FathomDEM, a new global 30 m DTM

Описание: Accurate digital elevation models (DEMs) are foundational inputs for a vast array of geomorphometry
applications, including natural hazard modeling, glaciology, and infrastructure planning. However, existing
global DEMs, such as Copernicus DEM (COPDEM), often contain surface features like trees and buildings, limiting
their effectiveness as Digital Terrain Models (DTMs). This talk introduces FathomDEM, a new global 30 m DTM
created using a novel machine-learning methodology. We utilized a hybrid vision transformer model within a U-
Net architecture to perform pixel-wise regression, analyzing and correcting the height biases in COPDEM. This
approach differs significantly from previous methods (like the pixel-by-pixel correction used for FABDEM) by
leveraging 2D spatial information (context) as an inductive basis, essentially employing 'computer vision' to
achieve more spatially coherent and robust corrections. FathomDEM was trained on extensive, diverse LiDAR
reference data and has been rigorously validated, demonstrating: Improved Accuracy, surpassing the accuracy
of existing best-ranked global DEMs; Excellent Performance in Specific Landscapes, showing reduced error even
when compared to specialized coastal DEMs (e.g., DeltaDTM); High Utility in Downstream Tasks, when utilised
in flood inundation modeling, FathomDEM achieves increased accuracy, approaching the performance levels of
models derived from high-resolution LiDAR data. Join this session for an informal discussion on the methodology
behind FathomDEM, its novel use of ML for artifact removal, and its potential to improve applied
geomorphometry tasks globally.