Interpretable mango leaf disease detection using a hybrid CNN–transformer model with GLCM features

Mango leaf diseases significantly hinder crop yield and food security in tropical regions, necessitating accurate and timely diagnostic tools. Traditional visual inspection methods are often subjective, time-consuming, and lack scalability, while existing deep learning approaches struggle with dataset imbalance, generalization limitations, and interpretability issues. To address these challenges, we propose ViX-MangoEFormer, a hybrid model that combines convolutional layers with self-attention mechanisms for robust classification of eight mango leaf conditions. The architecture incorporates MBConv4D and MBConv3D modules to capture both localized textures and global patterns, while GLCM-based statistical features enhance discriminatory power. Additionally, a stacking ensemble (MangoNet-Stack), comprising five pretrained models, is introduced as a comparative benchmark. Both models are trained and validated on a merged dataset of 25,530 images from four public sources, including balanced and imbalanced classes. Grad-CAM-based explainability is natively integrated to offer real-time visual rationales. Experimental results demonstrate that ViX-MangoEFormer achieves an F1 score of 99.78% and MCC of 99.34%, outperforming all baseline models. Furthermore, cross-domain tests reveal strong generalization to morphologically similar crops. A web application has been deployed to deliver real-time predictions with transparent explanations, providing an effective and interpretable solution for precision agriculture.