Generating high-resolution DEMs in mountainous regions using ICESat-2/ATLAS photons

• A spatial interpolation knowledge-constrained deep learning model (SikCGAN) is built. • SikCGAN assimilates sparse ICESat-2/ATLAS photons for effective fine-res DEM mapping. • SikCGAN has demonstrated effectiveness compared to benchmark approaches. • SikCGAN is particularly effective in mountain landscapes & low photon density regions. High-resolution (≤10 m) digital elevation models (DEMs) are essential for obtaining accurate terrain information and are integral to geographic analysis. However, a majority of currently available DEMs datasets possess a relatively coarse spatial resolution (≥30 m), which limits the terrain features and details that can be accurately represented. Furthermore, due to the substantial production costs associated with high-resolution DEMs, these products are often unavailable or difficult to obtain in numerous countries and regions, particularly in less developed areas. Here, we introduced a novel method named the Spatial interpolation knowledge-constrained Conditional Generative Adversarial Network (SikCGAN). This method can generate high-resolution DEMs from publicly available data sources, specifically the photons collected by the Advanced Topographic Laser Altimeter System (ATLAS) carried by the Ice, Cloud and land Elevation Satellite-2 (ICESat-2). SikCGAN takes ICESat-2/ATLAS photons as the single data source and incorporates spatial interpolation knowledge constraints into a Conditional Generative Adversarial Network (CGAN) to generate DEMs at a 10-m spatial resolution. A case study conducted in boreal mountainous regions demonstrates SikCGAN’s remarkable ability to produce high-resolution and highly accurate DEMs, with an MAE of 22.09 m and RMSE of 29.25 m, which reduced error by 37 %–46 % compared to benchmark methods. Additionally, the results reveal that SikCGAN has remarkable resiliece to interference, including variations in spatial distance, terrain slope, and ATL03 photon count, this further elucidates and substantiates the effectiveness of SikCGAN. These findings demonstrate that SikCGAN provides innovative solutions for generating new high-resolution DEMs products and potentially supplementing existing ones to overcome their limitations.