Spatial prediction of the soil-wheat cadmium bioconcentration factor with a knowledge-guided learning model

Cadmium (Cd) readily transfers from contaminated soils into wheat grain, posing dietary risks; however, predicting the spatial distribution of the bioconcentration factor (BCF) remains challenging due to interacting direct and indirect pathways. We developed a knowledge-guided learning model using structural equation modeling to quantify relationships governing BCF. These mechanistic insights guided Random Forest coupled with high-accuracy surface modeling for spatial prediction and informed management strategy evaluation through scenario simulations. The results showed that the BCF was jointly governed by direct soil buffering and Zn-Cd competitive uptake, and indirectly by industrial emissions and wind that degraded soil properties and increased soil Zn concentrations. Validated against a large field dataset from northern China, the model achieved R² of 0.44-0.84. Spatial predictions revealed a partial mismatch between soil Cd and BCF hotspots, with high-BCF zones occurring where neutral pH, moderate organic matter, and industrial emission-related deposition converged. Scenario simulations suggested that a combined strategy of source control and zinc fertilization could reduce the high-BCF area from 22.7% to 3.9% by 2050. We proposed a spatially differentiated zoning framework for targeted wheat production management. This knowledge-guided learning model provides a transferable methodology for precise control in heavy metal-contaminated agricultural regions.