Rice Leaf Disease Detection Using a Hybrid Deep Learning Architecture

Abstract

The occurrence of diseases in rice leaves is a crucial problem for agricultural productivity and food security, requiring the development of accurate algorithms to reliably detect these conditions. Although deep learning methods, especially Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs), have already shown promising results in detecting plant diseases, individual frameworks still struggle to extract both local texture information and the global context of the input data. Thus, the current study aims at designing a custom hybrid CNN+ViT model for classifying rice leaf diseases based on combining the ability of CNNs to extract spatial features and the global attention mechanism of ViTs. The proposed framework was compared against different transformer-based networks, including standard ViT, Fine-Tuned ViT, and Multiscale ViT models. In comparison to the baseline models, the proposed CNN+ViT framework showed the best classification accuracy with 98.47% for the training set, 98.86% for the validation set, and 97.34% for the test set.