Correcting SAR-derived DEMs with ICESat-2 using deep learning

Global digital elevation models (DEMs) generated from spaceborne synthetic aperture radar (SAR), such as the Copernicus 30 m DEM, provide exceptional coverage of the Earth's topography. However, SAR-derived DEMs struggle to accurately map terrain under forest canopies and in certain topographic conditions. In contrast, spaceborne laser altimeters like ICESat-2 can accurately measure ground elevations in areas where SAR sensors struggle, but the lack of dense coverage from laser altimetry precludes creation of complete global DEMs. This work aims to combine the accuracy of laser altimetry with the coverage of SAR by using deep learning algorithms. A convolutional neural network (CNN) is trained to correct Copernicus 30 m DEM using sparse but accurate ICESat-2 elevations in the south-east United States around South Carolina. Model inputs include temporally coincident imagery from Sentinel-2A, other SAR inputs from Sentinel-1B, as well as Copernicus 30 m DEM. The CNN is trained to correct the elevation of each individual pixel, allowing for the use of sparse ICESat-2 measurements. This allows the creation of a global DEM with the coverage of SAR and precision closer to that of laser altimetry. The resulting CNN model reduced ground elevation RMSE from 8.65 m to 2.62 m. The corrected DEM has potential to benefit numerous scientific endeavors requiring accurate global topographic information.