A DEM Image Super-Resolution Reconstruction Method Based on the Texture Transfer of High-Resolution Remote Sensing Images

Traditional methods for acquiring high-resolution Digital Elevation Models (DEMs) are costly and laborious. Deep learning-based image super-resolution (SR) offers a promising alternative, but requires substantial training data. High-resolution DEMs, however, are often scarcer than satellite images at the same resolution. Recognizing the strong correlation between DEM grayscale images and high-resolution satellite imagery, we propose a novel method called EMASA-SR: Enhanced DEM Image super-resolution Reconstruction using Texture Transfer. It leverages texture information from satellite images to enhance the resolution of low-resolution DEMs. We address limitations of existing texture transfer methods by integrating a pyramid pooling module (PPM) and selective kernel convolution (SKC) into the network. PPM strengthens feature extraction for complex terrain objects, while SKC minimizes texture loss and feature confusion. Our experiments used 10 m Sentinel-2 remote sensing images and AW3D30 DEM data to upscale 30 m DEMs to 10 m resolution. Validation with ground-truth elevation data and ICESat-2 laser altimetry data revealed significant improvements. Compared to the original DEM, EMASA-SR achieved a 21.42%-37.44% reduction in elevation RMSE and a 23.30%-38.99% decrease in MAE. Moreover, it outperformed other super-resolution methods, achieving a 2.87%-28.27% reduction in RMSE and a 7.83%-30.04/% decrease in MAE.