Improving digital elevation models from satellite images with layered height filtering
HLC-GS: Risk-Map-Guided Height-Layer Consistency Gaussian Splatting for DSM Reconstruction from Optical Satellite Imagery
Computer Vision and Pattern Recognition
Summary
Digital Surface Models (DSMs) map the Earth's surface heights and help in various mapping tasks. The authors found that current 3D modeling techniques sometimes mix up different elevation layers, causing errors in the height data. They created HLC-GS, a new method that detects risky pixels with unreliable height data and fixes these by keeping height layers consistent. This method produces more accurate and realistic elevation maps from satellite pictures than previous methods, according to tests on public datasets.
What this means in practice
- •For geospatial engineers: Generate more accurate elevation models from optical satellite images for mapping and analysis.
- •For urban planners: Use improved DSMs with fewer height errors for city planning and infrastructure projects.
Authors
Jie Yang, Yingdong Pi, Qiyan Luo, Xiaoyu Wang, Lekang Wen, Mi Wang
Abstract
A Digital Surface Model (DSM) is a fundamental geospatial data product for representing the elevation of the Earth's surface. Recently, 3D Gaussian Splatting (3DGS) has shown considerable potential for DSM reconstruction from multi-view optical satellite imagery due to its explicit scene representation and efficient optimization. However, in 3DGS-based DSM generation, alpha-weighted aggregation of Gaussian altitudes may blend splats from different height layers at the same rendered pixel or DSM sampling location, producing non-physical intermediate elevations and height-layer mixing errors. To address this problem, we propose HLC-GS, a risk-map-guided height-layer consistency Gaussian Splatting method for DSM reconstruction from optical satellite imagery. HLC-GS consists of a risk map module, a dominant-layer reliability correction module, and a secondary-layer suppression module. The risk map localizes high-risk pixels with abnormal height dispersion and unreliable dominant-layer responses, while the latter two modules regularize unreliable dominant-layer responses and suppress weakly supported far secondary-layer responses. Extensive experiments are conducted on the DFC2019 and IARPA2016 datasets. Compared with six state-of-the-art DSM reconstruction methods, HLC-GS achieves better overall accuracy. Compared with the latest and precision-enhanced EOGS, HLC-GS reduces the average MAE from 1.46 m to 1.18 m and the average RMSE from 2.78 m to 2.58 m over the evaluated scenes, while improving PAG$_{2.5}$ from 86.09\% to 88.61\%. Overall, these results demonstrate that explicitly modeling per-pixel height-layer consistency alleviates height-layer mixing and improves the geometric quality of 3DGS-based DSM reconstruction from optical satellite imagery.