Metric 3D Reconstruction from Satellite Imagery via Geometry-Constrained Gaussian Splatting

Digital surface model (DSM) generated from high-resolution satellite imagery plays an important role in various remote sensing applications. However, reconstructing accurate DSM from such data remains challenging due to sparse viewpoints and the extremely low base-to-height ratio. To address these issues, we propose a DSM reconstruction method based on 3D Gaussian Splatting (3DGS) with geometric constraints. An adaptive Rational Polynomial Coefficients (RPC)-based camera optimization module enables differentiable refinement of satellite imaging geometry, enhancing multi-view consistency under sparse viewpoints. Meanwhile, a Depth-guided Adaptive Supervision strategy focuses on optimizing scene geometry, improving the structural completeness and accuracy of DSM. Experiments on the DFC2019 dataset demonstrate that the proposed method outperforms existing state-of-the-art approaches. This framework provides an efficient and robust solution for high-precision 3D reconstruction from satellite imagery.