A Hybrid Deep Learning Framework with Attention Mechanism for Robust and Explainable Plant Disease Classification

Background: For agricultural productivity, Plant Diseases are a significant problem all over the world and which requires development of robust, accurate and scalable automated diagnostic systems. The most recent deep learning methods have been promising, but most current methods lack generalization since they were trained based on single-domain datasets and are not interpretable. Methods: A hybrid deep learning architecture is proposed that integrates convolutional neural networks, attention models and feature refinement with transformers to improve classification and localization of plant disease. Result: The proposed framework achieves 99.08% and 98.88% for validation and test accuracy, respectively, surpassing traditional baseline models, such as CNN, MobileNetV3, EfficientNet-B0 and ResNet18.