A008-0028
A Deep Learning Approach for Surface PM2.5 Estimations from Geostationary Satellite and Numerical Model Data
Monday, 7 December 2020
Poster
Manisha Khatri1, Muthukumaran Ramasubramanian1, Iksha Gurung1, Aaron S Kaulfus1, George Priftis2, Peiyang Cheng3, Pawan Gupta4, Manil Maskey1, Rahul Ramachandran5, Sundar A Christopher6 and Haeyong Chung1, (1)University of Alabama in Huntsville, Huntsville, AL, United States, (2)University of Alabama in Huntsville, Atmospheric Science, Huntsville, AL, United States, (3)University of Alabama in Hunstville, Huntsville, United States, (4)Universities Space Research Association Greenbelt, Greenbelt, MD, United States, (5)NASA Marshall Space Flight Center, Huntsville, AL, United States, (6)University of Alabama in Huntsville, Atmospheric and Earth Science, Huntsville, AL, United States
Abstract:
Fine particles released from activities such as vehicle exhausts, burning of fuels, forest fires are known to cause severe impacts on public health. Particulate matter (PM) with diameter less or equal to 2.5 μm, known as PM2.5, is known to cause or exacerbate cardiovascular and respiratory illnesses. The Environmental Protection Agency (EPA) monitors the levels of particulate matter using air quality monitors stationed throughout the Continental United States (CONUS). The measurements indicate concentrations from point locations, which may not always be available and may not represent the concentrations of surrounding areas. To overcome these limitations in air quality monitoring, Aerosol Optical Depth (AOD) products at high spatiotemporal resolution from geostationary satellites may be used to estimate the levels of surface PM2.5. However, additional meteorological factors that nonlinearly affect surface PM2.5 concentrations such as relative humidity, temperature, height of the planetary boundary layer must be considered. This information is accessible from numerical modeling such as from the National Oceanic and Atmospheric Administration’s (NOAA) High Resolution Rapid Refresh (HRRR) model which resolves near real-time atmospheric conditions over the CONUS.
The analysis of the data from these different sources and derivation of the relationship among them to estimate surface PM2.5 can be a tedious task and may require significant computation and processing. Deep learning approaches are appropriate for such complex estimation problems as it defines relations among multiple non-linear parameters. We propose a deep learning neural network model that would utilize the AOD retrievals from Geostationary Operational Environmental Satellite (GOES) 16 and factor in the data from the other aforementioned sources to estimate the PM2.5 levels. The presentation will illustrate the development of the neural network model and its ability to provide estimations of surface PM2.5 over the CONUS based on the high spatial temporal AOD retrievals from GOES 16.
