DEM Super-Resolution Using Frequency-Selective Cross Attention Assisted by Remote Sensing Images

Digital Elevation Model (DEM) plays a crucial role in geospatial analysis. However, high-resolution DEM is often costly to acquire and has limited spatial coverage. This letter proposes a DEM super-resolution method that integrates auxiliary information from remote sensing image (RSI): dual-branch bidirectional feature guidance network (DBFGN). The proposed method adopts the Swin Transformer as the backbone and incorporates a channel-spatial joint Transformer (CSJT) module to enhance both global and local feature extraction capabilities. To effectively extract consistent texture from RSI that strongly correlates with elevation variations, we design a mutual-guided network (MGN) based on frequency-selective cross-attention (FSCA), enabling adaptive extraction and fusion of high-frequency structure. Experiments conducted on the Swiss and Vaihingen datasets demonstrate that the proposed method significantly outperforms existing DEM and general image super-resolution methods in the challenging 8× upscaling task of DEM super-resolution, achieving RMSE and MAE of 3.108 m and 1.818 m, respectively. Compared to the second-best method, our approach improves accuracy by 8.37%, validating its effectiveness in complex height reconstruction tasks.