Generating digital elevation models (DEMs) using high-resolution satellite imagery is an efficient method for acquiring topographic information over vast areas. However, the accuracy of satellite-derived DEMs is affected by terrain-specific error mechanisms: steep topography amplifies geometric distortions, dense urban blocks create severe occlusion artifacts, vegetated nonurban land introduces canopy-related overestimation, and low-texture flat surfaces degrade stereo correspondence. Despite these well-known terrain dependences, in most refinement studies, only area-averaged accuracy metrics, which can obscure substantially different residual error structures across landscape types, are reported. To address this deficiency, in this study, we introduce a terrain-adaptive DEM refinement framework that integrates U-Net++ deep learning with low-rank matrix factorization and evaluates its effectiveness across four distinct terrain categories: mountainous, flat, urban, and nonurban areas. The framework operates through initial digital surface model reconstruction via U-Net, morphological filtering for nonground separation, low-rank decomposition for systematic bias elimination, and U-Net++ refinement for boundary and detail restoration. Validation experiments using Korea Multi-purpose Satellite-3 (KOMPSAT-3) stereo imagery over San Francisco, USA, and multiple Canadian sites reveal terrain-dependent improvement patterns. Urban areas exhibited the most dramatic gains, with root mean square error (RMSE) reductions exceeding 40–50% relative to three commercial photogrammetric software packages. Mountainous regions showed consistent improvement despite inherently higher baseline errors, while flat and nonurban terrains demonstrated stable sub-3 m RMSE convergence. Results of cross-site comparisons between the US and Canadian test areas suggest that the framework exhibits consistent terrain-specific improvement patterns under the conditions examined in this study.