Validation of global elevation models using ICESat-2 LiDAR data for floodplain modeling in the Ukrainian Carpathians

Abstract

Accurate topographic data underpin hydrological and floodplain modeling in mountainous environments where steep gradients and dense forest cover amplify vertical errors in global Digital Elevation Models (DEMs). This study performs a comprehensive validation of freely available DEMs — SRTM v3, NASADEM, ASTER GDEM v2, ALOS AW3D30, Copernicus GLO-30, FABDEM, and TanDEM-X — against high-precision Ice, Cloud, and land Elevation Satellite-2 (ICESat-2) LiDAR altimetry within the Ukrainian Carpathians. To ensure geodetic consistency, all DEMs and ICESat-2 observations were vertically transformed to the European Vertical Reference System (EVRS) using the high-resolution European Gravimetric Quasi-Geoid EGG2015 prior to analysis.

Elevation residuals were quantified using both classical (Mean Error, Root Mean Square Error) and robust (Normalized Median Absolute Deviation) statistical metrics, combined with terrain-stratified analysis based on slope, land cover, and hydrological position derived from the Height Above Nearest Drainage (HAND) model. The results demonstrate that DEM errors are strongly controlled by terrain steepness and vegetation cover, with non-linear error amplification observed in slopes exceeding 12° and in forested areas.

Among the tested datasets, FABDEM demonstrates the lowest mean error (≈1.5 m) and the highest stability across all slope classes. In contrast SRTM and NASADEM systematically overestimate elevations in forested terrain due to canopy effects. Copernicus GLO-30 and ALOS AW3D30 exhibit moderate accuracy but degraded performance beyond 15° slopes. ASTER GDEM displayed the largest variability and extreme errors, particularly in complex terrain.

Hydrological analysis revealed that DEM-related uncertainties propagate directly into floodplain modeling outputs. Within the critical HAND 0—6 m zone, vertical errors (5—10 m) were comparable to or exceeded typical flood depths, resulting in substantial discrepancies in inundation extent, channel geometry, and hydraulic parameters.

The study further demonstrates that compliance with international accuracy standards (INSPIRE, FEMA, LAWA) is generally limited to low-relief terrain, whereas most global DEMs fail to meet requirements in mountainous regions. These findings highlight the necessity of using DTM-type datasets or LiDAR-derived elevation models for regulatory flood-risk assessments.

To support reproducible and scalable analysis, the study introduces the GeoHydroAI framework — an integrated geospatial analytical environment combining ICESat-2 processing via SlideRule, DEM differencing using xDEM, terrain analysis with WhiteboxTools, and high-performance spatial querying with DuckDB.

This approach enables automated validation, terrain-stratified error analysis, and interactive exploration of DEM uncertainty across geomorphological and hydrological gradients. The proposed framework establishes a reproducible standard for DEM evaluation and provides a data-driven foundation for flood-risk assessment and hydrological modeling in data-scarce mountainous regions. Furthermore, the integration of geodetic referencing (EVRS/EGG2015), satellite altimetry (ICESat-2), and geomorphological analysis establishes a physically consistent framework for terrain representation in hydrological applications. This work positions DEM validation as a core component of GeoAI-driven environmental modeling, bridging geodesy, remote sensing, and hydraulic simulation within a unified analytical paradigm.