<![CDATA[Rice plant diseases represent a significant challenge to global food productivity. This study aims to optimize a rice disease classification pipeline using the EfficientNet-B0 architecture combined with intensive image augmentation strategies and transfer learning. The dataset comprises 10,407 rice leaf images categorized into 10 classes, including healthy conditions and nine types of diseases. Augmentation strategies such as random rotation, color jittering, and random resized cropping were implemented to enhance model robustness against diverse field conditions. Evaluation results demonstrate that the model achieved outstanding performance, with a Top-1 Accuracy of 96.25% and a Top-5 Accuracy of 99.90%. Grad-CAM++ analysis validated that the model accurately focuses feature extraction on pathological leaf areas. t-SNE visualization revealed clear feature cluster separation between classes, further supported by ROC curve AUC values reaching 1.00 for the majority of categories. This research confirms that the proposed pipeline is highly reliable for early rice disease detection and holds significant potential for mobile device implementation to assist farmers.]]>
Copyrights © 2025
