HS-USMR: stereo matching refinement with heterogeneous smoothness for urban satellite image 3D reconstruction

Accurate disparity estimation is crucial for urban 3D reconstruction from high-resolution satellite imagery as the quality of disparity maps directly affects the geometric fidelity of the reconstructed surface models. However, stereo matching on urban satellite images remains challenging, particularly in regions with occlusions, low texture, or disparity discontinuities. To address these challenges, we propose a general stereo matching refinement algorithm that integrates structural priors with heterogeneous smoothness. This method does not rely on large-scale training data, thereby overcoming the generalization limitations commonly encountered by deep learning approaches and exhibiting stronger applicability and robustness. Specifically, the method first models the initial disparity map using superpixel-driven segmentation and constructs a region-based graph prior in the form of a Markov random field to capture local and global spatial dependencies. A context-sensitive energy function is then applied to enforce consistency within regions and selective constraints across regions, enabling adaptive modeling of spatial relationships. During optimization, an iterative mechanism combining mini-batch gradient descent with edge-aware regularization dynamically adjusts edge weights, ensuring stable energy convergence while preserving geometric details. Experiments on two representative urban satellite image datasets, US3D and MVS3DM, demonstrate that our algorithm improves the quality of disparity maps and their corresponding digital surface model reconstructions. Compared with state-of-the-art stereo matching algorithms, our optimized disparity maps reduce the average endpoint error by 12.87%, D1 error by 18.92%, and root mean square error by 10.97%.