Unmixing frequency features for DEM super resolution

DEM super-resolution (SR) has recently been advanced by deep learning. The focus of existing works is mainly on the employment of various terrain constraints to force the general deep SR models to adapt to DEM data. However, we found that they leave a fundamental issue of terrain pattern confusion caused by the mixed frequency feature learning of deep neural networks, which leads to an inherent trade-off between the reconstruction of fundamental structures and the preservation of fine-grained terrain details. In this study, we propose a novel dual-frequency feature learning network (DuffNet) for high quality DEM super-resolution. The core idea of DuffNet is to directly learn the mapping relationship between low-resolution (LR) and high-resolution (HR) DEMs with meaningful frequency features, rather than the mixed convolutional features extracted from raw DEMs. Specifically, DuffNet deploys a dual-branch structure with a dedicatedly designed dual-frequency loss to enable the learning of high- and low-frequency features under the supervision of input HR DEM. An adaptive elevation amplitude refiner (AEAR) is then developed to dynamically adjust and optimize the amplitudes of the initial HR DEM synthesized by the integration of learned low-frequency and high-frequency terrain components. Extensive experiments conducted on TFASR30, Pyrenees, Tyrol, and the challenging TFASR30to10 datasets show that DuffNet can achieve state-of-the-art performance, outperforming other SoTA methods such as TTSR and CDEM by 19% and 29% respectively in RMSE-Elevation on the TFASR30to10 dataset. The dataset and source code are available at: https://github.com/Geo-Tell/DuffNet.