PVSAM: Adapting geometric prompts to segment anything model for photovoltaic detection in remote sensing imagery

Accurate photovoltaic (PV) detection from high-resolution remote sensing imagery plays a crucial role in assessing electricity generation potential and facilitating renewable energy management. While deep learning-based approaches have achieved significant performance in PV segmentation tasks, existing methods predominantly rely on single-scenario datasets to represent the specific feature distributions, limiting their capability to simultaneously generalize size and edge features of PV systems across diverse scenarios. To address this limitation, we propose PVSAM, a novel segmentation model that integrates zero-shot generalization capability of the Segment Anything Model (SAM) with geometric prompts tailored for PV panels. In PVSAM, we incorporate two specialized prompt modules as the knowledge-specific adapter to guide SAM for multi-scenario PV feature learning. Specifically, to improve adaptability to PV panels of various sizes, we construct a multi-scale prompt module that employs a multi-branch convolutional structure to effectively aggregate feature information with different receptive fields. To leverage the structural regularity of PV panels for refined semantic segmentations, we introduce an edge pyramid prompt module that explicitly reinforces multilevel shape features while strengthening the model's sensitivity to high-frequency boundary information. Extensive experiments on the PV01–03-08 (PV01, PV03, PV08), HRPVS and PVP datasets demonstrate that PVSAM can obtain the superior detection performance and outperform existing state-of-the-art methods with impressive F1 and IoU accuracy exceeding 90 % overall. Furthermore, the PVSAM method exhibits remarkable generalization performance in cross-scenario PV detection tasks, providing an effective solution for large-scale energy infrastructure monitoring.