Enhancing Rice Disease Detection with K-Means Segmentation and Comparison of EfficientNetB0, ResNet50, and Xception

Rice cultivation in Indonesia faces persistent threats from various diseases, posing significant challenges to food security and agricultural sustainability. This study addresses the crucial issue of disease identification in rice plants through advanced image classification techniques. Leveraging the EfficientNetB0 method and K-Means segmentation, we propose an optimized approach to classify five common rice diseases: Blast, Brown Spot, Hawar, Kresek, and Narrow Brown Spot. We collected a dataset comprising 1,275 images of diseased rice leaves acquired from rice fields across three provinces in Indonesia. Following manual classification and preprocessing, we applied K-Means segmentation to distinguish disease-affected areas. Subsequently, the EfficientNetB0 architecture was employed for disease classification. Our methodology achieved an accuracy rate of $\mathbf{8 9. 0 6 \%}$ on a test dataset of 128 images, with precision, recall, and $F 1-$ scores of 0.85,0.81, and 0.83 for Blast; $\mathbf{0. 8 6,} 0.75$, and 0.80 for Brown Spot; 0.96,0.96, and 0.96 for Hawar; 0.86, 1.00, and 0.93 for Kresek; and 0.92,0.86, and 0.89 for Narrow Brown Spot, respectively. Comparison with existing architectures revealed the superiority of EfficientNetB0 in terms of accuracy and loss. Additionally, evaluation metrics validate the robustness of our classification model. Our study contributes to the development of diagnostic systems for rice diseases, offering a valuable tool for farmers to identify and manage crop diseases effectively.