GF7S: A large-scale GaoFen-7 stereo dataset and self-supervised adaptation of FoundationStereo for high-resolution DSM generation
Rattachement africain : cn. Niveau de preuve : code pays fourni par la source.
Le résumé fourni par la source
Accurate inversion of high-resolution digital surface models (DSMs) is crucial for urban science and environmental monitoring. Fully supervised deep learning has been widely adopted for DSM generation from stereo satellite imagery, but its dependence on costly ground-truth labels limits scalability in label-scarce regions. Here, we introduce GF7S, a large-scale unlabeled GaoFen-7 stereo dataset comprising 121 K stereo pairs from 150 scenes at sub-0.8 m resolution, and evaluate a self-supervised adaptation of a large-scale foundation model, FoundationStereo (FS), for 1-m DSM generation. The adaptation uses established self-supervised stereo constraints, including spectral reconstruction consistency, left–right disparity consistency, and edge-aware disparity smoothness. Without using ground-truth disparity labels, self-supervised FSLarge (the large variant of FS, ∼357 M parameters with only ∼ 37 M trainable), achieves performance close to its fully supervised counterpart, and incurs modest increases in end-point error (EPE) from 1.215 to 1.502 pixels on WHU-Stereo (+0.287) and from 0.653 to 0.933 pixels on US3D (+0.280). Comparisons with existing methods further show that FSLarge provides the best self-supervised accuracy across the evaluated datasets. The 1-m DSMs demonstrate promising results in building-scale height estimation with an RMSE of 5.51 m evaluated using ICESat-2-derived heights and in tree canopy height mapping, showing good agreement (Pearson’s R of 0.45 and RMSE of 5.89 m) with a leading global canopy height dataset. This work presents a self-supervised adaptation of foundation model, FoundationStereo, for high-resolution satellite stereo matching, using established stereo constraints to guide training on the proposed GF7S dataset. Data and code will be available at https://github.com/lauraset/Self-Supervised-Satellite-Stereo-Matching-Model .
Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.
Le contrôle bibliographique ouvert
DOI retrouvé dans Crossref DOI retrouvé ; titre concordant.
- Titre Crossref
- GF7S: A large-scale GaoFen-7 stereo dataset and self-supervised adaptation of FoundationStereo for high-resolution DSM generation
- Date Crossref
- 01/10/2026
- Éditeur
- Elsevier BV
- Type
- journal-article
Ce recoupement confirme des métadonnées liées au DOI. Il ne confirme ni la méthode ni les conclusions de l’étude, et il ne compte pas comme une seconde source scientifique indépendante.
Où se fait cette recherche
-
Chinese Academy of Sciences pays non établi dans la noticeOrganisme public
-
Institute of Geographic Sciences and Natural Resources Research pays non établi dans la noticeStructure de recherche
-
University of Chinese Academy of Sciences pays non établi dans la noticeUniversité ou école supérieure
-
Beijing Academy of Artificial Intelligence pays non établi dans la noticeInstitution
-
Shanghai Artificial Intelligence Laboratory pays non établi dans la noticeStructure de recherche
-
Wuhan University pays non établi dans la noticeUniversité ou école supérieure
-
College of Resources and Environment pays non établi dans la noticeUniversité ou école supérieure
-
State Key Laboratory of Resources and Environmental Information System pays non établi dans la noticeStructure de recherche
-
School of Remote Sensing and Information Engineering pays non établi dans la noticeUniversité ou école supérieure
Chinese Academy of Sciences, Institute of Geographic Sciences and Natural Resources Research et University of Chinese Academy of Sciences, avec 6 autres affiliations.
Une affiliation ne permet pas de déduire la nationalité d’un auteur.