Research On Atmospheric PM_2.5 Neural Network Remote Sensing Retrieval Considering Spatiotemporal Characteristics

Atmospheric PM_2.5 pollution is a major environmental problem around the world.Towards the monitoring of atmospheric PM_2.5,fusing satellite remote sensing and ground station observation is a promising approach.Considering the complex non-linear relationship between atmospheric PM_2.5,satellite observation and other auxiliary variables,machine learning-based remote sensing retrieval methods of atmospheric PM_2.5 have attracted wide attention.However,the existing machine learning methods usually establish the numeric relationship between PM_2.5 and the influencing factors,with few consideration of the spatial and temporal characteristics of atmospheric PM_2.5.To this end,based on the neural network model and considering the spatial and temporal characteristics of atmospheric PM_2.5,this paper aims to develop atmospheric PM_2.5remote sensing retrieval method combining geoscience laws and learning models,so as to achieve the accurate and robust retrieval of atmospheric PM_2.5.The research contains the following four main aspects:(1)Deep learning-based atmospheric PM_2.5 retrieval method considering spatiotemporal autocorrelation of PM_2.5.Due to the relatively simple structure of the existing atmospheric PM_2.5 machine learning retrieval model,the nonlinear relationship is not sufficiently represented;besides,the spatiotemporal autocorrelation of atmospheric PM_2.5 is seldom considered.To this end,the novel deep learning technology is introduced to explore the complex nonlinear relationship between atmospheric PM_2.5 and influencing factors.Meanwhile,the spatiotemporal autocorrelation factors of atmospheric PM_2.5 are extracted by using the inverse distance weighting function and incorporated into the deep learning framework.On the basis,this paper proposes a deep learning-based atmospheric PM_2.5 retrieval method considering spatiotemporal autocorrelation of PM_2.5,and realizes the joint processing of spatiotemporal autocorrelation and nonlinear characteristics.The satellite aerosol optical depth(AOD)based atmospheric PM_2.5 retrieval experiment across China shows that the model estimates are highly consistent with the ground station observations.Furthermore,the model is extended to omit the AOD retrieval process and directly estimate atmospheric PM_2.5 from satellite reflectance.The results indicate that the model can effectively estimate atmospheric PM_2.5 directly from satellite reflectance.(2)Atmospheric PM_2.5 geographically and temporally weighted neural network retrieval method.The existing methods are incapable to simultaneously take into account the nonlinear characteristics and spatiotemporal heterogeneity of the AOD-PM_2.5 relationship,which often leads to poor local estimates.For this reason,combining with the classical spatiotemporal regression modeling strategy and non-linear neural networks,the geographically and temporally weighted neural network retrieval method is proposed.To establish the nonlinear relationship between PM_2.5 concentration and influencing factors,the model adopts the generalized regression neural network(GRNN)that has a strong nonlinear fitting ability.In view of the spatiotemporal heterogeneity of this relationship,the model is respectively set up for each location and time;meanwhile,the Gaussian weighting function of the spatiotemporal distance is introduced to describe the importance of samples,which is then incorporated into the GRNN network model,so as to solve geographically and temporally weighted neural network.The results of the national experiment show that this model has a great advantage over the classical spatiotemporally weighted regression model.(3)Atmospheric PM_2.5 global-local space-time neural network retrieval method.The neural networks considering the spatiotemporal heterogeneity are usually trained based on the limited samples in a local context,which easily leads to the instability of the model and unstable estimates.For this reason,the atmospheric PM_2.5 global-local space-time neural network(STNNF)retrieval method is proposed.First,the global neural network(NN)model is constructed using all samples from the whole study region and period to learn the overall effect of input variables on ground PM_2.5.Secondly,for the spatiotemporal heterogeneity of the AOD-PM_2.5 relationship,geographically and temporally weighted NN(GTWNN)models are established with the global NN as the initial condition,and the GTWNN models are fine-tuned by using spatiotemporally local samples.The national experiment shows that this model reports a great advantage over the classical spatiotemporally weighted regression models and performs better than the geographically and temporally weighted GRNN model in(2).(4)Remote sensing monitoring of atmospheric PM_2.5 across Wuhan 1+8 urban agglomeration.Wuhan 1+8 urban agglomeration is one of the pilot areas of"resource-saving and environment-friendly society"in China,and the dynamic monitoring of atmospheric environment in this area is of great significance.Taking Wuhan 1+8 urban agglomeration as an experimental area,this paper introduces the new generation of geostationary satellite(i.e.,Himawari-8 of Japan)and takes advantage of its high temporal resolution,then develops Wuhan 1+8 urban agglomeration atmospheric PM_2.5remote sensing real-time retrieval technical system,using deep learning-based atmospheric PM_2.5 retrieval method that considers spatiotemporal autocorrelation of PM_2.5 and ground station observation and other auxiliary data.The technical system realizes the automatic processing from data access to data production,and automatically generates PM_2.5 concentration data per hour;this technical system provides basic data and technical support for the fine-scale monitoring of atmospheric PM_2.5 in urban agglomeration.