Aboveground biomass (AGB) in grasslands is a key biophysical indicator for evaluating grassland productivity, ecosystem functioning, and carbon storage. However, accurate regional-scale AGB estimation in complex mountainous terrain remains challenging because of fragmented topography, strong environmental gradients, and heterogeneous grassland patches. This study evaluated the applicability of satellite embedding features for grassland AGB estimation in the Three Parallel Rivers region of Yunnan Province, China. Based on 135 field plots surveyed during the 2022 growing season, 64-dimensional annual satellite embedding features were extracted, and a conventional feature system derived from Sentinel-1, Sentinel-2, and DEM data was constructed for comparison. Four feature systems, namely Traditional-9, Traditional-40, Emb-9-PCA, and Emb-64, were evaluated using six regression models, including RF, SVR, GPR, XGBoost, LightGBM, and Elastic Net. Random five-fold cross-validation was used to compare feature systems and model combinations, while spatial cross-validation was further applied to assess model robustness under spatially independent conditions. The results showed that satellite embedding features outperformed conventional remote sensing features. Under random five-fold cross-validation, the Emb-64-based XGBoost model achieved the best performance, with an R2 of 0.7949 ± 0.0405, an RMSE of 0.1388 ± 0.0191 t/ha, and an MAE of 0.1141 ± 0.0203 t/ha. Under spatial cross-validation, XGBoost retained the highest mean performance, with an R2 of 0.7660 ± 0.1003, an RMSE of 0.1417 ± 0.0111 t/ha, and an MAE of 0.1165 ± 0.0124 t/ha. SHAP and Spearman correlation analyses further indicated that important embedding dimensions were associated with AGB and selected conventional environmental variables. Regional mapping showed that predicted grassland AGB ranged from 0.17 to 1.26 t/ha, with a mean value of 0.79 t/ha, and exhibited significant positive spatial autocorrelation. Bootstrap-based uncertainty analysis indicated that higher uncertainty mainly occurred in fragmented mountainous areas, grassland edges, and transition zones. These findings suggest that satellite embedding features provide a promising high-dimensional representation for grassland AGB estimation in complex mountainous landscapes, while their ecological interpretability and transferability still require further investigation.
Liu et al. (Tue,) studied this question.