Remote Sensing Inversion Of Fine Particulate Matter And Ozone Based On Spatio-temporal Model

Fine particulate matter(PM_2.5)and ozone are the two major pollutants（collectively referred to as"pollutants"）that affect air quality in China.The application technique of using remote sensing data to obtain parameters related to the target area is called remote sensing inversion.Remote sensing inversion can obtain a wide area and long time series pollutant concentration distribution with reliable accuracy,which is important for air pollution monitoring and management.Near-surface pollutants exhibit spatial and temporal correlations within the local area.Most works design fixed weight modules based on neighboring site values to extract spatio-temporal correlations,but the use of site values reduces the generalization ability of the models and increases the application limits of the models.In addition,most of the studies were conducted based on single pollutants.However,PM_2.5 and O₃ have a clear interaction and share common precursors.In order to solve these two problems,this dissertation proposes the near-surface pollutant inversion method based on spatial and temporal model.The main contents of this dissertation are as follows.（1）To address the deficiency that most of the work in single pollutant estimation based on the fixed weighting of neighboring site values,an estimation method based on 3-D extension of single image metadata set and model adaptive extraction of spatio-temporal correlation is proposed.The method considers the spatio-temporal correlation of pollutants from both data and algorithm levels.That is,the 3-D expansion of the single-image dataset is used instead of the neighboring site values,and the convolutional neural network and the long and short-term memory network are used instead of the fixed weight method based on Gaussian distance,respectively.（2）A joint estimation method based on heterogenous satellite data is proposed to address the shortcoming that PM_2.5 and O₃ are estimated separately as a single target in multi-pollutant estimation.The method is based on the idea of a priori knowledge and multi-tasking,constructing task-specific and shared inputs to jointly estimate PM_2.5 and O₃,and constructing a joint estimation framework based on the spatio-temporal model.The task interaction module and dynamic loss weight module are added in the model structure to collaborate and balance the two subtasks.In this dissertation,3-D extended dataset and multi-source satellite product dataset are constructed for experiments respectively.The results show that the designed spatio-temporal model can improve the estimation performance at the site level,and the generated pollutant concentration distribution maps have unique advantages compared with the mainstream products;the designed joint estimation model can improve the performance based on the single pollutant estimation,and the joint estimation performance is close to the current optimal model.