Benchmarking Earth Observation Embeddings for Pan-EU Aboveground Biomass Mapping
Abstract. The paper presents a comprehensive evaluation of pre-trained Earth Observation (EO) embeddings compared to traditional remote sensing approaches trained with standard multi-sensor satellite data (Sentinel-1, Sentinel-2, and ALOS PALSAR-2, both with and without the addition of multi-scale terrain derivatives) for mapping aboveground biomass density (AGBD). The rank result was evaluated on 4 criteria: point-estimate accuracy, spatial pattern reconstruction, saturation effects, and uncertainty quantification. EO features were specifically used with tree-based (Quantile Random Forest Regression; QRFR) and deep learning (Convolutional Neural Network; CNN) models to map AGBD across varied biogeographical region in Finland (≈ 73,000 km2), Spain (≈ 273,000 km2), and the Netherlands (≈ 2,200 km2). A proportional stratified sampling design using airborne laser scanning (ALS) estimates of AGBD was implemented to generate sample tiles for each country, and country-specific models were trained. The benchmarking results show that EO embeddings achieve predictive accuracy that is highly competitive when compared with models that use Sentinel-1, Sentinel-2, and PALSAR: the lowest RMSE values were recorded by AlphaEarth – QRFR with terrain variables in the Netherlands (42.1 Mg ha−1), SpheerFM – CNN in Spain (17.5 Mg ha−1), and TESSERA – CNN in Finland (13.4 Mg ha−1). Pairing these embeddings with a CNN model yielded clear performance gains: TESSERA improved across all three test countries, while SpheerFM showed distinct enhancements in Spain and Finland. This strong performance was further corroborated by spatial pattern reconstruction metrics. In addition, a rigorous uncertainty assessment of both the QRFR and CNN models using empirical metrics, Prediction Interval Coverage Probability (PICP) and Mean Prediction Interval (MPI), showed that both models derive unbiased, well-calibrated prediction intervals. Ultimately, TESSERA utilized CNN as downstream model comes as the highest-ranked configuration. The result from different predictors set is also marginal, implying no single set dominates. The choice of applying EO embeddings for large-scale AGBD mapping requires consideration also data licensing and transparency, data volume, and their existing spatial temporal coverage.