Forest biomass plays a central role in ecosystem functioning, carbon storage, and climate regulation, making accurate large-scale biomass estimation essential for forest monitoring and carbon accounting. Remote sensing approaches have traditionally relied on manually derived predictors based on optical, radar, and structural variables extracted from individual Earth observation sensors. Recently, geospatial foundation models have introduced an alternative paradigm based on satellite embeddings, where high-dimensional latent representations are learned from large volumes of multi-source Earth observation data. Despite their increasing availability, systematic evaluations of their predictive performance and operational applicability for forest biomass estimation remain limited.

In this study, we benchmark geospatial foundation model embeddings against conventional multi-sensor remote sensing predictors for the estimation of four forest biomass components: aboveground biomass density (AGBD), belowground biomass density (BGBD), shrub biomass density (SBD), and forest floor biomass density (FFBD). The analysis was conducted using 4,689 Spanish National Forest Inventory (SNFI) plots distributed across Mediterranean forest ecosystems. Three predictor families were evaluated within a unified modelling framework: (i) traditional predictors derived from Sentinel-1, Sentinel-2, and ALOS-2/PALSAR-2 observations, (ii) AlphaEarth embeddings, and (iii) TESSERA embeddings. Gradient boosted decision tree models were implemented using XGBoost, with hyperparameter adjustment, progressive feature selection, and bootstrap-based performance assessment. Model performance was evaluated through 10,000 bootstrap iterations using relative RMSE and relative MAE metrics. Results showed that embedding-based predictor families consistently matched or exceeded the performance of traditional predictors across most biomass components, although predictor-family behaviour varied depending on the target biomass pool. AlphaEarth embeddings achieved the lowest relative error values for AGBD, while TESSERA embeddings produced the best performance for BGBD, FFBD, and SBD. Feature-selection analyses further revealed differences in predictor-retention behavior between traditional and embedding-based predictor families. In addition to predictive benchmarking, selected predictor subsets were transferred to wall-to-wall raster products to generate spatially continuous biomass prediction maps across Mediterranean forest landscapes.

This study provides an integrated evaluation of geospatial foundation model embeddings for multi-component forest biomass estimation and contributes to the emerging assessment of foundation-model-based Earth observation products for operational forest monitoring applications.