污染源解析是土壤重金属污染防控和风险管理的必要前提，多变量统计分析、基于GIS空间分析和地统计学方法的源解析模型方法已有广泛的研究，但针对小尺度多来源多介质土壤重金属污染缺乏成熟的解析方法，其难点在于无法建立针对不同区域尺度和污染特性的普适性方法。本项目从自然地理、土壤地球化学及社会经济三个维度，研究不同污染途径和污染相态下的土壤重金属污染。通过识别镇域尺度农田土壤重金属铅和镉的主要污染途径、范围和程度，基于随机梯度提升模型结合区域土壤重金属污染特征与来源特性，应用交叉检验和随机森林模型进行验证与优化，建立小尺度多来源多介质农田土壤重金属污染估算模型, 实现农田土壤多途径污染源解析及其贡献率的定量分析，为合理防控土壤重金属污染及制定管理措施提供科学依据。其创新在于：通过应用随机梯度提升模型结合区域土壤重金属污染特征与来源特性，建立小尺度多来源多介质农田土壤重金属污染估算模型。

The identification and apportionment of pollution sources is a necessary prerequisite for pollution prevention and control and risk management of heavy metals in soils. Modelling approach of source identification and apportionment including multivariate statistical analysis, GIS-based spatial analysis methods and geostatistics have been extensively studied. However, mature analyzing methods are shortage aiming at multi-source and multi-media pollution of heavy metals in soils. The difficulty lies in that a universal method for meeting different regional scales and pollution properties is not able to be established. This study will assess the pollution of heavy metals in soils under different pathways and media from three dimensions, physical geography, soil geochemistry and socio-economy. We will identify the main pathway, the range and the extent of heavy metal (lead and cadmium) pollution in agricultural soils at a town scale. Stochastic gradient boosting is used in combination with regional pollution features and the characteristics of pollution sources. Random forest model and cross validation are applied for the validation and optimization of prediction results. A model for estimating the multi-source and multi-media pollution of heavy metals in agricultural soils is developed. Multi-channel pollution sources in agricultural soils are identified and apportioned quantitatively. This will provide a scientific basis for the rational prevention and the draft of management measures of heavy metal pollution in soils. The innovation of this study is: A model for estimating the multi-source and multi-media pollution of heavy metals in agricultural soils is developed based on stochastic gradient boosting in combination with regional pollution features and characteristics of pollution sources.