DBSF: dual-branch boundary-supplemented framework for photovoltaic power station extraction

As global efforts to address the energy crisis intensify, the adoption of photovoltaic (PV) power stations has increased rapidly, especially in China. This trend underscores the need for accurate monitoring of PV station distribution. Current methods encounter challenges in delineating boundaries, especially for closely spaced installations. The dual-branch boundary-supplemented framework (DBSF) is proposed to extract PV power stations from Landsat 8 imagery. DBSF features a dual-branch architecture that integrates convolutional neural networks (CNNs) for local feature extraction and transformers for global feature learning. Continuous morphological operations are applied prior to the local feature learning branch to enhance boundary features. The local feature learning branch incorporates a boundary enhancement module (BEM) and an auxiliary boundary loss function to improve boundary accuracy. We evaluated DBSF on Landsat 8 images from 16 regions in China. The results show that DBSF outperforms six existing models, with a 2.24% improvement in mean intersection over union (mIoU). The results highlight the effectiveness of DBSF in complex PV station extraction tasks. The implementation code will be publicly released to support practical PV station extraction applications upon acceptance of this manuscript.