You only look once-based neural network and grading method for peach leaf shot hole disease detection under natural conditions.

The frequent occurrence of peach leaf shot hole disease severely affects peach yield and fruit quality. Under natural conditions, the detection accuracy of existing methods is often compromised by substantial variability in disease characteristics, complex environmental backgrounds, and dense leaf coverage that obscures lesion features. To improve detection performance under such conditions and to enable quantitative assessment of disease severity, this study proposes a YOLO-PSH (You Only Look Once-peach shot hole) model for peach leaf shot hole disease detection. The proposed model addresses three critical challenges: multi-scale feature extraction, suppression of complex background interference, and robustness to overlapping and occluded leaves. First, the GhostConv and AConv modules are introduced to reduce computational redundancy and enhance feature compression while preserving fine-grained details. Second, the A2C2f and RepNCSPELAN4 modules, combined with a regional attention mechanism and parameter reorganization strategy, are used to strengthen semantic representation in lesion regions and mitigate background interference. Third, a multi-scale feature fusion neck that integrates SPPELAN (spatial pyramid pooling with efficient layer aggregation network) and C3k2 is designed to improve adaptability to overlapping and occluded disease features. In addition, a disease severity grading method is developed based on detection results to enable regional severity assessment. To evaluate model performance, a PSHData (peach shot hole data) dataset was constructed, consisting of 800 images of peach leaves affected by shot hole disease and containing 9, 183 annotated instances. Experimental results demonstrate that the proposed YOLO-PSH model achieves an average detection accuracy of 81.49%. The model contains 1.82 million parameters, requires 8.71 GFLOPs, and has a model size of 7.00 MB. The average inference time is 7.1 ms per image. For disease severity grading, the classification accuracies for healthy, mild, moderate, and severe categories are 81.82%, 60.00%, 72.22%, and 91.67%, respectively. These results indicate that YOLO-PSH exhibits strong robustness and adaptability under natural conditions, demonstrating its potential for accurate detection and severity assessment of peach leaf shot hole disease across varying scales, complex backgrounds, and dense leaf occlusion scenarios.