Interpretable GNN–Residual Kriging for spatial prediction and investigation-priority zoning of soil mercury in a karst agricultural region

Abstract

Introduction Soil mercury (Hg) in karst agricultural regions is commonly characterized by strong local heterogeneity, which limits the ability of conventional interpolation methods to delineate localized enrichment. Methods We developed an interpretable graph neural network–residual kriging (GNN-RK) framework for soil Hg prediction, uncertainty assessment, and investigation-priority zoning. The framework integrates 13 non-element environmental covariates, graph-based spatial learning, residual kriging, bootstrap uncertainty estimation, model-agnostic feature attribution, and GeoDetector analysis. It was applied to 149 surface-soil samples from agricultural land in Zhijin County, Guizhou Province, Southwest China. Results Soil Hg ranged from 0.0256 to 10.5983 mg kg −1 (mean 0.321; median 0.141 mg kg −1 ). Under the pH‐specific threshold scenario for non‐paddy agricultural land in GB 15618-2018, 147 samples were below the risk-screening values and two exceeded the risk-control values. In label-masked transductive five-fold spatial block cross-validation, GNN‐RK achieved the highest pooled performance among the evaluated models (R 2 = 0.632, RMSE = 0.695 mg kg −1 , MAE = 0.201 mg kg −1 ). Lithology was the primary explanatory factor, with BSI, distance to road, pH, nighttime light, NDVI, and elevation acting as secondary controls. Discussion The results suggest that localized Hg enrichment is associated with the spatial coupling of lithological background, surface exposure, and accessibility‐related disturbance. The spatially limited priority zones, mainly in eastern Zhijin, should be treated as targets for supplementary sampling, crop Hg testing, and field verification rather than as confirmed contaminated land.