基于集成学习方法的中国近地面臭氧浓度时空分布

Estimating ground-level ozone concentration in China using ensemble learning methods

自2013年大气污染防治行动以来,PM_(2.5)、PM_(10)、SO_(2)、NO_(2)等空气污染物浓度都有不同程度下降,但臭氧污染仍有上升趋势,臭氧污染已成为制约中国空气质量持续改善的关键问题。地基站点可以提供空间上特定点的臭氧浓度,但无法获得近地面臭氧连续的空间分布。由于臭氧大量分布于平流层,遥感卫星反演的臭氧柱浓度产品仅能反映整层臭氧柱浓度,但整层臭氧柱浓度与近地面浓度无明显相关性,因而无法体现近地面臭氧浓度。本文综合地基监测数据、再分析资料、卫星产品,采用不同的模型方法,得到近地面臭氧浓度的时空分布,结果表明集成学习方法可以准确估算近地面臭氧在空间上的分布状况和在时间上的变化趋势。本文对比了梯度提升回归树(GBRT)、极端随机树(ERT)、极端梯度提升器(XGBoost)3种不同的集成学习方法在近地面臭氧污染估算的效果表现,3种集成学习方法在2019年—2020两年的十折交叉验证R^(2)都在0.89以上,极端梯度提升器(XGBoost)方法在RMSE、MAE指标上有最好的表现,2019年—2020年两年的平均RMSE、MAE分别为15.77μg/m^(3)、10.53μg/m^(3),但基于极端随机树(ERT)方法获得的近地面臭氧空间分布更加连续自然。因此最终选择极端随机树(ERT)方法估算得到中国近地面臭氧浓度数据集,并在此基础上进行时空分析。由于中国政府实施积极的减排措施的及疫情影响,臭氧浓度多年来的上升趋势得到了逆转,2020年臭氧年均值为107.41±18.6μg/m^(3),较去年平均值109.26±19.71μg/m^(3),有所减少。时间上,每年的5—9月气温较高,光化学反应剧烈,因而臭氧高污染事件频繁发生。空间上,京津冀地区、长三角地区、珠三角地区、成渝地区等城市群显著高于周围其他区域,是臭氧污染防治的重点区域。

Following the successful implementation of the Air Pollution Prevention and Control Action Plan(2013—2017)and the Three-Year Action Plan to Win the Blue Sky Defense War(2018—2020),the concentrations of five major pollutants(i.e.,PM_(2.5),PM_(10),SO_(2),NO_(2),and CO),except for ozone,significantly dropped for most cities in China.The increasing ground-level ozone concentrations have been a key factor restricting the improvement in ambient air quality,especially during summer.Compared with the measurements from ground-based monitoring sites,satellite remote sensing technology can obtain spatially continuous total column ozone.However,given that ozone is abundantly distributed in the stratosphere,ground-level ozone has a very low contribution to the total column ozone observed from space.Therefore,the satellite total column ozone product provides limited information for estimating ground-level ozone concentrations.In this study,by combining TROPOMI ozone precursor(NO_(2)and HCHO)products,ERA5 meteorological parameters,and ground-based monitoring data,a machine learning model was developed to estimate the daily maximum 8-hour average ground-level ozone concentration over China from years 2019 to 2020.By comparing the performance of three ensemble learning methods,namely,extreme gradient boosters(XGBoost),extreme random trees(ERT),and gradient boost regression tree(GBRT),the averaged overall 10-fold cross-validation R^(2)of 2019 and 2020 are all larger than 0.89.Although the results estimated by XGBoost showed the best agreement between the model predictions and observations with an average RMSE and MAE of 15.77μg/m^(3)and 10.53μg/m^(3),respectively,the ERT method was eventually selected to model the daily maximum 8-hour average ground-level ozone concentration by considering the rationalization of spatial distribution.Due to the proactive emission reduction measures implemented by the Chinese government and the impact of the COVID-19 pandemic,the rising trend of ozone concentration over the years has been reversed.The annual average value of ground-level ozone concentration in 2020 reached 107.41±18.6μg/m^(3)over China,which is 1.85μg/m^(3)less than that recorded in 2019(109.26±19.71μg/m^(3)).Severe surface ozone pollution events frequently occur from May to September of every year because the high temperatures during these months can promote photochemical reactions.The estimated ground-level ozone concentrations in the Beijing-Tianjin-Hebei,Yangtze River Delta,Pearl River Delta,and Chengdu-Chongqing regions are significantly higher than those in their surrounding areas,making these regions the key areas for ozone pollution prevention and control.