LeafVisionNet: A deep learning-based approach for the classification of black gram leaf disease using attention-driven and semi-local feature boosted SqueezeNet

Agriculture is essential for sustaining the global food supply, supporting economies, and promoting environmental balance. Among pulse crops, black gram is highly valued for its nutritional content and contribution to agricultural economies. However, its productivity is significantly impacted by diseases that cause loss of yield. Current methods of detecting disease are labour-intensive, prone to error, and lack scalability. In this study, a deep learning framework for black gram leaf disease classification is developed. LeafVisionNet utilizes a modified SqueezeNet backbone, chosen for its lightweight and efficient architecture, enhanced with a Multi-Path Semi-Local Feature Enhancement (MPSLFE) block and a Coordinate Attention (CA) module. This architecture effectively captures integral semi-local features, while also focusing on salient regions through integrated spatial and channel attention mechanisms. The integration of XAI, further improves the interpretability of the network and provides insights into disease-affected regions. The proposed LeafVisionNet is the first work to be evaluated on the IDBGL dataset, reaching 99.02 % accuracy and 98.96 % F1-score, outperforming state-of-the-art architectures. The LeafVisionNet was validated across four additional datasets, BAU-Insectv2, MangoLeafBD, BPLD, and BananaLSD, demonstrating its robustness and generalizability, with accuracies of 97.14 %, 98.00 %, 98.09 % and 99.37 %, respectively.