Wang, Siyuan, Cai, Bowen, Luo, Zhiwen  ORCID: https://orcid.org/0000-0002-2082-3958, Hou, Dongyang, Zheng, Xiaoyu, Liu, Jinglei and Shao, Zhenfeng
2026.
Edge-preserving and uncertainty-guided domain adaptation for weakly supervised rooftop photovoltaic mapping.
ISPRS Journal of Photogrammetry and Remote Sensing
240
, pp. 31-53.
10.1016/j.isprsjprs.2026.07.008
|
Preview |
PDF
- Accepted Post-Print Version
Available under License Creative Commons Attribution. Download (3MB) | Preview |
Abstract
Accurate, large-scale mapping of rooftop photovoltaic (PV) systems from high-resolution remote sensing imagery is critical for urban energy planning and grid management. However, existing deep learning methods suffer from high annotation costs associated with pixel-level labeling and significant performance degradation due to cross-domain shifts in geography and sensors. To address these challenges, this paper presents an integrated edge-preserving and uncertainty-guided domain adaptation framework (EPUDA) for weakly supervised rooftop PV mapping. By transferring knowledge from densely annotated source domains to target domains with only sparse point annotations, EPUDA achieves low-cost cross-domain segmentation. First, a semantic-structural dual-path decoding network is employed to model regional semantics and edge structures separately. Subsequently, an edge-preserving adversarial adaptation mechanism aligns features by dynamically focusing on structurally informative boundaries and hard regions with inconsistent semantic predictions, thereby enhancing boundary stability across domains. Furthermore, to tackle the sparsity of point supervision, an uncertainty-guided pseudo-label consistency learning module is incorporated to generate more reliable pseudo-labels via probabilistic consistency evaluation and confidence-aware selection. To support this study, we construct the China Rooftop Photovoltaic Segmentation Dataset (CRPVS) for weakly supervised learning, covering 49 major cities. Experiments across four cross-domain scenarios demonstrate that EPUDA outperforms 15 representative comparative methods, improving the Intersection over Union (IoU) by 1.75%–2.65% over the strongest competing method in each scenario. This study offers a practical, low-cost solution for large-scale rooftop PV mapping.
| Item Type: | Article |
|---|---|
| Date Type: | Publication |
| Status: | Published |
| Schools: | Schools > Architecture |
| Additional Information: | RRS policy applied |
| Publisher: | Elsevier |
| ISSN: | 0924-2716 |
| Date of First Compliant Deposit: | 18 August 2026 |
| Date of Acceptance: | 15 July 2026 |
| Last Modified: | 18 Aug 2026 10:00 |
| URI: | https://orca.cardiff.ac.uk/id/eprint/188500 |
Actions (repository staff only)
 |
Edit Item |
Dimensions
Dimensions