An enhanced global terrain map using a vision transformer machine learning model

Digital Elevation Models (DEMs) describe the earth surface’s topography and are an important source of information for applications of physical modelling, engineering and many others. Flood inundation modelling, where water flows are determined by terrain slope, is also highly dependent on DEM quality. The most accurate DEMs currently available are sourced from airborne LiDAR, however these only cover a small fraction of the globe, leaving the majority of the globe sourced from satellite imagery. Satellite based DEMs have limitations and are considered Digital Surface Models (DSMs) which represent the surface of vegetation canopy, buildings and other objects, rather than the bare earth surface which is represented by a Digital Terrain Model (DTM). 

Due to this, we have developed FathomDEM, a DTM generated from the best global satellite based DSM, Copernicus DEM. FathomDEM uses a novel vision transformer technique to improve on previous attempts to generate a DTM from Copernicus DEM.  FathomDEM reduces the Mean Absolute Error and Root Mean Squared Error to half of our previous work, FABDEM, and quarter of Copernicus DEM, while also improving the spatial correlation. 

Flood simulations of inundation using a given DEM shows its use in a real world application and we present results showing flood inundation maps from different global DEMs and LiDAR. FathomDEM gives similar scores to LiDAR data when compared to benchmark flood extents, tested across multiple sites. FathomDEM therefore provides a significant advance when applied to flood inundation modelling in locations without LiDAR DEMs.