Transfer learning-based comparative study of cnn architectures for automated rice disease detection in Bangladesh

In the rural landscape of Bangladesh, rice stands as the cornerstone of sustenance, with 78% of the country's arable land dedicated to its cultivation. However, crop diseases pose a significant threat to this vital agricultural sector, endangering both farmers' livelihoods and national food security. Efficient disease detection is crucial for safeguarding rice yields, but traditional methods have been labor-intensive and time-consuming. The advent of deep learning presents a promising solution, rapidly surpassing conventional approaches in terms of performance. This study presents a transfer learning-based comparative analysis of convolutional neural network (CNN) architectures for automated rice disease detection in Bangladesh. A dataset of 4,330 rice leaf images encompassing three major diseases-bacterial blight, Blast, and tungro-was used to train and evaluate classical CNNs (VGGNet16, ResNet50, InceptionResNetV2), lightweight CNNs (MobileNetV2, EfficientNetB0), and classical machine learning (ML) models (Random Forest and Support Vector Machine). The experimental results reveal that VGGNet16 and InceptionResNetV2 consistently achieved outstanding performance, with accuracies above 98%, while MobileNetV2 reached 96% with a lightweight design suiTable for real-time applications. In contrast, ResNet50, EfficientNetB0, and traditional ML models performed poorly due to overfitting and feature-learning limitations. Among different optimizers tested, RMSProp yielded the best outcomes, particularly with VGGNet16 and InceptionResNetV2, highlighting the importance of optimization in CNN training. Additional experiments on potato, Tomato, and corn datasets demonstrated the models' generalizability, though with reduced accuracy in cross-crop classification. Compared with existing works, the proposed VGGNet16 with RMSProp outperformed most prior methods, establishing itself as a robust and scalable solution for automated rice disease detection and agricultural decision support.