Despite significant advancements in satellite stereo matching, existing deep learning models often struggle with practical challenges in generating accurate Digital Surface Models (DSMs), particularly facing issues of high computational costs and performance limitations in complex scenarios. This paper presents TriGeoNet, an efficient end-to-end architecture that addresses these limitations through three key innovations: (1) a Triple-Feature-Split cost volume strategy that simultaneously models similarity, difference, and complementarity relationships between features, enabling robust matching in weakly textured regions; (2) a geometry-statistics collaborative optimization framework that combines entropy-enhanced anisotropic geometric perception with cross-scale dynamic attention fusion, strengthening geometric constraints while reducing computational complexity; and (3) a gradient-guided uncertainty-aware edge refinement module that bridges low-level image gradients with high-level disparity confidence maps for precise boundary preservation. Extensive experiments demonstrate state-of-the-art performance with an endpoint error (EPE) of 0.983 pixels (D1-error: 4.92%) on the US3D dataset and 1.511 pixels (D1-error: 11.18%) on the WHU-Stereo dataset. In terrain reconstruction tasks, TriGeoNet generates DSMs with average RMSE and MAE improvements of 24.6% and 28.9% respectively across all test regions compared to other state-of-the-art methods. The proposed framework offers new insights for developing modern satellite stereo vision systems. Code available at https://github.com/CVEO/TriGeoNet . • Triple-Feature-Split cost volume captures multi-information for robust matching. • Entropy-enhanced geometric optimization balances directional constraints adaptively. • Cross-Scale Dynamic Fusion integrates multi-resolution features. • Gradient-guided uncertainty-aware refinement preserves sharp building boundaries.
Chen et al. (2026) studied this question.