Bare-earth DEM Generation in Urban Areas Based on a Machine Learning Method

Precise representation of global terrain is of great significance for estimating global flood risk. However, current Global Digital Elevation Models (GDEMs) are all Digital Surface Models (DSMs) in urban areas. As a result, GDEMs such as the Shuttle Radar Topography Mission (SRTM) will block floods due to the higher elevation than ground level, caused by buildings' reflection of the incoming radar signal. Although corrections of SRTM have been examined in many research studies, these corrections have yet to be developed for urban areas despite the locations being of greatest exposure and suffering most from large GDEM errors. Even in the most comprehensive corrected version of SRTM - "Multi-Error-Removed Improved-Terrain DEM" (MERIT DEM), bias caused by buildings still exists. Research about the quality of MERIT DEM in urban areas is rare. This is also true of the newly available open access 1" resolution SRTM and 3" resolution TanDEM-X data products. In all cases methods to correct these DEM data to bare earth in urban areas are lacking. By taking GPS and LIDAR data as terrain observations, GDEMs (including SRTM, MERIT DEM and TanDEM-X 3" resolution DEM) errors were analysed in several cities. In five studied cities, it was found that the RMSE of SRTM, MERIT and TanDEM-X errors are in the range of 3.4 m - 5.9 m, and that MERIT and TanDEM-X both outperformed SRTM. But the error comparison between MERIT and TanDEM-X varied between the studied cities. Generally, mean error of TanDEM-X is slightly lower than MERIT but TanDEM-X has more extreme errors. For cities which have experienced rapid development in the past decade, mean error and RMSE of MERIT DEM error is lower than that of TanDEM-X DEM. In order to estimate MERIT DEM error, we adopted a machine learning method to train a model based on 70% samples from four cities (Beijing, Bristol, London and Manchester) and outputs were assessed on the other 30% samples of the four cities and in Berlin. 14 factors from widely available datasets including Night Time Light Data, MODIS vegetation index, world population density data, Openstreetmap building data, slope, elevation and its neighbourhood elevation values are used in the regression. Validation of corrected MERIT DEM showed good consistency with ground truth DEM, where the RMSE of test data dropped from 6.2 m to 1.7 m, and from 4.6 m to 2.7 m in a second validation in Berlin. The corrected DEM showed significant improvement on the original MERIT DEM. Although validation in more study sites is needed for extending application, a higher accuracy global scale bare-earth DEM is potentially available for flood estimation.