SkySplat++: Hierarchical 3D Gaussian Splatting for generalizable sparse-view reconstruction from multi-temporal satellite imagery

Scene reconstruction from sparse-view satellite images remains a challenging task due to insufficient geometric constraints. While generalizable 3D Gaussian Splatting (3DGS) methods have shown high efficiency for novel-view synthesis, existing methods are often unsuitable for satellite imagery due to incompatibility with the rational polynomial coefficient (RPC) model and limited generalization capability. In this work, we present SkySplat++, a hierarchical generalizable 3DGS framework that extends our previous SkySplat to further improve scene reconstruction. SkySplat++ refines Gaussians in a coarse-to-fine manner, enabling robust modeling of both large-scale scene structures and fine-grained texture details. To improve reconstruction under sparse-view conditions, a Height-Guided Hierarchical Cost Volume (HG-HCV) module is integrated to enable more accurate cost volume construction. Additionally, a hybrid height supervision strategy is employed to stabilize height estimation under radiometric inconsistencies. Compared to per-scene optimization methods, SkySplat++ achieves up to 36 times speed up over EOGS through feed-forward inference after training, while achieving higher accuracy. It also outperforms generalizable 3DGS baselines, reducing mean absolute error by 10.38–14.81 m relative to HiSplat across two benchmarks, and demonstrates strong cross-dataset generalization on the MVS3D test set. The code will be released at https://nancheng2001.github.io/SkySplatV2/ .