Applicability-aware inference of salinisation-metal risk associations from sparse soil data

Provincial cropland databases often combine widely measured salinity-fertility variables with sparsely measured element panels, creating structural missingness. Using 7,282 cropland samples from Zhejiang Province, China, we transferred target variables only when supported by county-blocked validation, an explicit applicability domain and calibrated uncertainty. A composite salinity-fertility background gradient (SP) was derived from seven high-coverage variables and externally validated against measured topsoil salt content in 4,253 samples (Pearson r = 0.660). County-blocked validation retained B-MDS1, C-MDS1, C-MDS2 and log(RI + 1), with out-of-fold R 2 values of 0.164, 0.347, 0.179 and 0.171, respectively; scale-adaptive cross-fitted conformal 95% intervals achieved 95.3% out-of-fold coverage. Among 1,342 samples with observed five-metal RI, SP was weakly positively associated with log(RI + 1) (Pearson r = 0.168; Spearman rho = 0.258). A sensitivity model including the seven SP constituents yielded r = -0.038 and rho = 0.053, whereas adjustment for independent spatial and soil-landscape covariates yielded r = 0.282 and rho = 0.372. Thus, the association depended on the adjustment scheme and did not establish causality or source pathways. County-cluster bootstrap screening identified 82 low-SP/higher-observed-RI and 82 high-SP/lower-observed-RI units, but transfer uncertainty supported no directional labels outside the observed panel. Cd and Hg contributed median shares of 40.0% and 37.1% to RI. These results support targeted validation and monitoring-network redesign rather than site-level management zoning.