Photovoltaic System Segmentation Algorithm via Transformer and Multi-Scale Convolution

Semantic segmentation of photovoltaic (PV) system remote sensing images has been widely applied in distributed energy forecast and assessment. To improve PV segmentation accuracy, we propose a PV segmentation algorithm with a dual-branch encoder concatenating a Transformer and multi-scale convolution modules (MICM) to enhance global-local feature representation for remote sensing images. A three-stage progressive training strategy is proposed, coupled with confidence-thresholded pseudo-label refinement and multi-loss optimization, which leverages unlabeled data to mitigate overfitting under limited labeled samples while improving generalization performance. Comparative experiments are conducted to verify the segmentation precision of the proposed algorithm under different annotation ratios, with the proposed method showing a 6.84\% improvement over DeepLabv2 with full annotation ratios.