Building an Operational Framework for Leaf Disease Detection

Leaf disease detection (LDD) is a main component of crop management systems. An operational LDD system requires a method that enables leaf detection in plant images and an LDD model that identifies whether the detected leaves present signs of disease and if so which type. The use of Convolutional Neural Networks (CNN) has enabled the development of effective LDD models using available rich leaf disease datasets. More recently, Vision Transformers (ViT) and EfficientNets have emerged as a competitive alternatives to conventional CNNs. Progress in leaf detection has not proceeded at the same pace. In addressing the development of an operational LDD framework, the goal of this study is threefold: (1) provide a comparison of CNN, ViT and EfficientNet LDD algorithms using datasets totaling 48,546 leaf images from the lab and the field, (2) develop a YOLO-based detection model that identifies leaves in plant images, and (3) detail the development of a drone-based LDD framework that integrates models of leaf detection and leaf disease detection. Our experimental results show that while CNN, ViT and EfficientNet LDD models all perform well on standard LDD datasets, with EfficientNets showing a marginal advantage, the ViT classifier demonstrates superior performance when run in conjunction with leaf detection in an operational setting.