Hybrid Machine Learning Simulation Of Particulate Matter Mass Concentration Based On Image Features

Lanzhou is a typical valley city.Analyzing the characteristics of persistent pollution and reasonably assessing the mass concentration of particulate matter that plays an important role in the process of severe persistent pollution is helpful to provide reference for the prevention and control of air pollution and management strategies in Lanzhou.Various methods and techniques have been used to evaluate the mass concentration of pollutants in the air,such as using ground monitoring points for surface monitoring,evaluating the mass concentration of particulate matter through aerosol optical thickness or using numerical simulation methods,but the above methods have certain limitations and shortcomings.Based on the daily concentrations of AQI,PM2.5,PM10,SO2,NO2,O3(8h)(8-hour moving average of ozone)and CO during 2014-2021,the interannual and monthly variation characteristics of air pollution were analyzed in Lanzhou,a valley city,and the continuous characteristics of air pollution were emphatically analyzed in Lanzhou.Secondly,a low-cost and convenient method was studied to simulate air pollutant concentration through image machine learning.Three different outdoor images taken in Lanzhou City from 15:00on October 21,2021 to 20:00 on May 6,2022 were used,and nine different deep learning models of convolutional neural networks were adopted.The mass concentration of PM2.5 and PM10,which played a major role in the process of longterm severe and continuous pollution,was simulated,and the most suitable training images and models were selected for image feature extraction.Finally,based on the extracted image features and combined with meteorological elements and pollutant concentration characteristic factors,10 machine learning models were established to simulate the mass concentration of PM2.5 and PM10.A hybrid machine learning model was established by comparing and selecting the optimal machine learning model.Finally,the Res Net50 model was used to extract the image features of the building image,the random forest importance analysis was used to select the image features,meteorological element features and pollutant concentration features,and the multiple regression model was used to train three selected key features to simulate the particle mass concentration.The optimal Res Net50,random forest and multiple regression hybrid machine learning model was established.A test set was established using image data,meteorological factor data and pollutant concentration data from 06:00May 7,2022 to 10:00 June 6,2022 to evaluate the model.The main conclusions are as follows:(1)In recent 8 years,the annual average of AQI in Lanzhou showed a general trend of fluctuation and decline.The air quality was mainly good and mildly polluted,while the number of days with heavy and severe pollution was relatively small.The air pollution index fluctuated more in winter than in spring,and showed a downward trend.The levels of light pollution,moderate pollution and heavy pollution were more serious in spring than in winter;The level of severe pollution was more serious in winter than in spring,but the fluctuation of severe pollution process in spring was more stable.The monthly changes of AQI in eight years were similar,showing a trend of first decreasing and then rising.(2)Different levels of long-term and short-term continuous pollution processes are defined.The number of short-term mild continuous pollution processes,long-term moderate and severe continuous pollution processes was the most in winter;the number of short-term moderate and severe continuous pollution processes was the most in spring;the number of short-term moderate and severe continuous pollution processes was the most in summer;the number of long-term mild continuous pollution processes was more in autumn and winter.The pollutants that play a major role in the process of continuous pollution were different according to the season,duration and grade: In winter,PM2.5 was the main pollutant in the mild pollution,moderate pollution and short-term serious continuous pollution process,while PM2.5and PM10 were the main pollutants in the long-term serious continuous pollution process.PM10 was the main pollutant during the continuous pollution in spring.In summer,the main pollutant was O3(8h).In autumn,the main pollutant was PM10 in the short-term mild continuous pollution process,and PM10,PM2.5 and NO2 in the long-term mild continuous pollution process,respectively.(3)Nine convolutional neural network deep learning models were used to simulate PM2.5 and PM10 mass concentration through the sky images and the building images.Compared with the learning and training of the original images,the simulation effect of the sky images is worse.On the contrary,the simulation effect of the building images is better.From the simulation results of the verification set,it can be seen that when the building images are taken outdoors and Res Net50 model is used for training,the overall simulation effect of the verification set is the best,and the change of PM2.5 and PM10 mass concentration can be simulated more accurately.(4)The image features,meteorological features and pollutant concentration features selected by random forest importance analysis were brought into 10 machine learning models for simulation.Compared with the direct simulation of two kinds of particulate matter mass concentration by using outdoor building images,after adding meteorological features and pollutant concentration features,except SVR and Ada Boost,Other models have greatly improved the simulation ability of verification set,among which the multiple regression model has the best simulation effect.(5)According to the simulation results of the test set,it can be seen that the image features of building images were extracted by Res Net50 model,then three key features were selected by random forest importance analysis,and a hybrid machine learning model of particulate matter mass concentration was obtained by using multiple regression model training.Although the simulation ability of low concentration of PM2.5 in some periods is relatively poor,But the overall simulation effect of PM2.5 change is better;The model can simulate the change of PM10,and compared with PM2.5,the change of simulated value of PM10 is closer to the change of observed value.