Label-efficient global centralized and distributed photovoltaic segmentation via hierarchical CLIP-calibrated active learning

Accurate global photovoltaic (PV) mapping is essential for renewable energy planning, grid integration, and carbon-neutral transition assessment. AI-based remote sensing segmentation offers an efficient way to update PV inventories, but pixel-wise annotation of PV imagery remains expensive at large scale, especially across distributed rooftop systems and centralized utility-scale plants with different resolutions and geographic backgrounds. A key practical challenge is therefore to achieve high-accuracy PV segmentation with the smallest possible annotation budget under realistic labeling constraints. To address this challenge, we propose a hierarchical CLIP-calibrated active learning (AL) strategy for label-efficient global PV segmentation. Due to the heterogeneous appearance of PV scenes, CLIP is introduced as an external visual-semantic calibrator, and sample acquisition is organized into three stages: semantic coverage in the CLIP feature space, hybrid-uncertainty hard-example mining, and feature-level redundancy removal. Using SegFormer-B3 as the segmentation baseline, we evaluate the method on two worldwide benchmarks covering distributed rooftop PV at 0.1–1.0 m resolution and centralized utility-scale PV at 10–20 m resolution. Under the same annotation budget, our method consistently outperforms classic AL baselines, including Entropy, BALD, and Core-set, across both benchmarks. Practical deployment analysis shows that, for an IoU ≥ 90 % target, our method reduces the required annotation budget by 39.1–78.8% for distributed PV and 15.1–34.8% for centralized PV, corresponding to estimated manual-labeling savings of approximately 260 h/$5.2k and 114 h/$2.28k relative to Random sampling, respectively. These results support scalable, low-cost PV inventory updates for deployment monitoring, grid planning, and renewable energy planning under constrained labeling resources.