Deep Learning in Rice Disease Detection: A Systematic Analysis of Techniques, Strategies and Model Generalization

Abstract

Rice diseases significantly reduce crop yield and threaten global food security, particularly in developing agricultural systems. Deep learning approaches, mostly CNN, have shown promising performance in automated rice disease detection. However, despite high reported accuracies, concerns remain regarding model overfitting, limited dataset diversity, and insufficient evaluation of generalization performance under real-world conditions. This research provides a structured review of deep learning-based rice leaf disease detection methods published between 2018 and 2026. Using a structured search and predefined selection criteria, relevant studies were analyzed based on method/model architecture, and performance. The review identifies dominant adoption of transfer learning models such as VGG, ResNet, DenseNet, EfficientNet, and YOLObased detection frameworks. While most studies report classification accuracies above 90%, demonstrating the capability of deep learning techniques in accurately identifying rice diseases from images. The analysis further reveals that explicit application and systematic evaluation of regularization techniques are rarely emphasized. Similarly, data augmentation strategies are commonly applied but often lack detailed investigation regarding their impact on generalization performance. These gaps indicate the need for focused research on improving robustness and preventing overfitting in rice disease detection  systems. The findings provide a foundation for developing more generalized and reliable deep learning models for practical agricultural deployment.