基于机器学习和非参数估计的PM2.5风险评估

为开展区域风险评估，融合手机信令、气象和地理信息等多源数据，引入随机森林机器学习、非参数估计分位数图示法和非监督学习K-mean等方法，构建了区域PM_2.5风险评估及特征识别评价框架，在南京市区以0.3km分辨率网格为基础单元开展了案例研究.结果表明，该技术既可有效模拟PM_2.5浓度时空分布，十折交叉验证R²达到0.76，证明了准确度较高，并基于此识别出4种主要污染特征；也可有效捕捉短期人口流动导致的风险，在污染浓度不变的情况下短期人口流动会导致风险增加0.30~0.97倍.综合PM_2.5浓度和人口分布，识别出4种主要暴露风险模式，其中，研究区域6.5%的面积为高风险地区，23.0%的面积为低风险地区.“十四五”期间应加快现代科学技术在环境保护领域的应用，实施网格化和差异化的风险控制政策，维护人群健康.

Risk assessment of PM2.5 pollution based on machine learning and nonparametric estimation

A systematic approach of regional PM_2.5 risk and characterization assessment was developed in this study by integrating random forest model, Quantile-Quantile plot model, and K-mean model, based on multi-source data including mobile phone signals, meteorological data, geographic data, etc. This new approach was further applied in a case study of Nanjing at a 0.3km resolution grid. On the one hand, this new approach effectively simulated the temporal and spatial distribution of the PM_2.5 concentration with an10-fold cross-validation R² of 0.76 and screened out four major pollution characteristics. On the other hand, it effectively captured the short-term population mobility risk. Short-term population mobility increased the PM_2.5 exposure risk by 0.30~0.97 times, even keeping PM_2.5 concentration unchanged. After combining PM_2.5 concentration and population mobility simultaneously, four major PM_2.5 exposure risk modes were identified. 6.5% of the areas of Nanjing were at high risk, and 23.0% were at low risk. During the 14^th Five Year Plan, it is suggested that the government should speed up the application of modern science and technology in environmental protection and implement gridding and differentiated policies on air pollution risk control to promote human health.