Abstract:

[Objective] Digital Elevation Models (DEMs) are a critical data foundation for research in the Earth sciences. However, existing super-resolution (SR) methods often face a common challenge in reconciling global topographic structures with local fine-grained textures, failing to adequately preserve both macroscopic terrain trends and fine-scale details. [Methods] To address this limitation, this paper proposes a Topography-aware Hybrid Attention Network (THA-Net). This model introduces an innovative Topography-aware Hybrid Attention Block (THAB), which concurrently integrates a Swin Transformer-based global branch with a convolution-based local branch to enable the synergistic and efficient capture of multi-scale topographic features. [Results] Experimental results demonstrate that THA-Net achieves superior performance in evaluations on complex regions, such as the Himalayas. Compared to traditional Bicubic interpolation and advanced methods like EDSR and SwinIR, THA-Net reduces the Root Mean Square Error (RMSE) by 1.9% to 77.5% and the Mean Absolute Error (MAE) by 1.5% to 82.6%. Furthermore, the reconstructed terrain details and hydrological networks from our model exhibit the highest fidelity to the ground-truth topography, offering a valuable tool for complex terrain analysis and disaster simulation. [Conclusion] We conclude that this collaborative approach, which synergizes global and local feature learning, enhances the model’s topographical awareness and leads to superior reconstruction quality.

Key words: Digital Elevation Models, super-resolution, deep learning, Transformer, Convolutional Neural Network