Satellite Embeddings Improve GEDI-Based Forest Aboveground Biomass Mapping Across Sampling Designs in the Greater Khingan Mountains

Mapping forest aboveground biomass typically relies on sparse spaceborne lidar measurements combined with spatially continuous optical and radar data, involving substantial preprocessing and feature engineering. Biomass models also frequently compress predictions toward intermediate values, leading to overestimation at low biomass and underestimation at high biomass. Here, we evaluated whether Satellite Embedding V1 could provide a more effective analysis-ready predictor representation and whether targeted allocation of training samples could alleviate this biomass-range compression. Nearly 600,000 quality-filtered GEDI L4A observations acquired in 2020 across the Greater Khingan Mountains were linked to either 64 satellite-embedding dimensions or 36 conventional predictors. XGBoost models were trained using simple-random, balanced, and upper-tail-enhanced sampling and evaluated using both a common independent test set and spatially independent block validation. Across all three sampling designs, embedding models increased R2 by 0.08–0.09 and reduced RMSE by 3.28–3.66 Mg/ha relative to conventional predictors in the common test, and their predictive advantage remained under spatially independent evaluation. Sampling design affected different aspects of model performance. Random sampling provided the strongest overall spatial predictive performance, balanced sampling produced the lowest overall bias, and upper-tail enrichment substantially reduced underestimation in high-biomass forests. Feature-response analysis further showed that several conventional predictors lost sensitivity at medium-to-high biomass, whereas multiple embedding dimensions retained detectable gradients across the upper biomass range. The resulting 30 m map preserved the regional biomass pattern while resolving forest edges, burned patches, and local heterogeneity that were smoothed at 1 km resolution. Satellite embeddings therefore provide an effective analysis-ready representation for GEDI-based biomass mapping, while training-sample allocation should be selected according to whether the priority is overall regional accuracy or improved calibration in sparsely represented biomass ranges.