A Geometry–Appearance Collaborative 3DGS-SDF Framework for Metric 3D Reconstruction from Satellite Imagery

High-resolution satellite imagery is a key data source for large-scale 3D scene reconstruction. However, existing neural rendering methods remain limited in sparse-view satellite scenes due to complex imaging geometry, insufficient geometric constraints, radiometric inconsistency, and unstable structural boundaries. To address these challenges, we propose a geometry–appearance collaborative framework that combines 3D Gaussian Splatting (3DGS) with a Signed Distance Function (SDF). The 3DGS branch enables efficient explicit rendering and fine-grained appearance recovery, while the SDF branch models continuous surfaces to improve structural completeness and reduce floating artifacts.

Based on this hybrid representation, we introduce three modules tailored to satellite remote sensing scenes. First, an adaptive camera optimization module converts RPC models into differentiable affine camera initializations and optimizes rotation and translation increments during training, thereby reducing projection errors and improving multi-view geometric consistency. Second, a height supervision module aligns relative depth priors with rendered elevations and imposes adaptive depth-to-height constraints to strengthen metric surface geometry. Third, a multi-scale edge-gradient loss jointly compares Sobel gradient vectors and gradient magnitudes at progressively downsampled scales, together with scale-decayed weights and optional edge masks, to improve structural boundary stability.

Experiments are conducted on seven satellite scenes from the Data Fusion Contest 2019 (DFC2019) and IARPA MVS 3-D Mapping Challenge (IARPA2016) datasets, with EOGS adopted as the baseline method. Using Mean Absolute Error (MAE) as the metric for geometric reconstruction accuracy, our method achieves average MAEs of 1.4517 and 1.8510 on DFC2019 and IARPA2016, respectively, corresponding to reductions of 3.78% and 2.39% over the EOGS baseline. These results demonstrate that the proposed geometry–appearance collaborative framework consistently improves metric 3D reconstruction accuracy over the EOGS baseline in satellite imagery.