Digital elevation models (DEMs) are essential for applications such as hydrologic analysis and urban planning, and the demand for satellite image.based DEMs is growing with the spread of high-resolution imagery. However, conventional satellite-based DEM generation often suffers from local noise and global structural distortion in complex terrain and urban areas. This study proposes a hybrid DEM refinement method that combines traditional morphological filtering, deep learning.based U-Net/U-Net++, and low-rank matrix decomposition. Using KOMPSAT-3 imagery over San Francisco (USA) and Toronto (Canada), DEMs generated by three commercial software packages were compared with DEMs refined by the proposed method for mountainous and flat terrain, based on RMSE, MAE, and bias against a reference DEM. The proposed approach reduced RMSE and MAE by approximately 50.70% and constrained bias to within about ±1 m, with especially large improvements in mountainous areas, demonstrating the effectiveness of combining low-rank decomposition and U-Net++ for satellite image.based DEM refinement.



Keywords : DEM, KOMPSAT-3, Deep Learning, Low-Rank Matrix Decomposition, Morphological Filtering