NepalDTM v2: Improving on NepalDTM v1 with ICESat-2-Adjusted Stereo Strips for a National 8 m Terrain Model of Nepal
Year: 2026
Type: preprint
Abstract: NepalDTM v1 corrected the NASA High Mountain Asia (HMA) 8 m stereo mosaic to bare earth with a gradient-boosting model trained on satellite laser altimetry. Thirty-one recorded experiments showed that its error is set by the input surface, not by the model: added predictors moved the national error by at most 2%. NepalDTM v2 therefore rebuilds the input. All 512 HMA stereo strips over Nepal are adjusted jointly, with a plane per strip, against 18.5 million ICESat-2 ATL03 ground photons and 13.3 million strip-to-strip ties; and stereo cells that v1 discarded at cliffs, where Copernicus itself is wrong, are kept wherever two adjusted strips agree. On 19.8 million held-out ATL03 ground photons, from tracks never used in the adjustment, the v2 terrain model scores 5.26 m RMSE (0.85 m NMAD), against 6.93 m for v1, 8.69 m for Copernicus GLO-30, 8.40 m for FABDEM and 8.97 m for GEDTM30, and it is the most accurate of the five in every land-cover, slope and elevation class. Against an independent 2 m WorldView/GeoEye composite at Langtang the v2 surface error falls from 8.62 to 4.71 m. We also report what did not help. Data: https://doi.org/10.5281/zenodo.24327134; code: https://github.com/achyut/NepalDTM.
DOI: 10.31223/x5b23b
Author: Achyut Paudel
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236127396
