Drivers and threshold responses of surface structural vulnerability in the hilly dryland of southern China

Surface structural degradation, manifested as aggregate breakdown and crust formation, represents a critical constraint to sustainable dryland farming in the hilly regions of southern China. However, systematic quantification of surface structural vulnerability - the susceptibility of soil surfaces to such degradation - and its key controlling factors and threshold responses remains limited. Here, we collected 99 soil samples from dryland fields across the hilly regions of southern China and characterized 24 physicochemical indicators. We employed path analysis alongside machine learning algorithms - including decision tree, random forest, gradient-boosted regression tree (GBRT), and extreme gradient boosting - as well as SHapley Additive exPlanations and locally weighted scatterplot smoothing (LOWESS) regression to delineate the key drivers of surface structural vulnerability and quantify its nonlinear response thresholds. Path analysis revealed that sand content, CEC, total phosphorus, coarse macroaggregates, and alkaline phosphatase all exhibited significant direct or indirect associations with surface structural vulnerability. Machine learning models were subsequently constructed using these five soil properties. Leave-one-out cross-validation demonstrated that the GBRT model achieved the best predictive performance (R² = 0.76), with sand content emerging as the most influential predictor, followed by CEC. SHAP analysis revealed pronounced nonlinear interaction effects among indicators, with sand content and CEC identified as the two dominant variables modulating feature interactions. LOWESS regression further quantified critical response thresholds for each indicator: 55.5% for sand content, 5.8 cmol/kg for CEC, 0.06 mg/g/2h for alkaline phosphatase, 1.4 g/kg for total phosphorus, and 1.3% for coarse macroaggregates. These findings offer reference thresholds for monitoring surface structural vulnerability and inform future soil structural conservation strategies in regional dryland systems.