Metals source apportionment in farmland soil and the prediction of metal transfer in the soil-rice-human chain

Predicting Bioaccumulation of Potentially Toxic Element in Soil–Rice Systems Using Multi-Source Data and Machine Learning Methods: A Case Study of an Industrial City in Southeast China

“…The model constructed using RF yielded the best estimation of Cr bioaccumulation in soil-rice systems. The R 2 of this model was 0.79 (Table 3 and Figure 4), and it significantly outperformed the linear models constructed by Zhao et al (2009) [46] (R 2 = 0.13), Zeng et al (2011) [92] (R 2 = 0.22), and Deng et al (2020) [40] (R 2 = 0.46) (Table 5). Our models for predicting Cu, Zn, and Ni were also more accurate than those reported previously in similar case studies (Table 5).…”

“…We used 20 auxiliary variables, which have been proved to have effects on the transfer of PTEs from soil to the rice by previous studies to build a model to predict PTE bioaccumulation in soil-rice systems (Table 1) [36,40,[45][46][47][48][49]. The auxiliary variables included soil properties (such as soil organic matter, pH, soil bulk density, content of PTE in soil, cation exchange capacity, soil sand content, soil clay content, soil silt content, soil coarse fraction), climatic factors (such as annual temperature, annual precipitation), terrain attributes (such as elevation), agricultural management practices (such as amount of phosphate fertilizer applied annually, amount of organic fertilizer applied annually, amount of nitrogen fertilizer applied annually, amount of potash fertilizer applied annually), geology (soil group, parent material), and land use (land use types), and population density are detailed in Table 1.…”

“…The transfer and bioaccumulation of PTEs from soil to rice is a complicated process involving PTE release, dissolution, and bioaccumulation in different plant parts [36,37]. It is affected by various factors, including the total concentration of PTEs in soil, as well as soil physical and chemical properties [38][39][40]. Numerous studies have analyzed the shift of PTEs in soil-rice system [41][42][43][44].…”

“…Chen et al (2016) further used LR to predict the transfer of Cu, Pb, Zn, Cd, and Mercury (Hg) from soil to rice grains, using total PTE content, soil pH, and SOM as covariates, with a model R 2 of 0.24 to 0.63 [36]. Deng et al (2020) predicted the transfer of Cr, Pb, Cd, Hg, and As from soil to rice grains using MLR, with the total soil Cd concentration and soil pH as covariates, and an R 2 ranging from 0.109 to 0.456 [40]. Mu et al (2020) also predicted the Cd concentration in rice grain using LR, with the total Cd concentration, soil pH, SOM as covariates, and obtained the R 2 ranging from 0.42 to 0.66 [47].…”

“…In addition, RF and Cubist can manage quantitative and categorical variables simultaneously. Another advantage of machine learning methods is that they can provide information about the importance of variables in the model, which can aid the regulation of PTE pollution in soil-crop systems and reduce the health risks of PTE exposure [36,40,[45][46][47].…”