Remote Sensing Estimation And Spatiotemporal Characteristics Analysis Of Near-surface Ozone Concentration In China Based On Ensemble Learning

Since the State Council of China promulgated the"Air Pollution Prevention and Control Action Plan"in 2013,Chinses haze weather pollution has continued to improve,and the PM_2.5 concentration in most cities has dropped significantly,but the near-ground O₃ concentration has shown an increasing trend.High concentrations of ozone have adverse effects on human health,vegetation productivity,and crop yield.The realization of precise monitoring of large-scale ozone concentrations is a prerequisite for accurate assessment of its hazards.The near-surface O₃ concentration can be obtained in real time by using ground monitoring stations.However,most of the air pollution observation stations in my country are located in urban areas,so it is difficult to fully reflect the regional ozone pollution.Satellite remote sensing can only achieve the purpose of monitoring ozone atmospheric column concentration（stratosphere+troposphere）or tropospheric column concentration on a large scale.Therefore,how to effectively combine the advantages of satellite observation data and ground measurement data to obtain large-scale near-surface O₃concentration data waiting to be solved.With the advent of the era of big data,emerging machine learning algorithms can extract valuable features from massive amounts of information,providing researchers with feasible solutions for data mining and model building.Therefore,based on the ensemble learning algorithm,combined with Trop OMI satellite remote sensing observation data,ERA5 meteorological data and other auxiliary data,this paper established a near-surface O₃ concentration remote sensing estimation model,obtained a daily data set of near-surface O₃ concentration in China,and further explored China Spatial and temporal distribution patterns of near-surface O₃ concentrations.The main conclusions are as follows:（1）The satellite-observed tropospheric formaldehyde（HCHO）column concentration data were introduced as the characteristic factor of the ozone concentration estimation model.HCHO can represent the volatile organic compounds（VOCs）in the precursors of ozone generation,but it is rarely reported in the existing near-ground remote sensing estimation models.On this basis,15 characteristic factors affecting O₃ concentration were preliminarily delineated through the generation mode of ozone and related prior knowledge.Using the correlation coefficient method,the correlation between different characteristic factors and the near-surface O₃ concentration was obtained.Through comparison and screening,14characteristic factors were finally added to the model.The first five characteristic factors that have a greater correlation with the near-surface O₃ concentration are:net solar radiation（r=0.59）,2m temperature（r=0.57）,2m dew point temperature（r=0.43）,Trop OMI-NO₂ concentration（r=-0.285）,Trop OMI-HCHO concentration（r=0.219）,10m wind speed（r=0.135）.The selected characteristic factors were matched with the observation data of O₃ concentration at the ground station in time and space to generate a near-surface O₃ concentration estimation data set.（2）The decision tree algorithm,the typical random forest algorithm in the bagging category and the latest Light GBM method in the boosting category were selected respectively,and a remote sensing estimation model of near-ground O₃ concentration was established.The coefficient of determination（R²）,root mean square error（RMSE）,mean square error（MSE）,and mean absolute error（MAE）were used as indicators to evaluate the model.The models all passed ten-fold cross-validation,and the results showed that the decision tree model R² was 0.76,RMSE was 22.37（μg/m3）,and MAE was 16.24（μg/m3）,the random forest model R² reached 0.86,the RMSE was 16.88（μg/m³）,and the MAE was 12.05（μg/m³）;the model established by Light GBM reached 0.90,and the RMSE was 13.94.（μg/m³）,MAE was 10.06（μg/m³）.In this paper,the characteristic factor of HCHO is introduced,and the near-ground O₃ concentration remote sensing estimation model based on Light GBM is constructed.（3）The performance of the near-ground ozone concentration estimation model based on Light GBM on different time scales was compared and analyzed.Compared with the daily-scale estimation results,the model has a more significant improvement in the estimation accuracy of the monthly and annual mean values of near-surface O₃ concentrations.Prediction of monthly mean,R²:0.97,RMSE:6.27（μg/m³）,MAE:4.46（μg/m³）;for prediction of annual mean,R²:0.97,RMSE:2.54（μg/m³）,MAE:1.78（μg/m³）.It is of great significance for long-term air quality research.（4）Based on the Ligth GBM model,a dataset of daily average near-surface O₃ concentrations in my country in 2019 was constructed to explore the temporal and spatial distribution characteristics of ozone distribution.The map analysis shows that there is a seasonal change in the concentration of O₃ near the ground,and the trend throughout the year is to increase first and then decrease;the average concentration in December is 70.05μg/m³,which is the lowest in the whole year;it reaches its peak in June,and the ozone pollution is serious,and the monthly average value is 116.54μg/m³,and the local concentration reaches226.54μg/m³;in terms of spatial distribution,the O₃ concentration in central China is the highest,with an annual average of 111.24μg/m³,followed by East China（110.31μg/m³）,Northwest China（98.51μg/m³）,North China（95.51μg/m³）,South China（93.54μg/m³）,Southwest（91.88μg/m³）,Northeast（81.21μg/m³）.The O₃ concentration level of each major urban agglomeration is closely related to population and economy.The top five provinces with the highest O₃ concentration are Shandong（118.57μg/m³）,Henan（117.59μg/m³）,Jiangsu（116.82μg/m³）,Shanghai（115.95μg/m³）,Anhui（114.50μg/m³）.