Spatial prediction and drivers of surface soil organic carbon density in subtropical montane forests.

Soil organic carbon (SOC) represents the largest terrestrial carbon (C) pool, and even small changes in its stock can exert great influences on regional and global C cycles. Quantifying the spatial distribution of forest SOC and identifying its key drivers are therefore central to climate change science. In this study, we collected 150 topsoil samples (0-20 cm) across six vegetation types spanning an elevation gradient of 500-1,900 m in Baishanzu National Park, Zhejiang Province. Soil organic carbon density (SOCD) was determined alongside litter stock, and the activities of three cellulolytic hydrolases-α-glucosidase (AG), β-glucosidase (BG), and cellobiohydrolase (CBH). Ten environmental covariates derived from remote sensing, climate, and topographic data were used to predict SOCD. Three machine learning models-random forest (RF), boosted regression trees (BRT), and eXtreme Gradient Boosting (XGBoost)-were evaluated using repeated nested five-fold cross-validation and nested nearest neighbor distance matching cross-validation (NNDM-CV), with hyperparameter tuning restricted to the corresponding outer-training partitions. The three algorithms showed broadly comparable predictive performance across the two validation schemes. Based on the XGBoost model, the predicted mean surface SOCD across the park was 108.6 t C hm