Lightweight object detection model for False Smut detection in rice leaf with Eigen-cam Interpretability

Rice is an essential world food crop, especially in Asia, and sustains food security and rural livelihoods. False smut, a widespread fungal disease, has a drastic impact on rice yield and quality if not detected early. Manual detection methods by conventional means are time-consuming, error-prone, and impractical for mass deployment. In this paper, we proposed a lightweight and interpretable deep learning model for the computer-aided detection of false smut. We have compared various YOLO-based object detection models, investigating performance metrics like Mean Average Precision (mAP), inference time, and model complexity. Among the models investigated, YOLOv8n has the optimal balance between accuracy and computational cost and is hence best suited for real-time deployment on resource-restricted platforms. To make the model more interpretable, Eigen-CAM is used to project attention regions responsible for the model’s predictions. The proposed approach facilitates accurate, real-time, and interpretable disease detection with high potential for real-world application in precision agriculture.