A large Scale Digital Elevation Model Super-Resolution Transformer

The Digital Elevation Model (DEM) super-resolution approach aims to improve the spatial resolution or detail of an existing DEM by applying techniques such as machine learning or spatial interpolation. Convolutional Neural Networks and Generative Adversarial Networks have exhibited remarkable capabilities in generating high-resolution DEMs from corresponding low-resolution inputs, significantly outperforming conventional spatial interpolation methods. Nevertheless, these current methodologies encounter substantial challenges when tasked with processing exceedingly high-resolution DEMs (256×256,512×512, or higher), specifically pertaining to the accurate restore maximum and minimum elevation values, the terrain features, and the edges of DEMs. Aiming to solve the problems of current super-resolution techniques that struggle to effectively restore topographic details and produce high-resolution DEMs that preserve coordinate information, this paper proposes an improved DEM super-resolution Transformer(DSRT) network for large-scale DEM super-resolution and account for geographic information continuity. We design a window attention module that is used to engage more elevation points in low-resolution DEMs, which can learn more terrain features from the input high-resolution DEMs. A GeoTransform module is designed to generate coordinates and projections for the DSRT network. We conduct an evaluation of the network utilizing DEMs of various types of terrains and elevation differences at resolutions of 64×64,256×256 and 512 × 512. The network demonstrated leading performance across all assessments in terms of root mean square error (RMSE) for elevation, slope, aspect, and curvature, indicating that Transformer-based deep learning networks are superior to CNNs and GANs in learning DEM features.