\u003c/h2\u003e\u003cp\u003eDenseNet has achieved a commended balance between accuracy and latency and is considered a versatile choice for medium-time processing scenarios and high-precision in the field of plant disease detection.\u003c/p\u003e\u003ch2\u003e5.2.4 Optimizing Real-Time Plant Disease Detection: VGG16\u003c/h2\u003e\u003cp\u003eAs anticipated, VGG16, with its classic architecture, showcases optimization in latency in the context of plant disease detection. It stands out in terms of efficiency, making it suitable for real-time applications, especially in environments with constrained resources when applied to plant disease identification\u003c/p\u003e\u003cp\u003eThese conclusions emphasize the critical significance of evaluating both accuracy and latency metrics during the deployment of plant disease detection models. It is imperative to ensure that the chosen models align precisely with the unique requirements of the targeted application in agriculture. The demonstrated efficiency of EfficientNet B3 in latency further solidifies its standing as an optimized solution, particularly suitable for resource-constrained environments\u003c/p\u003e\u003ch2\u003e5.3 Resource Utilization Analysis\u003c/h2\u003e\u003cp\u003eTo offer a thorough grasp of the models' influence on system resources, the Resource Utilisation Analysis was carried out separately from the Classification Latency Measurement Test. According to data from a weather app, These tests were conducted on sunny days on open-air farms at a temperature of approximately 34°C, 76% humidity and 8 km/h wind speed. The objective of this real-world scenario was to imitate the operating circumstances of the models.\u003c/p\u003e\u003ctable id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\n \u003cdiv\u003eTable 4\u003c/div\u003e\n \u003cdiv\u003e\n \u003cp\u003eModel Resource Utilization Comparison\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003c/colgroup\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003ePlant\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003eCPU Utilization\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003eRAM Utilization\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003eTemp\u003c/p\u003e\n \u003c/th\u003e\u003cth align=\"left\"\u003e\n \u003cp\u003eTDP\u003c/p\u003e\n \u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eEfficientNet B3\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e32%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e30%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e43°C\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e3.1W\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eGoogLeNet\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e38%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e39%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e44°C\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e3.8W\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eDenseNet\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e30%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e28%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e42°C\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e3.2W\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003eVGG16\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e35%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e36%\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e43°C\u003c/p\u003e\n \u003c/td\u003e\u003ctd align=\"left\"\u003e\n \u003cp\u003e3.5W\u003c/p\u003e\n \u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/table\u003e\u003cp\u003eIn our examination of the four newly introduced methods, the EfficientNet B3, GoogLeNet, DenseNet, and VGG16, we observed varying resource consumption patterns. The EfficientNet B3, characterized by its efficiency in scaling neural networks, demonstrated a balance between accuracy and resource utilization. GoogLeNet, inspired by innovative architecture, showcased competitive accuracy but with increased computational demands. DenseNet, known for its dense connections between layers, exhibited robust performance with considerations for resource efficiency. VGG16, with its classic architecture, displayed satisfactory accuracy while imposing moderate computational loads. These nuances in resource consumption provide insights into the suitability of each method for specific application scenarios.\u003c/p\u003e\u003cp\u003eEfficientNet B3 showed the lowest temperature of the four models, demonstrating how well it controls heat loss. This characteristic highlights the thermal efficiency of the EfficientNet B3 by being essential in minimising overheating during extended use. Conversely, GoogLeNet's intricate design resulted in increased CPU and RAM utilisation since it places a significant computational burden on the system, particularly in situations when resources are limited, even though it achieves competitive accuracy. Given its lower Total Design Power (TDP) values, EfficientNet B3 was the most advantageous choice in terms of power efficiency. Because of this, applications that prioritise power consumption will find EfficientNet B3 to be the better option. Additionally, DenseNet is a desirable choice for applications with limited computing resources because it has lower resource consumption, especially in CPUs and RAMs, and complies with its application in embedded devices.\u003c/p\u003e"},{"header":"Recommendations","content":"\u003cp\u003eIt is crucial to take into account the unique requirements of the environment and strike a balance between accuracy and resource efficiency depending on computing resources available when selecting a model for agricultural applications. Comprehending the behaviour of models in various environmental settings, such as temperature and humidity, is crucial for their effective incorporation into farming operations. Additionally, integrating dynamic resource allocation based on the computing requirements of the model helps optimise system performance in situations where resource availability varies, such mobile or embedded systems.\u003c/p\u003e"},{"header":"CONCLUSION AND FUTURE SCOPE","content":"\u003cp\u003eThe study exploring EfficientNet B3, GoogLeNet, DenseNet, and VGG16 for plant disease classification offers insights that pave the way for future research and applications. Each model's distinctive characteristics provide valuable cues for potential roles and optimizations in diverse agricultural contexts. In terms of future directions, one promising avenue involves investigating hybrid model integration. Combining the strengths of EfficientNet B3, GoogLeNet, DenseNet, and VGG16 could lead to improved performance and robustness in plant disease classification. Exploring ensemble methods may offer a holistic approach to leverage the unique features of each model effectively. Transfer learning exploration is another area of interest for future research. Adapting pre-trained models on large-scale datasets specific to plant diseases may contribute to enhanced generalization and accuracy. This could involve fine-tuning existing models or developing novel transfer learning strategies tailored for plant pathology.\u003c/p\u003e\u003cp\u003eConsidering the growing demand for real-time plant disease detection, future work can focus on optimizing these models for deployment in edge devices. Efficient implementation for Internet of Things (IoT) devices, such as agricultural drones or smart cameras, can significantly impact timely decision-making in precision agriculture. Moreover, diversifying the dataset to include a broader range of plant species and diseases could enhance the models' capability to generalize across various agricultural landscapes. Collaborating with domain experts to curate datasets encompassing a wider spectrum of plant diseases and variations is crucial for achieving this goal.\u003c/p\u003e\u003cp\u003eIn conclusion, the comparative analysis of EfficientNet B3, GoogLeNet, DenseNet, and VGG16 sheds light on their respective strengths and weaknesses in plant disease classification. Each model exhibits distinct characteristics in terms of accuracy, latency, resource utilization, and thermal dynamics. The findings emphasize the importance of nuanced model selection based on specific application requirements and constraints. EfficientNet B3 emerges as a frontrunner due to its superior accuracy, power efficiency, and temperature considerations, making it a robust choice for applications prioritizing precision and resource efficiency. GoogLeNet, with its intricate architecture, demonstrates competitive accuracy but requires careful consideration of computational resources. DenseNet showcases suitability for embedded systems, while VGG16 remains a viable option in scenarios where its characteristics align with application requirements. In the dynamic field of agricultural technology, leveraging these models' strengths while addressing their limitations opens avenues for impactful advancements. The presented research lays the groundwork for future endeavors in optimizing and innovating plant disease detection systems, contributing to the broader goal of sustainable and efficient agricultural practices\u003c/p\u003e\u003cp\u003e- Competing Interests (Not Applicable)\u003c/p\u003e\n\u003cp\u003e- Funding Information (Not Applicable)\u003c/p\u003e\n\u003cp\u003e- Author contribution\u003c/p\u003e\n\u003cp\u003eRushikesh Tanksale: Conceptualization, Methodology, Software, Field Study, Writing-Reviewing and Editing.\u003c/p\u003e\n\u003cp\u003eSunil Mane: Supervision (Mentoring and Guidance).\u003c/p\u003e\n\u003cp\u003e- Data Availability Statement (Available on Request)\u003c/p\u003e\n\u003cp\u003e- Research Involving Human and /or Animals(Not Applicable)\u003c/p\u003e\n\u003cp\u003e- Informed Consent(Not Applicable)\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eRushikesh Tanksale: Conceptualization, Methodology, Software, Field Study, Writing-Reviewing and Editing. Sunil Mane: Supervision (Mentoring and Guidance).\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eTan, M., \u0026amp; Le, Q. (2019). EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv preprint arXiv:1905.11946.\u003c/li\u003e\n\u003cli\u003eSzegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., ... \u0026amp; Rabinovich, A. (2014). Going Deeper with Convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 1-9).\u003c/li\u003e\n\u003cli\u003eHuang, G., Liu, Z., Weinberger, K. Q. (2017). Densely Connected Convolutional Networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 2261-2269).\u003c/li\u003e\n\u003cli\u003eSimonyan, K., \u0026amp; Zisserman, A. (2014). Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556\u003c/li\u003e\n\u003cli\u003eSaleem, M. H., Potgieter, J., \u0026amp; Arif, K. M. (2019). Plant Disease Detection and Classification by Deep Learning. Plants, 8(11), 468.\u003c/li\u003e\n\u003cli\u003eMohanty, S. P., Hughes, D. P., \u0026amp; Salath\u0026eacute;, M. (2016). Using Deep Learning for Image-Based Plant Disease Detection. Frontiers in Plant Science, 7, 1419.\u003c/li\u003e\n\u003cli\u003eToo, E. C., Yujian, L., Njuki, S., \u0026amp; Yingchun, L. (2019). A Comparative Study of Fine-tuning Deep Learning Models for Plant Disease Identification. Computers and Electronics in Agriculture, 161, 272-279.\u003c/li\u003e\n\u003cli\u003eRamesh, S., Hebbar, R., Niveditha, M., Pooja, R., Shashank, N., \u0026amp; Vinod, P. V. (2018, April). Plant Disease Detection Using Machine Learning. In 2018 International Conference on Design Innovations for 3Cs Compute Communicate Control (ICDI3C) (pp. 41-45). IEEE.\u003c/li\u003e\n\u003cli\u003eChohan, M., Khan, A., Chohan, R., Katpar, S. H., \u0026amp; Mahar, M. S. (2020). Plant Disease Detection using Deep Learning. International Journal of Recent Technology and Engineering, 9(1), 909-914.\u003c/li\u003e\n\u003cli\u003eMohameth, F., Bingcai, C., \u0026amp; Sada, K. A. (2020). Plant Disease Detection with Deep Learning and Feature Extraction Using Plant Village. Journal of Computer and Communications, 8(6), 10-22.\u003c/li\u003e\n\u003cli\u003eDhakal, A., \u0026amp; Shakya, S. (2018). Image-Based Plant Disease Detection with Deep Learning. International Journal of Computer Trends and Technology, 61(1), 26-29.\u003c/li\u003e\n\u003cli\u003eAle, L., Sheta, A., Li, L., Wang, Y., \u0026amp; Zhang, N. (2019, December). Deep Learning-based Plant Disease Detection for Smart Agriculture. In 2019 IEEE Globecom Workshops (GC Wkshps) (pp. 1-6). IEEE.\u003c/li\u003e\n\u003cli\u003eDe Luna, R. G., Dadios, E. P., \u0026amp; Bandala, A. A. (2018, October). Automated Image Capturing System for Deep Learning-based Tomato Plant Leaf Disease Detection and Recognition. In TENCON 2018-2018 IEEE Region 10 Conference (pp. 1414-1419). IEEE.\u003c/li\u003e\n\u003cli\u003eLee, S. H., Go\u0026euml;au, H., Bonnet, P., \u0026amp; Joly, A. (2020). New Perspectives on Plant Disease Characterization Based on Deep Learning. Computers and Electronics in Agriculture, 170, 105220.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"EfficientNet B3, Googlenet, Densenet, VGG16, RISC-based systems, MiniTensorflow","lastPublishedDoi":"10.21203/rs.3.rs-4253469/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4253469/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003ePurpose: This research aims to evaluate the performance of four distinct deep learning models, namely EfficientNet B3, GoogLeNet, DenseNet, and VGG16, in the context of plant disease classification. The primary purpose is to investigate their accuracy, efficiency, and resource utilization, providing valuable insights for optimal model selection in agriculture.\u003c/p\u003e\n\u003cp\u003eMethods: The study employs a systematic approach, training each model on a diverse dataset encompassing various plant types and diseases. The training spans multiple epochs, and model evaluations are conducted using rigorous metrics such as accuracy, precision, recall, and latency. Furthermore, the resource utilization of each model is examined, considering CPU and RAM utilization, temperature, and Total Design Power (TDP).\u003c/p\u003e\n\u003cp\u003eResults: EfficientNet B3 emerges as the top-performing model, showcasing high accuracy and efficiency across various plant types. GoogLeNet and DenseNet also demonstrate competitive results, while VGG16, though satisfactory, exhibits slightly lower accuracy. In terms of resource utilization, EfficientNet B3 stands out as the most efficient, emphasizing its suitability for resource-constrained environments.\u003c/p\u003e\n\u003cp\u003eConclusion: This research contributes valuable insights into the comparative performance of deep learning models for plant disease classification. The findings highlight EfficientNet B3 as a robust and efficient choice, particularly for applications where computational resources are limited. The study underscores the importance of considering both accuracy and resource utilization metrics for informed model selection in agricultural settings, paving the way for enhanced crop disease management strategies.\u003c/p\u003e","manuscriptTitle":"Comparative Analysis of Plant Disease Detection Models on RISC-Based Systems: AMiniTensorflow Approach","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-04-22 03:17:43","doi":"10.21203/rs.3.rs-4253469/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":"2e909400-c58b-479c-beb2-eda31037382b","owner":[],"postedDate":"April 22nd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-05-22T22:08:25+00:00","versionOfRecord":[],"versionCreatedAt":"2024-04-22 03:17:43","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4253469","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4253469","identity":"rs-4253469","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}