Estimation And Mapping Of PM_2.5 By Integrating Multi-Source Sensing Data Using A Deep Learning Method

Fine particulate matter(PM_2.5)is one of the main air pollutants endangering human health,which has a great impact on climate,production and living order.It is essential to obtain fine spatiotemporal distribution of PM_2.5 concentration data for epidemiological research,environmental monitoring and governance.However,the distribution of traditional ground observation stations is sparse and uneven,which cannot meet the requirements of large-scale continuous monitoring.In addition,both the diversity and the complex nonlinear relationships of PM_2.5 influencing factors pose challenges for accurate mapping.To address these issues,this paper innovatively combined social sensing data with remote sensing data and other auxiliary variables,which can bring both natural and social factors into the modeling.This paper applied geospatial analysis methods and remote sensing image processing methods to extract features from the multi-source data.Meanwhile,the deep learning model was used to learn the non-linear relationship,and finally an hourly PM_2.5concentration data at a spatial resolution of 0.01°was generated.The main work and results of this paper are as follows:（1）Feature extraction of dynamic social sensing data with geospatial analysis methods.In this paper,web crawler technology was adopted to obtain real-time check-in data and traffic index data respectively from the“Tencent Location Big Data”service and the Nav Info Traffic Index platform.For such abstract data,effective feature extraction was carried out by means of buffer analysis,kernel density analysis,spatial interpolation and other geospatial analysis methods,so as to provide hourly dynamic social sensing data supporting for the estimation model of PM_2.5 concentration.（2）Estimation of hourly resolution PM_2.5 concentration based on deep learning.In this paper,a deep belief network model was introduced to integrate the multi-source sensing data.Experimental tests were conducted to optimize the network parameter settings and the best combination of variables.Taking the central urban area of Wuhan as an application,the quantitative simulation absolute coefficient R²was 0.850,and the validation R² was 0.832.The method was also applied to the estimation and mapping of PM_2.5 concentration in Beijing,Xi’an,and Shenzhen,so as to verify its feasibility and accuracy.The validation R² is 0.95,0.90,and 0.73,respectively;the mapping results can reflect the distribution characteristics and change details of PM_2.5 concentration at different time scales.（3）Exploration of the effects of variables based on statistics and mapping.This paper made statistical analysis to reveal the characteristics of multi-source sensing variables and their correlation with PM_2.5 concentration in the central urban area of Wuhan.Based on the quantitative results and mapping feedback,the effects of social sensing variables and aerosol optical thickness on estimation and cartography were explored.It is shown that dynamic social sensing data can improve the estimation and mapping accuracy.Considering large gaps in the remote sensing data in practical scenarios,the aerosol optical thickness data should be dialectically selected.