Network analysis and GMM-informed machine learning for scale-dependent soil quality assessment in a Mollisol watershed

Soil quality assessment in rolling hilly Mollisol watersheds faces two challenges. Traditional principal component analysis (PCA)-based minimum data set (MDS) selection is variance-oriented rather than function-oriented, and conventional machine learning models rarely accommodate spatial non-stationarity. Unlike PCA, which selects indicators based on statistical variance, network analysis (NA) identifies hub indicators through their topological position within interconnected soil property networks, providing a more function-oriented basis for MDS construction. This study integrated NA with Gaussian mixture model (GMM)-informed extreme gradient boosting (XGBoost) to address these challenges in the Jianghu watershed (29.17 km 2 ), a typical rolling-hill Mollisol landscape in Northeast China. A nested sampling design collected 63 topsoil samples across the watershed and 30 from a sub-watershed; 13 soil physical and chemical indicators were measured. NA using Louvain community detection and eigenvector centrality identified scale-dependent functional hub indicators. At the watershed scale, capillary porosity and soil organic matter served as hubs ( n = 2). At the sub-watershed scale, the MDS shifted to sand, available potassium, available phosphorus, and bulk density ( n = 4), reflecting reorganization of the soil indicator network. Four soil quality index (SQI) versions were generated using total data set (TDS) and MDS under linear and non-linear scoring schemes. GMM-derived spatial priors were incorporated into XGBoost to predict SQI at 30 m resolution, and Shapley Additive exPlanations (SHAP) analysis quantified predictor contributions. SQI_TDS_Linear achieved the highest predictive performance (R 2 = 0.47), and linear scoring consistently outperformed non-linear scoring. Topographic factors accounted for 62.29–78.89% of model importance, with elevation as the dominant predictor. Predicted SQI maps revealed high-quality soils concentrated in convergent lower-slope positions and degraded soils on ridge tops and gully margins. This framework, integrating network-based indicator selection with environmentally informed machine learning, provides a scale-sensitive and interpretable tool for soil quality assessment in erosion-prone agricultural watersheds.