AHM-YOLO: an adaptive hierarchical multi-scale model for rice disease detection

Rice disease detection is crucial for ensuring crop yield and food security, yet remains challenged by lesion scale variation, high inter-class similarity, and weak early-stage features. To address these issues, this paper proposes Adaptive Hierarchical Multi-scale YOLO (AHM-YOLO) based on YOLOv8. The model introduces a lightweight adaptive downsampling module (ADown) to mitigate information loss and feature aliasing, enhancing robustness to scale and morphological variations. A hierarchical channel attention mechanism (HCA) is designed to capture subtle pathological cues and improve discrimination among similar disease categories. Additionally, an attention-guided multi-scale switchable atrous convolution module (AGMSAConv) is designed to enhance spatial continuity and adaptive receptive field modeling, enabling joint learning of fine-grained textures and mid-scale structures. Experiments on RD1 and RD2 demonstrate that AHM-YOLO stably surpasses YOLOv8n. It achieves F1-score improvements of 1.1 and 1.4 percentage points, as well as mAP50:95 increases of 3.8 and 2.1 percentage points on the two datasets respectively, with its GFLOPs only accounting for 76.5% of YOLOv8n. These results demonstrate a favorable balance between accuracy and efficiency, achieving consistent performance across the two evaluated datasets.