Can Neural Networks Learn to See Airborne Dust and Sand in Thermal Satellite Imagery?

Aerosols have a major impact on the climate and human health. Today scientists use a plethora of satellite, airborne, and ground-based instruments to measure aerosols in the atmosphere. Satellites provide the broadest coverage and are used widely to generate global aerosol data sets, but radiometers have difficulty separating the contributions of different aerosol types to the total amount. Infrared bands on radiometers such as the VIIRS instruments, which have flown in space since late 2011, are sensitive to coarse elevated aerosols like mineral dust, but not other smaller-sized aerosols. Due to the potential for infrared measurements to help distinguish aerosol types, and the availability of infrared measurements during both day and night, we investigate a dust aerosol optical thickness measurement using thermal infrared bands M14, M15, and M16 from VIIRS. Utilizing simulated VIIRS infrared measurements over a natural range of dust aerosol optical thickness and other atmospheric parameters, we trained a neural network model to quantify dust in the atmosphere. After evaluating the model skill on simulated test dataset, and assessing its sensitivity to factors such as dust amount, height, size, and other atmospheric constituents, we aim to produce a global dust aerosol optical thickness estimate, including uncertainty, from actual VIIRS imagery.