Multi-modal DEM super-resolution using relative depth: A new benchmark and beyond

Abstract

Learning-based Digital Elevation Model (DEM) super-resolution (SR) remains a challenge due to the complexity of real-world terrains. Existing approaches typically treat DEMs as digital grids or triangulated irregular networks, solving numerical fitting problems to densify points through learning models. However, these methods often overlook the spatial context and structural textures inherent in the terrain. To address this limitation, we propose utilizing relative depth maps derived from open-source remote sensing images by a foundational Depth Anything Model (DAM), which provide complementary structural information about the terrain and enhance the elevation details in DEMs. A novel DEMSR dataset, DEM-OPT-Depth SR, is constructed, pairing open-source remote sensing images, DEMs, and their corresponding relative depth maps. Additionally, we present a benchmark method, the Multi-modal Fusion Super-Resolution (MFSR) network, which extracts features through multi-branch pseudo-siamese networks and performs multi-scale feature fusion. Extensive experiments on the DEM-OPT-Depth SR dataset demonstrate a 24.63% improvement in RMSE-Elevation, a 22.05% improvement in RMSE-Slope, and an 11.44% improvement in RMSE-Aspect, showing the superiority and generalization capabilities of the MFSR model over previously proposed state-of-the-art baselines in DEMSR tasks. The code and dataset can be accessed at https://github.com/hwj0711/MFSR.