RSD-YOLO: An improved YOLOv7-tiny framework for oat disease severity identification with integration of ReXNet and decoupled head

In precision agriculture, accurate and timely identification of plant disease severity is essential for optimizing crop yield and health. However, current methods often face challenges such as high computational cost and reduced accuracy in resource-constrained environments, limiting their practical use on farms. To address these limitations, we propose RSD-YOLO, an improved YOLOv7-tiny model that integrates a Regularized Xception-based Network (ReXNet), a Slim-Neck module, and a Decoupled Head—together forming the RSD design. We construct a dataset of 1,010 oat leaf images, categorized into five severity levels and annotated by experts. RSD-YOLO achieves 91.6% precision, 90.8% recall, and 88.5% mAP@0.5, significantly outperforming YOLOv7-tiny by up to 10%, while maintaining a computational cost of only 11.2 GFLOPs. Recent studies have applied lightweight models such as EfficientSAM and SwiftFormer for crop health monitoring on drones and edge devices. However, these models often struggle to balance accuracy and efficiency. In contrast, RSD-YOLO achieves higher performance with lower computational cost, making it well-suited for real-time deployment in agricultural environments. • RSD-YOLO integrates ReXNet, Slim-Neck, and Decoupled Head for superior efficiency. • Dataset covers five disease levels, augmented for robust model generalization. • RSD-YOLO, an improved YOLOv7-tiny model, enhances oat disease severity detection. • RSD-YOLO has 91.6% precision, 90.8% recall, 88.5% mAP@0.5 for identification. • RSD-YOLO is ideal for deployment on low-power devices or edge systems in farms.