PlantOHealth: Comparative Evaluation of Deep Learning Models for Plant Disease Detection Using Leaf Images

Abstract

The agricultural sector worldwide faces significant challenges from plant diseases that threaten crop yields, food security, and economic stability. This study introduces "PlantOHealth," a detailed comparative analysis of five models for detecting plant diseases using leaf images, leveraging advancements in deep learning. The models evaluated include a basic Convolutional Neural Network (CNN) and four transfer learning frameworks: VGG16, VGG19, MobileNetV2, and ResNet, all utilizing the PlantVillage dataset with balanced classes achieved through oversampling and undersampling techniques. MobileNetV2 emerged as the most effective, achieving an accuracy of 99.40% while maintaining computational efficiency for resource- constrained environments, followed by the CNN with an accuracy of 98.68%. VGG19 and VGG16 attained accuracies of 98.92% and 97.24%, respectively, while ResNet recorded the lowest at 96.24%. Graphical analyses provided deep insights into model performance and highlighted the trade-offs between accuracy and computational demands. "PlantOHealth" contributes to the integration of AI in agriculture, offering actionable insights for researchers and practitioners, while future work will focus on exploring advanced techniques like ensemble learning to enhance plant disease detection systems further, ultimately supporting sustainable agricultural practices and improving food security.