Remote Sensing, Vol. 17, Pages 3625: A Transformer-Based Residual Attention Network Combining SAR and Terrain Features for DEM Super-Resolution Reconstruction

Remote Sensing doi: 10.3390/rs17213625

Authors:
Ruoxuan Chen
Yumin Chen
Tengfei Zhang
Fei Zeng
Zhanghui Li

Acquiring high-resolution digital elevation models (DEMs) over across extensive regions remains challenging due to high costs and insufficient detail, creating demand for super-resolution (SR) techniques. However, existing DEM SR methods still rely on limited data sources and often neglect essential terrain features. To address the issues, SAR data complements existing sources with its all-weather capability and strong penetration, and a Transformer-based Residual Attention Network combining SAR and Terrain Features (TRAN-ST) is proposed. The network incorporates intensity and coherence as SAR features to restore the details of the high-resolution DEMs, while slope and aspect constraints in the loss function enhance terrain consistency. Additionally, it combines the lightweight Transformer module with the residual feature aggregation module, which enhances the global perception capability while aggregating local residual features, thereby improving the reconstruction accuracy and training efficiency. Experiments were conducted on two DEMs in San Diego, USA, and the results show that compared with methods such as the bicubic, SRCNN, EDSR, RFAN, HNCT methods, the model reduces the mean absolute error (MAE) by 2–30%, the root mean square error (RMSE) by 1–31%, and the MAE of the slope by 2–13%, and it reduces the number of parameters effectively, which proves that TRAN-ST outperforms current typical methods.

Source link

Ruoxuan Chen www.mdpi.com