Terrain Variation Aware Surface Height Estimation Network for High-Precision Multi-view Satellite Stereo

Deep learning–based multi-view stereo (MVS) has shown great potential for large-scale 3D reconstruction from satellite imagery. However, existing methods struggle with the large depth ranges and complex scenes inherent to this domain. By relying on rigid range-guided sampling and local consistency assumptions, these approaches often neglect specific terrain variations. This leads to misallocated hypothesis budgets and excessive smoothing across discontinuities in complex terrain. To address these limitations, we propose a terrain variation–aware multi-view satellite imagery surface height estimation network that introduces surface normal information to construct an end-to-end terrain-guided height estimation framework. To effectively capture terrain variation, the network computes normal maps from height maps and measures local terrain variation within the neighborhood of reference pixels based on height and normal information. The network further incorporates a local terrain-guided height hypothesis module and a terrain variation–aware feature enhancement module that optimize height hypothesis sampling precision and enhance the terrain-aware features in regions with significant terrain variation, respectively. Additionally, we propose a joint height-normal loss to implicitly optimize height estimation. Extensive experiments conducted on the WHU-TLC and MVS3D data sets demonstrate that our method achieves state-of-the-art performance across all evaluation metrics. Notably, in urban areas and mixed scenes, the proposed approach achieves a significant improvement in the proportion of high-precision pixels compared to existing methods, with the maximum improvement reaching 30.25%.