Multi-class soiling recognition method for photovoltaic panels based on improved U-Net

As photovoltaic (PV) power plants expand, module surface contamination critically reduces their efficiency and reliability; however, precise classification of contamination types remains underexplored. To address this, we propose an enhanced U-Net-based deep learning model for accurately identifying surface deposits on PV panels. Our method employs a two-stage semantic segmentation approach consisting of coarse recognition (CR) and fine recognition (FR). In the CR stage, a convolutional block attention module highlights areas of contamination, while an atrous spatial pyramid pooling module effectively captures multi-scale features, achieving a pixel accuracy (PA) of 95.2%. Subsequently, the FR stage integrates a conditional random field recurrent neural network to improve boundary pixel identification, ensuring smoothness and consistency at segment edges, thereby increasing PA to 95.7%. We constructed a specialized dataset of PV panel surface contaminants, evaluated model performance using five metrics—pixel accuracy (PA), Recall, Precision, F1 score, and mean Intersection over Union, and conducted comparative experiments involving five neural network architectures. Experimental results demonstrate that the proposed model achieves competitive or superior performance across all metrics, providing valuable insights for intelligent monitoring and maintenance of PV power plants.