Adaptive smooth constrained stereo matching and disparity refinement for satellite DSM generation

High-resolution satellite images often exhibit substantial geometric and radiometric differences, making it difficult for conventional stereo matching methods to achieve high matching accuracy while suppressing noise and preserving object boundaries. To address this issue, this study improves both stereo matching and disparity post-processing in the satellite digital surface model (DSM) generation workflow. First, within the semi-global matching (SGM) framework, a mixed matching cost is constructed by combining the Census transform and gradient descriptors. Image gradients and edge cues extracted by a pretrained PiDiNet are then incorporated to adaptively control the strength of the smoothness constraint during cost aggregation, thereby improving matching reliability while balancing noise suppression and disparity-boundary preservation. Second, we propose a disparity post-processing method that combines improved weighted least-squares disparity refinement with matching-cost-based disparity fusion. The method first dynamically adjusts the smoothing weights to reduce texture-copying artifacts and edge blurring and then fuses disparity using matching costs to alleviate staircase artifacts caused by excessive smoothing. Experiments were conducted on ZY3-02 and WorldView-3 satellite imagery, with SGM, S2P-HD, Agisoft Metashape software, guided aggregation net, and FoundationStereo included for comparison. Across three test scenes dominated by urban and agricultural areas, the proposed method increased the proportion of pixels with absolute elevation errors below 2 m by an average of 4.12 percentage points and reduced RMSE by an average of 0.11 m compared with standard SGM. Qualitative results further show that the proposed method effectively suppresses noise in relatively flat regions while preserving building boundaries.