Traditional methods for acquiring high-resolution digital elevation models (DEMs) are costly and laborious.Deep-learning-based image superresolution (SR) offers a promising alternative but requires Membrane Switch 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 SR reconstruction using texture transfer.
It leverages texture information from satellite images to enhance the resolution of low-resolution DEMs.We address the 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 Shipping Crates 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 SR methods, achieving a 2.87%–28.
27% reduction in RMSE and a 7.83%–30.04/% decrease in MAE.