Spatial structure and multiscale environmental associations of arsenic in karst agricultural soils of Zhijin County, southwestern China

Abstract

Soil arsenic (As) patterns in karst landscapes reflect geological, topographic, hydrological, and anthropogenic conditions, but spatial explanatory power must be distinguished from predictive contribution and causality. We analyzed agricultural topsoil from 144 sampling sites in Zhijin County, southwestern China, using descriptive statistics, Yeo-Johnson transformation, semivariograms, ordinary kriging, Geodetector, an extreme gradient boosting regressor (XGBRegressor), SHapley Additive exPlanations (SHAP), and sample-level SHAP maps. Soil As concentrations ranged from 1.79 mg kg⁻¹ to values exceeding 40 mg kg⁻¹, with a mean of 17.7 mg kg⁻¹. The transformation reduced skewness, although normality remained rejected (p = 0.003). The exponential semivariogram provided the best fit (R² = 0.820), with a nugget-to-sill ratio of 0.627 and a range of 37.02 km. Back-transformed kriging indicated relatively high values mainly in northern, northeastern, and central-eastern areas. Hg, soil organic carbon, distance to mining sites, elevation, distance to rivers, and lithology had nominal Geodetector p ≤ 0.003; Hg had the largest q statistic (0.4241), and the Hg-lithology interaction had the largest joint q (0.5030). Among five candidate regressors, XGBRegressor achieved the highest pooled out-of-fold R² (0.4197) and the lowest root mean square error (6.7793 mg kg⁻¹) and mean absolute error (4.5831 mg kg⁻¹). SHAP ranked Hg, elevation, and lithology highest; Hg and elevation remained among the top three across all five folds. Sample-level maps showed heterogeneous signed contributions. Predictive performance was moderate, and spatial transferability remains unverified. These results support targeted verification and zoned monitoring without implying causal attribution.