[Source Apportionment of Heavy Metals in Soils Based on Machine Learning Algorithms and Receptor Model].

Abstract

To analyze the source apportionment and influence factors of heavy metals in soils surrounding a coal gangue heap in Chongqing, three machine learning algorithms （decision tree （DT）, random forest （RF）, and support vector machine （SVM）） and the absolute principal component scores-multiple linear regression （APCS-MLR） receptor model were used. The surface soil results showed that the average values of Cd, Hg, As, Pb, Cr, Cu, Ni, and Zn were 0.44, 0.18, 9.92, 32.3, 129, 100, 72.8, and 148 mg·kg^-1. Combined profile soil data showed that Cd, Hg, As, Pb, Cr, Cu, Ni, and Zn were affected by human activities to varying degrees. Using machine learning algorithms analysis, RF was better than DT and SVM, and R² values of Cd, Hg, As, Pb, Cr, Cu, Ni, and Zn were 0.783, 0.728, 0.528, 0.753, 0.753, 0.853, 0.822, and 0.756. "The number of coal gangue units" （X₁）, "the vertical height difference between the sampling point and coal gangue heap" （X₂）, and "the distance between the sampling point and the coal gangue heap" （X₃） were the key driving factors by human activities. Combined with APCS-MLR model analysis, the soil in the study area was affected by natural sources, mining sources, and mixed sources （including atmospheric deposition, agricultural production, life and traffic emissions, etc.）, with contribution rates of 42.5%, 37.1%, and 20.4%, respectively. The combined application of the machine learning algorithms and receptor model can make the results of source apportionment more comprehensive, accurate, and reliable.