Accuracy assessment of digital bare-earth model using ICESat-2 photons: analysis of the FABDEM

Digital elevation models play a crucial role in many earth sciences-related disciplines and are essential for spatial data infrastructure in a geospatial computational environment. A digital terrain model, a.k.a bare-earth model, excludes the bias due to the heights from urban and forests will enable the quantitative terrain characterization. Digital bare-earth models exemplify themselves as an essential input in numerous environmental analyses, especially in floodplain mapping. A forest and buildings removed Copernicus digital elevation model, referred to as FABDEM, is a first, global, open-access, machine learning-based, and 30 m spatial resolution data that simulates a bare-earth model. This article presents a novel approach to validate a digital bare-earth model by considering FABDEM as a case. Highly accurate elevations from the ground reflected photons from Ice, Cloud, and land Elevation Satellite-2 (ICESat-2) were used as a reference to check the agreement of FABDEM's tendency as a bare-earth model. Visual analytics done using the elevation profiles and error metrics confirms the FABDEM's successful reduction of bias due to urban and forests to a certain extent from its source DEM. However, the error metric shows a positive offset of similar to 3 m while validating the FABDEM's building removal algorithm, indicating a scope to decrease the building heights to achieve its anticipated objective of generating a bare-earth model in urban areas. In the case of the forest removal algorithm of FABDEM, it has successfully reduced the canopy heights to approximately 50% of its source. Still, the error metrics show a mean absolute error of similar to 14 m, similar to 10 m, and similar to 3 m in the test sites that fall in mountainous areas, rolling hills, and flat regions, respectively, that host different tree types and canopy structures. Our research has also investigated the possible sources of uncertainties and performance factors of FABDEM; these include the predictor variables, the number of regions of training sets, and errors that can accumulate from its original elevation source, i.e., Copernicus GLO-30 DEM.