| 摘 要： |
Soil pollutants emitted from pesticide sites seriously threaten human health, and predicting the human health risk (HHR) is necessary to the control and management of these sites. A database of 1933 pesticide sites in China, including 32 surveyed site samples, was created through big data searching and cleaning. First, six risk predic-tion indicators were screened through a correlation analysis, ANOVA, and cardinality test. The six indicators included soil pollution proba-bility, proportions of the population over 60 and under 15 years old, soil sand at the 30-100 cm layer, elevation, and proportion of arable land area within 1 km of the sites. Second, the surveyed samples were divided into training and testing sets at a ratio of 8:2, and the synthetic minority oversampling technique was used to achieve sample size rebalancing in the training set between positive and counter examples. Two semi -supervised learning models based on the self-training method and label propagation algorithm (SSL-LPA) were trained and tested. SSL-LPA was screened for the prediction of HHR probability, and it had the highest prediction accuracy with a cross-validation accuracy of 93% and recall rate of 100%. Lastly, about 3.7% of the un-surveyed sites were predicted by SSL-LPA as risky sites with an HHR probability of over 0.94, and they were mainly distributed in northern and eastern China. These sites were predicted as risky due to the effects of soil pollution proba-bility and the spatial distribution of pesticide sites, especially in the surveyed risky sites. This study provides technical support for HHR control and management of pesticide sites in China. |
在线留言
联系我们
所长信箱