DGS-CF: a DSM and sparse-view 3DGS co-optimization framework for georeferenced 3D building reconstruction
Satellite-based DSMs enable large-scale 3D reconstruction but lack facade geometry. Existing methods rely on dense close-range images, incurring high costs and limited applicability. Addressing the challenge of facade geometry recovery for DSMs under sparse close-range images, we propose a DSM and Sparse-View 3DGS Co-Optimization Framework for Georeferenced 3D Building Reconstruction (DGS-CF). DGS-CF leverages DSMs to provide spatial priors that enhance Gaussian initialization quality. It augments the training samples through controllable view derivation, and employs a 3DGS that combines a depth-augmented view joint optimization with a dual‑Gaussian pruning strategy to achieve high fidelity recovery of the missing facade from extremely sparse-views. The completed 3D‑DSM preserves the georeferenced attribute of the original DSM, leading to georeferenced 3D building reconstruction. Experiments on public and self-collected data demonstrate that under only four views, our method outperforms the mainstream approaches with an average improvement of 21.73–27.91% in SSIM. Under the same initialization, the PSNR is increased at least 2.953 dB in the self-collected datasets with poorer DSM quality, while the restored 3D-DSM achieves improvements of at least 273.8% in quantity and 48.2% in accuracy. This research provides an effective technical pathway for 3D building reconstruction and facade geometry completion of DSMs from sparse-views.
Authors
- Y. Li (ORCID: https://orcid.org/0009-0005-8490-8431)
- Jiaxuan Song (ORCID: https://orcid.org/0009-0007-5180-5262)
- Dazhao Fan (ORCID: https://orcid.org/0000-0001-7100-8454)
- Ji Song
- Yang Dong
Institutions
- PLA Information Engineering University (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-17
- DOI
- https://doi.org/10.1038/s41598-026-80964-6
- Primary Topic
- 3D Surveying and Cultural Heritage
- Type
- article
- Field-Weighted Citation Impact
- 0.00