Abstract
Recovering 3D structure from multi-view satellite images is fundamental to urban planning, disaster assessment, and land monitoring. 3D Gaussian splatting (3DGS) offers an explicit geometric representation and real-time rendering, making it a promising alternative to stereo matching and neural radiance fields for satellite photogrammetry. Existing satellite 3DGS methods couple geometric optimization tightly with photometric signals. This coupling often leads to incomplete and overly smooth surfaces in vegetation and weak-texture regions. In this paper, we propose elevation-consistent 3D Gaussian splatting (ECGS), a method that breaks this coupling at the level of primitive creation and removal through three modules. First, we propose a multi-view elevation consistency regularization to establish cross-view geometric anchors, addressing elevation ambiguity without relying on photometric consistency. Second, a decoupled structure-aware densification mechanism is designed with an independent geometric pathway that keeps adding primitives where photometric gradients are unreliable. Finally, we introduce an importance-guided pruning strategy that removes the exposed redundancy through importance-weighted scores, while a geometric safety lock protects structural details. Experiments on the DFC2019 and IARPA 2016 benchmarks show that ECGS achieves better elevation estimates and higher-quality novel views. By jointly exploiting the radiometric appearance of multi-view imagery and the geometric evidence of cross-view correspondences, ECGS demonstrates that fusing these two modalities is an effective path toward intelligent remote sensing interpretation.