A novel approach to enhance rice foliar disease detection: custom data generators, advanced augmentation, hybrid fine-tuning, and regularization techniques with DenseNet121

Abstract

Rice leaf diseases impact crop yield, leading to food shortages and economic losses. Early, automated detection is essential but often hindered by accuracy challenges. This study contributes to improving model robustness against diverse and adversarial inputs by proposing a custom data generator that applies Albumentation-based advanced augmentations, such as Gaussian blur, noise addition, brightness/contrast adjustments, and coarse dropout, to enhance model generalization. Five deep learning architectures—simple convolutional neural network (CNN), ResNet50, EfficientNetB0, Inception v3, and DenseNet121—were evaluated for classifying six categories: bacterial blight, brown spot, leaf blast, leaf scald, narrow brown spot, and healthy leaf. A hybrid model approach is proposed, fine-tuning the DenseNet121 model by unfreezing its last 20 layers, which balances transfer learning benefits with domain-specific feature extraction. Regularization techniques, including L2 regularization and a reduced dropout rate, are incorporated to control overfitting. Additionally, a custom learning rate scheduler is proposed to promote stable training. DenseNet121 achieved the highest performance, with an accuracy of 98.41%, demonstrating the effectiveness of these advanced augmentation and tuning strategies in rice leaf disease classification.