Hybrid Machine Learning and Geostatistical Approaches for Forest Aboveground Biomass Estimation in a Subtropical Region of China

Accurate aboveground biomass (AGB) estimation in subtropical forests is critical for regional carbon accounting and sustainable forest management. However, standardized multi-source feature screening and integrated machine learning–geostatistical analysis of AGB remain limited. This study integrated six heterogeneous datasets: Landsat-8 optical imagery, Sentinel-1 SAR, topographic, meteorological, soil data and the 2014 National Forest Inventory (NFI), and established 48 predictors in subtropical forests of Anhui Province, China. A two-stage variable selection framework was applied, with Pearson correlation screening reducing the initial 48 predictors to 36 less-correlated variables, followed by the recursive feature elimination (RFE) with 5-fold spatial block cross-validation for further predictor selection. Random Forest (RF), eXtreme Gradient Boosting (XGB), Empirical Bayesian Kriging Regression Prediction (EBKRP), hybrid RF_EBKRP and XGB_EBKRP models were evaluated. Stand age and stand density were dominant predictors in both RF and XGB, contributing 33.6% and 24.0% in RF and 36.5% and 17.3% in XGB, respectively. Elevation, precipitation, and canopy cover showed secondary importance, whereas vegetation indices contributed relatively little. RF_EBKRP achieved the highest prediction accuracy (R2 = 0.77), reducing RMSE by 17.20% and 43.75% compared with RF and EBKRP, respectively. This study provides a reproducible RF–EBKRP workflow integrating nonlinear machine-learning prediction with geostatistical residual correction, supporting improved subtropical forest AGB mapping and management.