AGC-YOLO: adaptive gated channel network for rice leaf disease detection

Accurate detection of rice leaf diseases is critical for intelligent crop protection, yet class imbalance, degradation of multi-scale feature information, and the limited discriminative capability of feature fusion continue to hinder effective monitoring under field conditions. To address these challenges, this study presents AGC-YOLO, an improved detection model based on the YOLOv8s framework. First, a semantic-aware data augmentation network termed SAICP Net is devised to alleviate class imbalance and the scarcity of cross-class co-occurrence samples through a weakly supervised soft spatial prior and semantically constrained cross-category instance pasting. Second, a gated hybrid pooling module named GHP-SPPF is proposed to adaptively integrate max pooling and soft pooling via a learnable gating mechanism, thereby strengthening the discriminative representation of multi-scale lesions. Finally, a channel-adaptive fusion module termed CAFF is constructed to replace standard concatenation with channel-wise learnable weights and gated residual refinement, boosting the discriminative power of cross-scale feature fusion. Experimental results demonstrate that AGC-YOLO achieves an mAP@0.5 of 93.5 ± 0.6% on the rice disease dataset, representing a 3.0 percentage-point improvement over the baseline, and additional experiments on two public datasets further demonstrate the effectiveness of the proposed method for disease detection.