• PPS-SAM was proposed for PPS extraction in 10 m Sentinel-2 under sparse samples. • MSPV-Dataset was built to cover diverse terrains with multispectral features. • SEencoder and MaskDecoderHQ were combined to improve segmentation accuracy. • PPS-SAM outperformed baseline models in multi-scene extraction under sparse samples. • PPS area in China increased 69% from 2022 to 2024 while clustering weakened. Driven by global carbon neutrality initiatives, China’s rapid expansion of photovoltaic (PV) power generation necessitates large-scale and precise extraction of photovoltaic power stations (PPS) for effective resource management. However, in 10 m resolution remote sensing imagery, PPS targets frequently exhibit multi-scale and fragmented spatial distributions. Such characteristics often lead to limited model generalization, high omission rates, and elevated commission errors, particularly under sparse-sample conditions. To address these challenges, this study introduces PPS-SAM, a spectrum- and structure-aware extraction framework developed for sparse-sample scenarios based on the Segment Anything Model (SAM). PPS-SAM integrates a Spectral Enhancement Encoder to fuse near-infrared (NIR) and shortwave-infrared (SWIR) bands, thereby improving the spectral separability of PV targets from heterogeneous backgrounds. It also incorporates a High-Quality Mask Decoder to maintain edge integrity and delineate fragmented arrays more effectively. Evaluated on a newly developed 10 m multispectral PPS dataset (MSPV-Dataset), PPS-SAM demonstrated robust segmentation performance with only 11 training samples (F1: 91. 47% ± 0. 13%; mIoU: 91. 40% ± 0. 08%), notably surpassing baseline models trained on the complete 634-sample dataset. Ablation and generalization assessments indicate the effectiveness of each module in enhancing foreground detection and background suppression, with stable performance across diverse, unseen terrains and environmental disturbances (F1: 92. 19%; mIoU: 92. 19%). Applying this framework, nationwide PPS distribution maps for 2022 and 2024 were generated (excluding distributed rooftop systems). The results indicate that the total PPS area expanded from 3, 486. 41 km2 to 5, 900. 89 km2, representing an increase of approximately 70%. Regional analysis shows that Northwest China was characterized by predominantly centralized growth (78. 77%), whereas eastern and southwestern regions exhibited significant distributed expansion. Although national spatial clustering weakened slightly (Global Moran’s I decreased from 0. 4155 to 0. 3112), it intensified within central and southwestern provinces. Land-use transition analysis suggests that new PV installations primarily originated from grasslands (59. 04%) and barren lands (96. 40%), highlighting substantial land-cover conversion. These spatio-temporal patterns underscore regional disparities in PV expansion and shifting spatial structures. This study offers a robust methodological framework for high-precision PPS identification at 10 m resolution under sparse-sample constraints, supporting efficient renewable energy management and evidence-based policy formulation. The dataset is publicly available via Figshare (https: //doi. org/10. 6084/m9. figshare. 29618429).
Zhang et al. (Thu,) studied this question.