Abstract
Atmospheric pollution of fine particulate matter is a major concern for its deleterious effects on human health and climate. Over the past fifty years, Equatorial Asia has experienced significant land use change and urbanization, which have contributed to more intense and frequent extreme particulate matter (PM) concentrations associated with increased anthropogenic and wildfire emissions. Recent advances in remote sensing instrumentation and retrieval protocols have enabled effective monitoring of PM from space in near real time with almost global coverage. In this study, long-term satellite-based observations of key chemical and physical parameters, integrated with ground-based concentrations of PM with aerodynamic diameter < 10 μm (PM10) measured at 52 stations, are used to develop a machine learning approach for continuous PM10 monitoring. As PM atmospheric pollution, like most of environmental processes, is highly non-linear and influenced by numerous variables, machine learning approaches seem very suitable. Herein, Deep Neural Networks are developed and tested over different temporal scales and used to map PM10 over Equatorial Asia during the period 2005-2015. The proposed model captures both PM10 seasonal variability and the occurrence of extreme episodes, which are found to impact air quality on the regional scale. The modeled annual mean fine PM (PM2.5) concentrations are used to estimate long-term premature mortality. This study indicates that the region is experiencing increasing mortality rates related to long-term exposure to PM2.5, with 150,000 (108,000-193,000) premature deaths in 2005 and 204,000 (145,000 – 260,000) in 2015. This is mostly due to air quality worsening and high population growth in urban areas, although the contribution of years of intense wildfires results as well significant.