Improved Monitoring and Nowcasting of Particulate Matter based on Machine Learning through the Synergistic Use of Satellite and Model Products over East Asia

Abstract

This thesis proposes as comprehensive view of satellite-based particulate matter (PM) monitoring and nowcasting using deep learning over East Asia. This thesis seeks to 1) monitor satellite-derived PM10 and PM2.5 (PM with aerodynamic diameters less than 10 and 2.5 micrometers (μm), respectively) concentrations using historical datasets under all-sky condition, 2) monitor satellite-derived PM10 and PM2.5 concentrations using real-time learning (RTL) based on machine learning, 3) nowcast PM10 and PM2.5 concentration spatiotemporally using deep learning (i.e., U-net) using gap-filled PM concentrations. This dissertation consists of following six chapters. Chapter 1 summarizes the harmful of PM10 and PM2.5 concentrations on human health and the environment and overview of the thesis research. In Chapter 2, two random forest sub-models were used to estimate spatially continuous PM concentrations using historical datasets. The seamless AOD was modeled under two different conditions whether satellite-retrieved AOD exists or not. Spatially continuous PM10 and PM2.5 concentrations were then modeled using the AOD model output. However, the historical dataset is not enough to estimate current PM concentrations because they were not reflected the current meteorological and aerosol conditions. In Chapter 3, ground-level hourly PM10 and PM2.5 concentrations were estimated using satellite-derived aerosol products through the random forest (RF)-based RTL approach. Unlike previous studies that used accumulated historical datasets (i.e., offline-models), this study improved the model performance through hourly updated modeling (RTL-models). However, there is a limitation depicting a spatially continuous distribution of RTL-based PM concentrations since satellite-derived aerosol products contain missing values due to bright surface or cloud contamination. In Chapter 4, there are the two steps to predict future PM10 and PM2.5 concentrations. Firstly, simple gap-filling of hourly PM10 and PM2.5 concentrations estimated by RTL-models was conducted through hybrid-machine learning and regression kriging (Hybrid-RFK) and Gaussian filtering to solve the missing value problems. In second step, the hourly gap-filled PM10 and PM2.5 concentrations were calculated as daily averages and used for prediction. The spatiotemporal nowcasting was conducted using the spatially full-covered PM concentrations by deep learning (i.e., U-net). Chapter 5 provides a summary of this dissertation. The future works of this study were discussed in Chapter 6.