Evaluation of Earth Observation Embeddings for Aboveground Biomass Mapping across European Forests

Accurate estimation of aboveground biomass is critical for comprehensive forest monitoring. This study evaluates the predictive performance and uncertainty of two novel 10-m global time-series Earth Observation embeddings, TESSERA and Google Satellite Embedding (AlphaEarth), against a baseline of combined Sentinel-1 and Sentinel-2 data. Reference biomass estimates were derived from National Forest Inventory (NFI) and airborne lidar scanning. To predict aboveground biomass (AGB) and quantify model uncertainty, we employed a Quantile Random Forest Regressor (QRFR). Additionally, we assessed the impact of integrating a global terrain height model (GEDTM) into the predictive feature space. Our results demonstrate that both the TESSERA and AlphaEarth embeddings achieve comparable accuracy to the traditional Sentinel-1 and Sentinel-2 baseline, measured by RMSE, MAE, R², and ME. Furthermore, incorporating the GEDTM topographical features improved model accuracy and reduced overall prediction uncertainty, quantified by the Mean Prediction Interval (MPI). Notably, AlphaEarth did not benefit significantly from the addition of terrain variables, likely because its underlying architecture inherently captures these features from its training data. These findings validate the efficacy of global EO embeddings for high-resolution biomass mapping while highlighting that integrating additional features, such as topographical features, improves model performance.