PM_2.5 Remote Sensing Estimation And Prediction Of Emergency Room Visits For Respiratory Diseases Based On Deep Learning

In recent years,air pollution incidents caused by fine particulate matter(PM_2.5)have frequently occurred.The research and analysis on the effective monitoring of PM_2.5 and its health effects have gradually become the focus of attention of governments,scholars and ordinary people.In China,the existing monitoring data for PM2 5 mainly depends on the statistics of the ground monitoring stations.The existing method relying on traditional ground monitoring sites can not fulfill the need of acquisition and updating of PM_2.5.The emergence of satellite remote sensing monitoring is an applicable approach to solve the estimation of the concentration of PM_2.5 even in the full space.Satellite remote sensing monitoring data has obvious advantages such as spatiality and continuity which is of great value to the monitoring of atmospheric environmental changes.On the other hand,in addition to real-time and effective monitoring of PM_2.5 concentration,the harm caused by PM_2.5 to human health also needs special attention.Studies have shown that with the increase of PM_2.5 in the atmosphere,the incidence and mortality of many respiratory-related diseases such as asthma,bronchitis,bronchitis,and chronic obstructive pulmonary disease will increase greatly.Therefore,how to achieve high-efficiency and reasonable estimation of near-surface PM_2.5 concentration based on satellite remote sensing monitoring data,and reasonable analysis and reasoning for the health effects caused by PM2 5 could supply a new way to help government departments formulate relevant measures and reduce the harm of PM_2.5.This paper takes Beijing’s main urban area as a research area to carry out a remote sensing estimation of near-surface PM_2.5 and the forecast of emergency room visits for respiratory diseases in Beijing’s main urban area.The specific research contents and conclusions are as follows:（1）Based on the data of MODIS AOD,the ground-based measured PM_2.5 data and meteorological data,the deep belief network（DBN）model for PM_2.5 remote sensing estimation in Beijing’s main urban area is constructed.In order to improve the experimental efficiency,the model is optimized by the tree structure Parzen estimator（TPE）hyperparameter optimization algorithm.Then,based on the constructed model,the average daily concentration of PM_2.5 near the ground in the main urban area of Beijing in 2013 is estimated.The results show that the models MAE,RMSE,and R are 40μg/m3,57μg/m3,and 0.63,respectively.（2）Based on the data of PM2 5 estimation,meteorological data and emergency room visits data for respiratory disease in Beijing’s main urban area in 2013,quantitative analysis of cumulative lag effect of PM_2.5 on respiratory diseases in urban residents of Beijing is carried out by using the Distributed Lag Nonlinear Model（DLNM）.This study considered the relationship between PM_2.5 exposure-lag response function and the effect of confounding factors on cumulative lag effect analysis.The results show that when the PM2 5 exposure-lag response function is nonlinear,and the confounding factors are adjusted in the model,the model analysis results pass the statistical significance test.With the increase of PM2 5,the relative risk rate（RR）of residents in the main urban area increased by 0.1744（95%CI:1.000286～1.003203）,1.003059（95%Cl:1.000684～1.005439）,1.003936（95%CI:1.000938～1.006942）,1.004357（95%Cl:1.000938～1.009563）,1.004299（95%CI:1.00013～1.008487）and the maximum effect value is obtained within lag 0～3 days.On the contrary,after ignoring the influence of confounding factors in the model under the nonlinear function relationship,the cumulative lag effect analysis results do not pass the statistical significance test;and when the PM_2.5 exposure-lag response function relationship is specified to be linear,regardless of whether the confounding factors are adjusted in the model or not,the analysis results do not pass the statistical significance test.（3）Based on the PM_2.5 estimation data,meteorological data,and hospital emergency room visits data for respiratory diseases in Beijing’s main urban area in 2013,Long short-term memory network（LSTM）is used to forecast the emergency room visits for respiratory diseases in the main urban area.The prediction effects of three forecasting models including long short-term memory network（LSTM）,integrated autoregressive moving average（ARIMA）and multi-layer perceptron（MLP）are compared and analyzed.The results show that the LSTM model show the best perfomance when the time step is 3,and the MAE,MAPE,RMSE,and R of the model are 24,10.87%,31,and 0.87,respectively,which match the cumulative lag effect.The reslut indicates that LSTM can predict and process important events with relatively long latency in time series.LSTM can be used as a new auxiliary means for the analysis of cumulative lag effect.The MLP forecasting model show the best performance in the time step of 1.The MAE,MAPE,RMSE and R of the model are 25,11.01%,32,and 0.85,respectively.MLP cannot process delay events in time series compared with LSTM;ARIMA extrapolated the prediction according to the time series of emergency room visits for respiratory diseases.It is unable to analyze the influencing factors related to respiratory diseases in the model.Compared with other models,the performance of ARIMA is worse.The MAE,MAPE,RMSE,R of the model are 78,21.10%,102 and 0.84,respectively.