Laser-altimetry-enhanced foundation model embeddings for global elevations (Invited Paper)

Foundation models have demonstrated substantial performance gains across domains. In Earth science, Google’s AlphaEarth Foundation (AEF) model has outperformed conventional spaceborne imagery in many applications. Pre-trained on diverse datasets—including Landsat, Sentinel-2, Wikipedia vector layers, GEDI canopy heights, and Copernicus DEM—AEF shows strong capability in 2D classification tasks. However, its utility for 3D applications, particularly integration with airborne laser scanning (ALS) or spaceborne altimetry, remains less explored. This study evaluates AlphaEarth embeddings for predicting terrain elevation and canopy height across three ecoregions. Using an XGBoost regressor, models were trained with ALS or ICESat-2 as reference data. Results indicate that AlphaEarth features alone are insufficient for accurate terrain prediction without additional inputs, despite elevation-related pretraining. In contrast, canopy height predictions show strong standalone performance. Comparisons with multispectral approaches highlight the promise of foundation models, while underscoring the continued need for high-fidelity laser data to calibrate and validate these methods.