Rice leaf disease classification: A comparative evaluation of CNNs and vision transformers

Abstract

Rice is a staple crop worldwide but is highly vulnerable to leaf diseases that threaten food security. Traditional manual inspection is slow, labor-intensive, and error-prone, highlighting the need for automated detection. This research introduces a comparative analysis between convolutional neural networks (CNNs) and vision transformers (ViTs) for multi-class rice leaf disease classification. Using an 11-class dataset of bacterial, fungal, viral, pest-related, and healthy leaves, we evaluated ResNet50, MobileNetV2, BEiT, DeiT, Swin Transformer V2, and Swin-Tiny V2. Results show transformer models outperform CNNs, with the quantized Swin-Tiny V2 Transformer achieving 99.56% accuracy while reducing model size to 26.89 MB, making it suitable for mobile and IoT deployment. To enhance interpretability and understand the model's decisions, Grad-CAM visualizations highlight important regions, ensuring transparency in model predictions. The findings demonstrate the potential of lightweight, explainable transformers for realworld agricultural IoT applications in precision farming.