Risk Assessment And Spatial Prediction Of Heavy Metals For Farmland Soils At Regional Scale

Farmland soils have long been affected by human activities such as application of pesticide and fertilizer,sewage irrigation,and emissions of industry and mining,which have resulted in accelerated accumulation of heavy metals and their potential risks.Therefore,study on risk assessment and spatial prediction of heavy metals in farmland soils is an important basis for ensuring soil ecological security,crop and human health.The traditional methods of risk assessment and spatial prediction of heavy metals in soil are guided by“sink”,in which the correlation factors of heavy metals are less considered,and the“linear-nonlinear”response relationship between the correlation factors and heavy metals is unclear.For this end,this study integrates the multi-source factors closely related to soil heavy metals(such as topography,climate,soil,vegetation,emission sources,socio-economic,and remote sensing),and constructs the index system suitable for risk assessment and the set of auxiliary variables for spatial prediction,respectively.Combining with multiple index model,index scale AHP model,BNs model,RSM,and linear-nonlinear spatial prediction model,the risk degree and risk source of heavy metals in regional farmland soils were analyzed.Then the variability of heavy metals was predicted by linear-nonlinear prediction and spatial mapping,and the driving force of auxiliary variables was analyzed by dose-effect analysis.The study results were more consistent with the first(spatial autocorrelation)and second(spatial heterogeneity)laws of geography,which provides technical support and scientific basis for regional farmland-agricultural product safety control and management decision-making.The main results of this study were as follows:1.Risk assessment of multiple index of heavy metals in farmland soilsResults of descriptive statistical analysis showed that the mean values of As,Cd,Cr,Hg and Pb were 12.45 mg kg^-1,0.22 mg kg^-1,62.99 mg kg^-1,0.22 mg kg^-1and 53.15 mg kg^-1,respectively.And the maximum value did not exceed the limit of the risk screening value of the national standard(GB15618-2018)by Chinese Environmental Protection Agency,but it exceeded the maximum background levels in varying degrees in comparison with risk screening values for soil heavy metal of PRD(DB 44/T1415-2014)by Administration of Quality and Technology Supervision of Guangdong Province.Results of single factor index evaluation showed that content of Cr did not exceed the standard in different soil p H interval thresholds,while other heavy metals slightly exceeded the threshold values in accordance with“GB15618-2018”.On the whole,the health condition of farmland soils in the study area was good.Evaluation results of Nemerow comprehensive pollution index showed that the proportion of safe,alert pollution,and light pollution level was 37.5%,32.61%,and 29.89%for all heavy metals in accordance with“GB15618-2018”,respectively.In addition,risk evaluation results of pollution load index for all heavy metals were in safety level.According to“GB15618-2018”,evaluation results of Hakanson potential ecological risk index showed that 3.8%of samples of Hg were medium ecological risk level,and other heavy metals were low ecological risk level.The comprehensive risk level of Hakanson potential ecological risk index for all heavy metals was slight ecological risk level.2.Multi-factors risk assessment of heavy metals in farmland soilsThe risk assessment results of index scale AHP model showed that the heavy metals in farmland soils ranged from risk-free to medium risk level,and the accuracy(correct rate)of As,Cd,Cr,Hg and Pb is 70%,64%,58%,50%and 72%,respectively.The risk evaluation results of BNs model showed that the evaluation accuracy of heavy metals was from high to low:As(76%)>Pb(74%)>Cd(72%)>Cr(66%)>Hg(64%),which was better than the evaluation accuracy of index scale AHP model.The evaluation accuracy of Bayesian network model for As and Pb was at risk-free to medium-risk level,Cd and Hg were from risk-free to high-risk level,and Cr was from risk-free and slight-risk level.In addition,the high risk area of soil heavy metals was around the core area of the city.3.Quantitative analysis of heavy metal risk sources in farmland soils at regional scaleComparing linear traffic source(road)and point traffic source(intersection and parking lot),it was found that the response distance of content of soil heavy metals to point traffic source was longer and the concentration of heavy metals decreases slowly.This was related to the emission and wear of vehicle which caused by long stay in the intersection or parking lot.The integrated response distances of roads,factories,transportation facilities and rivers to heavy metals were 60-1680 m,25-710 m,27-2000 m and 10-2310 m,respectively,and the synergistic effects of various distance sources were uncertain.The results of RSM analysis showed that the integrated point source exhibited the highest contribution rate for As(33.95%)and Pb(63.13%).The application of chemical fertilizers and pesticides was the most important source input for Cd and Hg,which the contribution rate was 29.53%and 33.55%respectively.And the contribution rate of agricultural water consumption for Cr reached 62.84%.4.Dimension reduction of auxiliary variables and spatial prediction of geostatistical modelAccording to the analysis results of VIF-Monte-Carlo-Pearson,the set of variables corresponding to As(45),Cd(50),Cr(44),Hg(47)and Pb(45)was selected.Based on results of PCA(i.e.cumulative contribution rate greater than 80%),the first seven PCs of heavy metal As,the first ten PCs of Cd and Hg,and the first eight PCs of Cr and Pb were obtained as inputs for the subsequent spatial prediction models.Compared with SK interpolation results,the OK interpolation accuracy of all heavy metals was higher,the order of R²from high to low was 56.64%(Hg)>50.51%(Cd)>48.72%(Pb)>44.69%(Cr)>41.86%(As),respectively.The results of CK with the first three PCs as covariate respectively explained the spatial variability of 62.56%(Cd),59.45%(Hg),46.26%(Cr),44.17%(Pb),and 30.33%(As)in the test data set,and the regions with high content were concentrated around the urban area.5.Prediction by machine learning method for the spatial variability of soil heavy metalsThe prediction accuracy of the ELM model for soil heavy metals in the test data set was from high to low:Cd(RMSE=0.091 mg kg^-1,R²=72.03%)>Pb(RMSE=11.79 mg kg^-1,R²=71.72%)>Hg(RMSE=0.104 mg kg^-1,R²=70.22%)>As(RMSE=5.57 mg kg^-1,R²=69.44%)>Cr(RMSE=21.63 mg kg^-1,R²=61.61%).Compared with the above-mentioned models,the prediction accuracy of ELM model was higher and the predicted range of the content of soil heavy metals was closer to the real values of samples.The spatial mapping results showed that the high content of soil heavy metals in the study area were distributed in the central-western and southwest regions for As,mid-west and north-west regions for Cd,south-central regions for Cr,west-west and south-west regions for Hg,and mid-west regions for Pb.The results of driving force analysis using machine learning models to auxiliary variables showed that remote sensing and vegetation variables contributed greatly to the fitting power of Cd,Hg and Pb,and the influence of emission source variables on As and Cr was significant.That is,Hg and Pb were driven by natural variables and human activity variables simultaneously,and As and Cr were mainly driven by human activity variables,while Cd was mainly driven by natural variables.6.Hybrid geostatistical prediction of spatial variability of soil heavy metalsThe prediction results of hybird geostatistical models(i.e.SLROK,SVMOK,CARTOK,RFOK,and ELMOK)indicated that,except for As(the accuracy of R²by CARTOK was 79.62%),the prediction performance of ELMOK model on Cd(RMSE=0.072 mg kg^-1,R²=79.53%),Cr(RMSE=19.48 mg kg^-1,R²=68.68%),Hg(RMSE=0.092mg kg^-1,R²=76.98%)and Pb(RMSE=11.21 mg kg^-1,R²=79.83%)was better than other models,which indicated that ELMOK had a good fitting ability to the"linear-nonlinear"relationship between soil heavy metals and auxiliary variables.The mapping results of the hybird geostatistical models showed that the SLROK and SVMOK models were overestimated to some extent on the content of soil heavy metals,while the spatial mapping results of CARTOK,RFOK and ELMOK models were more close to the measured values.The high concentration regions of soil heavy metals by CARTOK,RFOK and ELMOK models models were concentrated in the surrounding areas of city proper and south of the study area.Compared with geostatistical method and machine learning method,hybird geostatistical method had the strongest fitting performance for complex relationship between auxiliary variables and soil heavy metals,especially the prediction accuracy of ELMOK model and its spatial mapping results were better.And the mapping of spatial variability of soil heavy metals was more consistent with the first(spatial autocorrelation)and second(spatial heterogeneity)laws of geography.