Machine learning method for predicting cadmium concentrations in rice near an active copper smelter based on chemical mass balance.

Identification the sources of heavy metals can effectively control and prevent agricultural soil pollution. Here we performed a three-year mass balance study along a gradient of soil pollution near a smelter to quantify the potential contribution and net cadmium (Cd) fluxes and predict Cd concentration in rice grains by multiple regression (MR) and back propagation (BP) neural network. The Cd inputs were mainly from the irrigation water (54.6–60.8%) in the moderately polluted and background sites but from atmospheric deposition (90.9%) in the highly polluted site. The Cd outputs were mainly from the surface runoff (55.8–59.5%) in the moderately polluted and background sites, but from Sedum plumbizincicola phytoextraction (83.6%) in the highly polluted site. The soil Cd concentrations, the annual fluxes of atmospheric deposition, pesticides and fertilizers, irrigation water, surface runoff, and leaching water were selected as the dependent factors to predict Cd concentrations in rice grains. The genetic algorithms (GA)-BP neural network model gives the best prediction accuracy compared to the BP neural network model and multivariate regression analysis. The major implication is that the health risks through the consumption of rice can be rapidly assessed based on the Cd concentrations in rice grains predicted by the model. [Display omitted] • A 3-year study of Cd mass balances were performed by an accurate plot experiment near a smelter. • Irrigation-water Cd input contributed 55–61% to background and moderately polluted cultivated soils. • Surface-runoff and leaching-water outputs were the main output pathways of Cd in agroecosystem. • Rice grain Cd were accuracy predicted by GA-BP neural network based on mass balance factors. [ABSTRACT FROM AUTHOR]