Paddy Leaf Disease Classification using an Enhanced CoAtNet with MobileNetV2 and Triple Attention

This paper presents a novel deep learning approach for paddy leaf disease classification that integrates MobileNetV2, CoAtNet, and a Triple Attention mechanism. The work focuses on addressing the challenges of class imbalance and limited dataset diversity by applying advanced data preprocessing steps, including duplicate removal and powerful augmentations using the Albumentations library. The model is implemented in PyTorch and trained with AdamW optimizer, Focal Loss, and a CosineAnnealingLR scheduler, while 5-fold cross-validation ensures reliable evaluation. Experimental results on the Kaggle Paddy Disease dataset demonstrate that the proposed architecture significantly outperforms baseline models, achieving 91.6% accuracy, along with improvements in precision, recall, and F1-score. This research highlights the importance of integrating lightweight feature extractors with attention-based architectures to improve robustness and efficiency in agricultural disease detection. Future directions include expanding datasets, exploring self-supervised learning, and deploying the model to mobile/edge devices for real-time field applications.