Study On Estimation Of Near-surface PM_2.5 Concentration Based On Neural Network Model

With the rapid development of China’s social economy,PM_2.5fine particulate matter,as a kind of air pollutant,has become a major problem of environmental governance in China in recent years.The establishment of monitoring stations to monitor the near-surface PM_2.5 concentration was the earliest action taken by China.However,due to the late establishment and sparse spatial distribution of surface environmental monitoring stations in China,the long-term and regional continuous distribution changes research of PM_2.5distribution is greatly restricted.Using the relationship between satellite remote sensing data and PM_2.5 to estimate PM_2.5 concentration can effectively break through the time and space limitations of PM_2.5 observation.Therefore,how to simulate this relationship and solve the temporal and spatial differences of PM_2.5 distribution is a major topic of this thesis.The main research content of this thesis is to introduce the capsule structure and dynamic routing algorithm into the PM_2.5 concentration estimation task,and build a capsule network model to solve the spatial differences of PM_2.5 concentration distribution.Long short-term memory network was introduced to deal with the changes of PM_2.5 concentration distribution over time,and a two-level model was formed with the capsule network.Finally,the daily mean PM_2.5 images were produced according to the existing data and the temporal and spatial variation characteristics of PM_2.5 in China were analyzed.The main research results of this thesis are as follows:First,the deep neural network model with capsule structure and dynamic routing algorithm can achieve a higher estimation accuracy than the ordinary neural network model,and the estimation value of PM_2.5 of monitoring stations in sparse areas can achieve a higher fitting accuracy.The estimation ability of the model is analyzed in terms of time,and the model can obtain good fitting results in both seasonal and annual scales.Second,Learning the weight of time through the long short-term memory network,the estimation results of the capsule network can be applied with the time weight to improve the overall estimation accuracy.The improvement effect was investigated in different seasons,and it was found that the estimation results of summer improved greatly,while those of spring,autumn and winter improved less.Third,PM_2.5 concentration distribution in China is mainly related to China’s topographic features and regional socio-economic development level.PM2.5 is mainly concentrated in China’s plains,basins and other flat areas,where the frequency of economic and social activities is high and pollutants are not easy to diffuse due to geographical environmental factors,resulting in the accumulation of pollutants.The characteristics of PM_2.5 in China are as follows:the pollution is most serious in winter,the pollution is least in summer,and the pollution degree in spring and autumn is between winter and summer.In the past 10 years,the average annual concentration of PM_2.5 in China has shown a downward trend,and the decline is most obvious in the last three years.