Inherited reference artefacts in Google’s GEM-15: orbit-geometry diagnosis, destriping and slope-adaptive fusion, validated with ICESat-2 and airborne lidar in drylands
Year: 2026
Type: preprint
Abstract: Google’s near-global 15 m optical-stereo elevation dataset GEM-15 has not been assessed independently in drylands. We evaluated GEM-15 and five global DEMs over Saudi Arabia with 313,466 ICESat-2 ATL08 v007 heights and in the south-western USA with 3DEP airborne lidar, and diagnosed their systematic errors from orbit geometry. GEM-15 was well registered horizontally (0.98 m) but carried tile offsets of −4.4 to +5.0 m and periodic stripes of about ±0.6 m. In Saudi Arabia, Egypt, the USA and Australia the stripes had the orientation (N12.5–13°E) and spacing (47–53 km) predicted by the 671-path reference grid of ALOS PRISM. AW3D30, also made from ALOS PRISM, carried the same stripes at the same positions (profile correlation 0.63–0.78), and its tile biases correlated with those of GEM-15 (r = 0.87). We attribute these errors to AW3D30, the producers’ prior for sensor-model refinement. An orbit-guided destriping against stripe-free Copernicus GLO-30 removed two thirds of the 35–70 km power and reduced the GEM-15 NMAD by 9–12% against lidar and ICESat-2. First-order estimates suggest that uncorrected tile seams (up to 4.8 m) and stripes could displace water lines by 0.3–1.1 km on flat plains. Against lidar, the GEM-15 DSM was more accurate than its DTM and than Copernicus GLO-30 above about 10°, while ICESat-2 underestimated steep terrain by 1.3–3.2 m. A slope-adaptive blend of corrected GEM-15 DSM and Copernicus GLO-30, calibrated with ICESat-2 only, reduced the RMSE by 21% in Saudi Arabia and, transferred unchanged, from 2.28 to 1.59 m (30%) against lidar in the USA.
DOI: 10.31223/x5dv5n
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236710951
