Above-ground Biomass Estimation with Geospatial Foundation Models

Accurate estimation of above-ground biomass (AGB) from satellite imagery underpins large-scale carbon monitoring, yet remains a difficult global-scale regression problem. Geospatial Foundation Models (GFMs) promise general-purpose Earth-observation representations, but their value for regression tasks is largely untested, as most benchmarks target classification and segmentation. We benchmark GFMs for global AGB regression on the AGBD dataset, distinguishing two ways in which GFMs reach practitioners: model weights run as frozen encoders (11 models, within the PANGAEA framework), and pre-computed embedding products (AlphaEarth Foundations and TESSERA). Frozen weight-distributed GFMs underperform with respect to the supervised baseline, whereas pre-computed embeddings surpass it: an MLP on AlphaEarth embeddings already beats the state-of-the-art supervised model trained on satellite imagery, and training the same model on the embeddings is best overall, while generalizing better across space and time and matching the operational ESA CCI product on independent reference data. Frozen GFM features can thus be highly informative for biomass regression, when the underlying model is trained on rich multi-modal data and served as an accessible embedding layer.