Title: IMPROVING ICESAT-2 ATL08 LAND ELEVATION ACCURACY IN COMPLEX HILLY TERRAIN USING MACHINE LEARNING MODELS AND SPATIAL ANALYSIS
|
Authors: Zhang Shengpeng, Zhang Yongming, Zhang Yongying, Hao Guangjing and Li Hanmei
|
DOI: 10.13168/AGG.2026.0015
|
Journal: Acta Geodynamica et Geomaterialia, Vol. 23, No. 2 (222), Prague 2026
|
Full Text:  PDF file (1.7 MB)
|
Keywords: ICESat-2 ATL08; GF-7; DEM; Geodetector; Spatial validation; Machine learning
|
| Abstract: Accurate evaluation of ICESat-2 ATL08 terrain elevations in rugged mountainous areas is important for reliable geodetic applications. In this study, ATL08 Version 006 100 m terrain segments in the upper Yellow River Basin were evaluated against a 2 m GF-7 stereo DEM after converting ATL08 ellipsoidal heights to orthometric heights to ensure vertical-datum consistency. After quality screening, 15,904 segments were retained, and the mean absolute elevation difference was 14.08 m. Geodetector analysis identified terrain_slope and h_te_std as the dominant explanatory factors, whereas snr, n_ca_photons, h_te_uncertainty, and h_te_interp_minus_median played secondary but still significant roles. To reduce spatial leakage, the retained segments were partitioned into spatially disjoint blocks of approximately 10 km and systematically assigned to five folds. Six machine-learning models were then compared for predicting absolute elevation error under spatially blocked cross-validation. HistGradientBoosting achieved the best overall performance, with a mean test R² of 0.452 ± 0.016, RMSE of 10.65 ± 0.64 m, and MAE of 7.24 ± 0.40 m across the five spatial folds. The results indicate moderate but stable predictive skill in geographically unseen areas and support terrain-aware screening of ATL08 segments for DEM validation and related geodetic applications. |
Archived website - The current official IRSM website can be found here - Archived website 
Intranet
Exchange Mail
Fileserver
Owncloud
About Institute
Acta Geodynamica et Geomaterialia 
