Lightweight dual-stage feature refinement for black gram leaf disease classification using ConViTSE

Black gram, also known as urad bean, is an economically crucial crop widely cultivated in India, particularly in the central and southern regions. However, black gram is highly prone to multiple leaf diseases, resulting in considerable crop losses and economic challenges for farmers. Manual disease identification is slow and often unreliable, necessitating the development of automated disease detection methods. In this study, we propose ConViTSE, a lightweight hybrid deep learning architecture specifically designed for black gram leaf disease classification. ConViTSE integrates ConvMixer, Vision Transformer (ViT), and Squeeze and Excitation (SE) blocks to effectively extract and refine both local and global features. The model introduces Local Channel Attention Refinement (LCAR) and Global Channel Attention Refinement (GCAR) modules to enhance feature representation at different hierarchical levels. Extensive studies show that ConViTSE achieves a leading classification accuracy of 99.30% on the black gram dataset, outperforming traditional deep learning models. Furthermore, ConViTSE exhibits robust cross-domain generalization, achieving accuracies of 98.75% for rice, 98.20% for maize, and 95% for wheat, highlighting its potential for widespread adoption in precision agriculture. ConViTSE enhances disease detection accuracy while remaining computationally efficient, making it a practical tool for real-time disease management in diverse agricultural environments.