A Scalable Slope-Dependent Vertical Bias Correction for Global Digital Elevation Models Using ICESat-2

Advancements in Earth observation satellites integrating stereo, multispectral, and radar imaging have enabled the three-dimensional mapping of landscapes at unprecedented resolutions, driving the creation of global Digital Elevation Models (DEMs) such as ALOS AW3D30 and Cartosat-1. Despite their utility, these products suffer from systematic, terrain-induced vertical biases that limit local hydrological and environmental applications. While recent literature proposes complex machine learning or multi-source data fusion frameworks to correct these errors, such approaches introduce heavy computational overhead, parameters that are difficult to interpret, and severe dependencies on auxiliary, often outdated global land cover products. To address these bottlenecks, this study presents a simple, highly scalable, and self-contained slope-dependent continuous linear regression framework using spaceborne laser altimetry from the ICESat-2 ATL08 product.

By utilizing only the continuous slope map derived directly from the target DEM, the proposed method isolates and corrects systematic vertical bias without requiring external datasets or high-compute iterations. To ensure rigorous, unbiased evaluation, we implement a spatial block-splitting strategy (50/50 training/validation split) to evaluate correction offsets on geographically distinct subsets of terrain. The framework was validated against high-resolution airborne LiDAR over New Zealand (Wellington and Waikato test sites) and across 29 geographically diverse Areas of Interest (AOIs) in India representing flat, moderate, and highly rugged topographies.

The results demonstrate that the correction framework successfully shifts the elevation error distributions to center precisely near zero. Across the 29 application sites, the overall category-wide systematic mean bias was reduced by 97.7% (shifting from 4.15 m to 0.10 m), with the overall Root Mean Square Error (RMSE) improving by 11.4% (11.74 m to 10.40 m) and the Mean Absolute Error (MAE) decreasing by 13.8% (8.96 m to 7.72 m). Category-specific analysis revealed mean bias reductions of 98.0% in flat terrain (1.73 m to 0.04 m), 99.2% in moderate terrain (6.58 m to 0.05 m), and 94.8% in highly rugged terrain (4.16 m to 0.22 m). By achieving substantial vertical accuracy improvements while maintaining a computationally light, single-parameter model, this framework offers a highly practical, globally scalable baseline solution for rapid large-scale DEM pre-processing and calibration.