Study On Distribution Of Ground-level Ozone In China Based On Machine Learning Approaches

Ground level ozone,due to its oxidation,can cause damage to people’s health,agricultural production and ecosystem.Ozone data products with long time scale,national scale,high precision and high spatial and temporal resolution are needed as support for optimizing environmental risk assessment and management.At present,the spatial-temporal coverage of ozone surface monitoring data is insufficient in China,and the simulation results of chemical transmission models are of high uncertainty.Statistical model methods,especially machine learning methods,do not rely on emission inventory data.Monitoring data,meteorological observation data and land use data used in modeling are easier to obtain,and the prediction accuracy is higher,which is more in line with the purpose of this study.However,at present,the application of such methods in China is still few.There is a strong correlation between ground level ozone concentration and meteorological parameters,and the statistical models for PM2.5,NO2 and other atmospheric pollutants concentration have achieved good performance.Therefore,based on the above data,it is technically feasible to extend the study of ozone concentration estimation from urban regional scale to national scale.The XGBoost model is more suitable for this study under the condition that the prediction accuracy is the primary consideration and the speed and interpretability of the model are taken into account.Therefore,the purpose of this study is to develop a statistical model based on XGBoost algorithm by using ground monitoring data,meteorological data,chemical transport model simulation results and land use data,so as to estimate the temporal and spatial distribution of O3 concentration in China with long time scale and high spatial and temporal resolution.In addition,the associated factors of temporal and spatial variation of O3 concentration in China are discussed.Based on the regionalization method of the "Eight Comprehensive Economic Zones",this study establishes a single-year model and a time extrapolation model in each region,evaluates the performance of the models,and finally makes a statistical analysis based on the results of the model prediction.The main conclusions are as follows:(1)Single-year models with high prediction accuracy can be established by incorporating the simulation of chemical transport model as covariate into XGBoost model and using 10-fold cross-validation as model evaluation method.By incorporating the results of atmospheric chemical transport model simulation as covariates rather than building dependent variables,the accuracy of the model can be improved.The model performance varies in different seasons and regions:that in autumn and in northern coastal and southern coastal economic zones is better.(2)By incorporating chemical transport model simulation as covariates into XGBoost model and using by-year cross-validation as model evaluation method,a model with high time extrapolation performance can be established.According to the results of time extrapolation,the prediction accuracy of the model built in this study will not decrease significantly with the expand of time interval.The indirect evidence also shows that it is applicable to the prediction of the whole country and has good stability and reliability.It is helpful to incorporate the simulation results of chemical transport model as covariates to improve the ability of time extrapolation of the model,and to a certain extent,to reduce over-fitting.In addition,although using monthly mean for modeling can save time and computation,the final performance is not as good as using daily concentration modeling directly when comparing with the same index,and the former can not provide information on daily basis.(3)North China,Jiangsu,Zhejiang and Shanghai,Pearl River Delta,Sichuan Basin,Henan Province and Northeast China belong to the areas with serious pollution.Among them,pollution in Jiangsu,Zhejiang,Shanghai and Pearl River Delta areas is serious in most of the time,while North China and Northeast China have obvious seasonal pollution characteristics,and the average concentration of them in winter is low.In terms of long-term changes,ozone concentration in most areas of China shows an upward trend,and ozone pollution is intensified and spreading.In terms of concentration change during single year,except for a few areas,the concentration in May to August is higher in China as a whole,but lower in January and December.The largest annual variation occurred in the northern coastal area.In terms of associated factors,daily maximum temperature and sunshine duration are the two most important meteorological parameters for predicting ozone concentration.Meanwhile,the importance of coordinates and annual DOY in the model reflects the spatial and temporal heterogeneity that has not been explained by other variables in the model.