Rice Leaf Disease Classification Using Convolutional Neural Networks Enhanced with Convolutional Block Attention Module (CBAM)

Abstract

Rice (Oryza sativa L.) is a staple food for over 270 million people in Indonesia, yet its productivity is continuously threatened by major diseases such as Tungro, Blast (Pyricularia oryzae), and Bacterial Leaf Blight (BLB), which can reduce yields by up to 70% and lead to crop failure. Traditional disease identification relies on manual visual observation, which is subjective, expertise-dependent, and inefficient for large-scale farming. This study aims to develop and compare four deep learning model variants ResNet-50, EfficientNet-B4,ResNet-50+CBAM,and EfficientNet-B4+CBAMfor automated classification of rice leaf diseases from digital images. A quantitative experimental approach was employed using a dataset of 5,702 images across four classes: Healthy, Tungro, Blast, and BLB. All models utilized transfer learning with ImageNet-pretrained weights, and the Convolutional Block Attention Module (CBAM) was integrated to enhance feature discrimination through channel and spatial attention mechanisms. The results demonstrate that ResNet-50 + CBAM achieved the best performance with 98.86% accuracy, 98.8% precision, 98.8% recall, and 98.8% F1-score, significantly outperforming the baseline ResNet-50 (50.6% accuracy). It can be concluded that the integration of CBAM with CNN architectures substantially improves classification accuracy by directing the model’s focus toward disease-relevant leaf regions while suppressing irrelevant background features. These findings provide a scalable and accurate diagnostic framework with strong practical implications for the development of mobile-based diagnostic tools to support real-time disease detection and precision agriculture decision-making by farmers in the field.