Learning with image guidance for digital elevation model super-resolution

Acquiring higher-resolution Digital Elevation Model (DEM) data is important for many land-surface applications. Creating high-resolution DEMs from freely available low-resolution satellite-derived DEMs, along with supplementary data about the land surface, is a possible solution that may provide the data needed, particularly in areas without high-resolution DEMs. This work investigates sparse-to-dense depth completion approaches for DEM super-resolution. It proposes a neural network based on joint image filtering and spatial propagation networks, reconstructing high-resolution DEMs deriving from low-resolution DEMs and guidance images. Experiments demonstrate that our results are improved by 56%/38% in accuracy and 9%/4% in reconstruction quality compared to Bicubic/EDSR methods.