Agriculture plays a vital role in global food security, necessitating early and accurate detection of plant diseases to mitigate yield losses. This research leverages deep learning techniques to develop a robust plant disease detection system, utilizing the PlantVillage dataset focusing on potato and bell pepper leaves. Multiple approaches were implemented, including Support Vector Machines (SVM), fully connected neural networks, convolutional neural networks (CNNs), and transfer learning with pre-trained models like ResNet50, EfficientNetB0, and MobileNetV2. The models were trained on preprocessed and augmented image datasets to enhance generalization and mitigate overfitting. The CNN achieved a test accuracy of 99.61% for potato diseases, while transfer learning models achieved up to 100% accuracy for bell pepper disease classification. This study demonstrates the potential of integrating machine learning into agriculture, enabling proactive disease management and paving the way for sustainable farming practices. Future work involves refining these techniques to accommodate diverse crop types and deploying models in real-world scenarios.
Detecting plant diseases is a critical aspect of agriculture, enabling timely interventions to prevent disease spread and protect crop yield and quality. This project aims to address this challenge by developing an effective plant disease detection system using deep learning models that leverage neural networks to identify subtle patterns indicative of diseases. Utilizing a subset of the PlantVillage dataset, which includes standardized 256x256-pixel images of potato and bell pepper leaves under varying conditions, the study focuses on key classes such as healthy and diseased leaves. This dataset, comprising categories like Potato Healthy (152 images), Potato Early Blight (1000), Potato Late Blight (1000), Pepper Bell Healthy (1478), and Pepper Bell Bacterial Spot (997), provides a robust foundation for training and evaluating the proposed models.
To address the critical challenge of plant disease detection, our research explored a variety of machine learning techniques to create a robust and efficient system capable of identifying diseases in crops. Beginning with conventional approaches such as Support Vector Machines (SVMs), we progressively advanced to fully connected neural networks, Convolutional Neural Networks (CNNs), and transfer learning utilizing pre-trained models like MobileNetV2, ResNet50, and EfficientNetB0. The PlantVillage dataset served as the foundation of this study, comprising a curated set of images representing healthy and diseased leaves from potato and bell pepper plants. To ensure optimal model performance, we employed comprehensive data preparation techniques, including resizing, rescaling, and data augmentation, leveraging TensorFlow's advanced preprocessing tools. These steps not only standardized the dataset but also introduced variability to improve generalization and mitigate overfitting. Each model was rigorously trained, validated, and tested, with CNNs achieving superior results due to their ability to learn hierarchical features. The integration of transfer learning further enhanced efficiency, leveraging the pretrained knowledge of deep models trained on extensive datasets. This research underscores the transformative potential of artificial intelligence in agriculture, enabling early disease detection and empowering farmers with actionable insights to improve crop yield and quality. Future efforts will focus on expanding the system's applicability to other crop types and integrating the models into real-world farming workflows.
In our experiments aimed at plant disease detection, we systematically evaluated the performance of various machine learning models using the PlantVillage dataset, which includes images of healthy and diseased potato and bell pepper plant leaves. Starting with Support Vector Machines (SVMs), we reduced the image resolution to 20x20 pixels to analyze the model's ability to extract features despite significant information loss. This was followed by testing fully connected neural networks, where we introduced preprocessing steps such as resizing, rescaling, and data augmentation to assess their impact on the model's generalization capabilities. Moving to more advanced architectures, we implemented Convolutional Neural Networks (CNNs), experimenting with different configurations and hyperparameter tuning strategies, including grid search and KerasTuner, to optimize accuracy. Further, we incorporated transfer learning by fine-tuning pre-trained models such as MobileNetV2, ResNet50, and EfficientNetB0 to leverage their inherent feature extraction capabilities. Each experimental setup was evaluated on training, validation, and test datasets, with CNNs and transfer learning models demonstrating strong performance and robustness. These experiments highlight the strengths and limitations of different methodologies, providing critical insights into their suitability for scalable, real-world plant disease detection systems.
The results reveal a detailed comparison of various machine learning and deep learning models applied to the detection of diseases in potato and bell pepper plants:
This research highlights the potential of artificial intelligence in agriculture, enabling early detection and prevention of crop diseases to enhance yield and quality. Future work will focus on extending the approach to diverse crop types, optimizing resource efficiency, and deploying these models in real-world farming systems for practical use. The integration of advanced computing resources and cloud-based solutions will also enable broader experimentation and scalability of the proposed methodologies.
