GeoGS: Geospatial Gaussian Splatting for Robust 3D Reconstruction from Sparse Satellite Imagery

Han-Gyeol Kim, Seonghyeon Yun, JaeWan Park, Darongsae Kwon; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2026, pp. 7980-7989

Sparse-view 3D reconstruction from satellite imagery remains a significant challenge due to data acquisition costs and the inherent geometric instability of 3D Gaussian Splatting (3DGS) in satellite environments. We propose Geospatial Gaussian Splatting(GeoGS), a robust framework that achieves high-fidelity reconstruction from as few as 3 to 7 images without reliance on external Digital Surface Models (DSMs). To prevent structural collapse, GeoGS integrates foundation model-based depth priors from Depth Anything V2 as a geometric anchor. Furthermore, a shadow-aware appearance module, utilizing DINOv3 semantic features and physics-based hillshade, effectively disentangles static terrain from dynamic illumination transients. Coupled with a three-phase curriculum learning strategy, GeoGS ensures stable radiometric consistency and geometric convergence. Experiments on DFC2019 and IARPA datasets demonstrate that GeoGS significantly outperforms state-of-the-art models, maintaining superior structural integrity even at N=3. Cross-sensor validation on SkySat imagery of Seville and Jamsil further proves the model's exceptional generalizability for real-world geospatial applications.