A Comparative Study for Rice Blast Disease Detection Using Deep Transfer Learning

Abstract

Rice is a fundamental food crop for more than half of the global population, especially in agrarian countries like Bangladesh. However, fungal diseases such as leaf blast, neck blast, and node blast, caused by Magnaporthe oryzae, threaten rice production and endanger food security. This study proposes an automated rice leaf disease detection system using deep transfer learning and multi-level feature extraction to improve early diagnosis accuracy. A real-time dataset of 1,500 annotated rice leaf images was collected from the fields of Bogura and Joypurhat and categorized into three major disease classes. Preprocessing techniques, including resizing, normalization, grayscale conversion, Gaussian blur, and advanced augmentation methods, were applied to enhance dataset quality and diversity.Six state-of-the-art pre-trained Convolutional Neural Network (CNN) models—EfficientNetV2S, ResNet50V2, MobileNetV2, VGG16, DenseNet121, and Xception—were fine-tuned and evaluated. Feature extraction was performed at multiple levels (shallow, texture-based, and deep semantic layers) to capture detailed disease characteristics. Among the models, EfficientNetV2S achieved the highest classification accuracy of 99.28%, outperforming others in both generalization and training stability. The proposed system offers a scalable, high-performance solution suitable for real-time deployment in rural and low-resource environments. This work contributes significantly to smart agriculture, enabling farmers to detect diseases early, reduce crop loss, and adopt sustainable management practices.