Paddy Leaf Disease Classification Using EfficientNet-B0 with Squeeze-and-Excitation Attention Outperforms B1 and B2 for Multi-Class Tasks

Abstract

Paddy leaf diseases cause significant yield losses in rice production, threatening global food security. This study presents a comparative evaluation of EfficientNet B0, B1, and B2 architectures integrated with Squeeze-and-Excitation (SE) attention for multi-class paddy leaf disease classification using the Paddy Doctor dataset (10,407 images, ten disease classes). Models were fine-tuned using aggressive data augmentation, three-phase progressive training (head-only, partial unfreezing, full fine-tuning), and cosine annealing learning rate scheduling. EfficientNet-B0 with SE attention achieves 94.50% validation accuracy, substantially outperforming B1 (89.29%) and B2 (90.93%) despite having only 5.3M parameters. B0's superior performance is attributed to optimal parameter efficiency and effective channel-wise feature recalibration, minimizing overfitting on moderately sized datasets. Grad-CAM visualizations confirm disease-relevant region focus. This lightweight solution (5.3M parameters) is suitable for precision agriculture in resource-limited settings.