关键词: 数字高程模型, 超分辨率重建, 地形因子, 先验引导, 特征融合

Abstract:

To overcome the limitations of existing digital elevation model (DEM) super-resolution reconstruction methods in model generalization and terrain feature awareness, a terrain-factor-guided approach is proposed. First, a paired high- and low-resolution DEM dataset covering global coastal regions is constructed based on the open-source DEM data from GEBCO and ETOPO-1. Then, an end-to-end DEM super-resolution model is designed with a high-low frequency feature decoupling module to separate and fuse terrain details and global features. A terrain guidance module is introduced, incorporating prior terrain knowledge, such as surface segmentation depth, to enhance the perception of terrain structure. Additionally, dynamic weight adjustment based on homoscedastic uncertainty is employed to adaptively assign loss weights according to the importance of different terrain features, thereby improving reconstruction accuracy and stability. Comparative experiments with seven state-of-the-art methods demonstrate that the proposed approach achieves superior performance in RMSE_elevation, MAE_elevation, and RMSE_slope metrics, accurately reconstructing complex features such as fragmented ridges and low seamounts. These results validate the effectiveness and robustness of the method in terrain feature reconstruction and precision enhancement.

Key words: digital elevation model, super-resolution reconstruction, terrain factors, prior-guided, feature fusion