Alpha Foundation Satellite Embeddings and GEDI Integration for Mapping Forest Aboveground Biomass in Alpine Areas via Google Earth Engine: An Application in Valle d'Aosta, NW Italy

Quantifying forest aboveground biomass (AGB) is critical for regional carbon accounting and climate change mitigation. While traditional remote sensing approaches rely heavily on handcrafted spectral indices from optical sensors, vision foundation models offer self-supervised, multi-dimensional latent representations that remain largely unexplored for quantitative biophysical regression in complex mountain ecosystems.

This study evaluates an advanced machine learning pipeline developed within Google Earth Engine (GEE) using the 64-dimensional Alpha Foundation Satellite Embeddings (2024 annual composite) to map forest AGB density. The model was calibrated using spaceborne LiDAR footprints from the Global Ecosystem Dynamics Investigation (GEDI L4A) dataset as the target variable. GEDI measurements were filtered to exclude high-uncertainty data (relative standard error > 50%) and complex terrain with slopes steeper than 30% derived from the Copernicus GLO-30 DEM. Predictors and target variables were co-registered and resampled to a 100-meter analysis grid. A Random Forest regression model (50 trees) was trained utilizing the full 64 embedding dimensions (A00 to A63) as independent features. Finally, a 1-meter high-resolution binary forest canopy mask was integrated via modal reduction to accurately isolate forest surfaces, compute the absolute forest area (in m² and ha), and upscale the total regional AGB (Mg).

The embedding-driven regression model effectively mapped spatial variations in biomass density across the topographically rugged landscape. By applying the predictive model exclusively over the masked forest surface, the total forest AGB was precisely quantified at 14.07 million Megagrams (Mg), yielding an average biomass density of 180.75 t ha⁻¹.

This work demonstrates that Alpha Foundation Satellite Embeddings can completely bypass traditional manual feature engineering (such as spectral index calculation), as their latent layers natively encode complex structural and biophysical vegetation characteristics. The cloud-based workflow provides a highly automated, scalable paradigm for continuous biomass inventories in mountain environments.