Advancing air quality monitoring through multi-satellite observations and artificial intelligence

Abstract

This thesis proposes an artificial intelligence (AI)-based framework to advance satellite-based air quality monitoring through the integration of multi-satellite observations. The framework is designed to improve both the retrieval of satellite-derived air quality products and the estimation of surface-level pollutant concentrations from observations that primarily provide column-integrated and radiance-based information. As no single satellite sensor can simultaneously achieve high spatial, temporal, and spectral resolution, satellite-based air quality monitoring inevitably involves trade-offs among these dimensions, motivating the integration of complementary satellite observations. To address these challenges, this thesis is organized into six chapters that progressively develop and evaluate machine learning and deep learning approaches across multiple satellite platforms. Chapter 1 introduces the research background, motivation, and overall structure of the dissertation. Chapter 2 examines the estimation of surface-level nitrogen dioxide (NO2) and ozone (O3) concentrations from polar-orbiting satellite column densities through machine learning, demonstrating the feasibility of data-driven approaches to capture nonlinear column-to-surface relationships. Chapter 3 focuses on the retrieval of aerosol optical depth (AOD) from hyperspectral observations of the Geostationary Environment Monitoring Spectrometer (GEMS) using deep learning, improving the accuracy and robustness of satellite-derived aerosol products for air quality applications. Chapter 4 extends the framework by directly estimating hourly PM2.5 concentrations from top-of-atmosphere reflectance measured by Geostationary Ocean Color Imager (GOCI-I and GOCI-II), enabling high-temporal-resolution surface monitoring while reducing reliance on intermediate physics-based AOD retrievals. Chapter 5 develops a deep learning-based multi-satellite fusion framework that integrates the hyperspectral radiance of GEMS with the finer spatial resolution of GOCI-II observations. Using a self-supervised learning strategy, this approach generates super-resolved hyperspectral radiance at sub-kilometer scales and examines its implications for subsequent air quality retrievals, addressing the inherent spatial–spectral trade-offs of geostationary satellite observations. Chapter 6 provides an overall summary of the dissertation and discusses directions for future research. This dissertation brings together multiple methodological components—satellite retrieval, surface-level estimation, and data fusion—into a unified framework for air quality monitoring. The results highlight the role of AI in improving the spatial and temporal representation of satellite-derived air quality information.