Improved YOLOv8 method for photovoltaic panel surface state detection integrating spatial-to-depth convolution and Wise-IoU

Photovoltaic panel surface coverage by dust and snow can reduce energy yield, thereby weakening the effective utilization of solar power and the environmental benefits of photovoltaic (PV) systems. Accurate and efficient visual inspection is therefore important for sustainable PV operation and maintenance. This study proposes an improved YOLOv8-based model for PV panel surface state detection to address fine-grained feature loss during downsampling and unstable bounding-box regression under low-quality samples. A cleaned and relabeled public dataset was constructed and divided into training, validation, and test sets, with three categories: Dust, Non Defective, and Snow. The proposed method replaces part of the stride-2 downsampling convolutions in the backbone with Spatial-to-Depth Convolution (SPD-Conv) modules and introduces Wise Intersection over Union (WIoU) loss in the bounding-box regression branch. Under a unified training setting, the proposed model achieved the best overall performance among the compared lightweight models, with mAP50, mAP50-95, Precision, and Recall of 0.7657, 0.5562, 0.7696, and 0.7531, respectively. Compared with YOLOv8n, mAP50 and mAP50-95 improved by 0.0150 and 0.0225. Ablation experiments further show that SPD-Conv and WIoU provide complementary gains in feature representation and localization quality. The proposed method offers an effective and compatible solution for PV surface-state monitoring and supports more reliable, resource-efficient, and sustainable solar energy system operation.