Abstract
The Earth’s terrain is linked to many physical processes, and gaining the most accurate representation is key to work in many sectors from engineering to natural hazards modeling and ecology. Existing global digital elevation models (DEMs) are widely used, however often suffer from systematic biases caused by trees, buildings and instrumentation error, ultimately limiting their effectiveness. We present here, FathomDEM, a new global 30 m DEM produced using a novel application of a hybrid vision transformer model. This model removes surface artifacts from a global radar DEM, Copernicus DEM, aligning it more closely with true topography. In addition to improving on other global DEMs, FathomDEM also has reduced error compared to coastal-focussed DEMs such as the recent DeltaDTM. This demonstrates its impressive capacity to perform for specific landscapes, while being trained globally to model a wide range of terrain types. FathomDEM has been tested on the downstream task of flood modeling, showing increased accuracy compared to those run with the previous best global DEM, FABDEM, approaching the performance of LiDAR based flood modeling. This improvement is attributed to FathomDEM’s smaller error and substantial reduction in artifacts. This shows the suitability of FathomDEM for applied tasks and strengthens our evaluation compared to one based on vertical error alone.
Original language | English |
---|---|
Article number | 034002 |
Number of pages | 13 |
Journal | Environmental Research Letters |
Volume | 20 |
Issue number | 3 |
Early online date | 11 Feb 2025 |
DOIs | |
Publication status | Published - 1 Mar 2025 |
Bibliographical note
Publisher Copyright:© 2025 The Author(s). Published by IOP Publishing Ltd.
Keywords
- remote sensing
- terrain
- vision transformer
- machine learning
- digital elevation model