On the Detection of High-Latitude Dust Using Deep Learning Methods

Continuous monitoring of Earth's atmosphere is important, in part, because it provides a climatological record of atmospheric phenomena that can be leveraged to improve scientific understanding. Satellite observations facilitate this purpose, especially over remote areas with limited observational record. High latitude dust (HLD), confined to latitudes poleward of 40^oN and 40^oS, is a significant source of polar atmospheric aerosol concentrations and surface deposition. Current methods of dust detection rely on spectral sensitivity at visible and infrared wavelengths. In the former case, those methods levelarge changes in the refractive index at different wavelengths, while in the latter, are based on threshold values of the brightness temperature difference. Considering that HLD processes and optical properties are different than dust in the mid-latitudes and (sub-)tropics, the above methods are not sufficient to account for the variation in the land properties at high latitudes. Therefore, a more robust approach to detect and infer HLD needs to be devised to account for inaccuracies of current methods.

In recent years, advancements in artificial intelligence, and more specifically machine learning, has allowed for the development of trained algorithms that enable detection of atmospheric phenomena. Algorithms are trained using hundreds to thousands of images, which enables statistical results that are not constrained to a specific area. In this initiative, a pixel-based convolutional neural network is used to detect HLD events using daily satellite imagery at the optical wavelength bands. Leveraging the optical properties of dust, true color images (RGB) from the Moderate Resolution Imaging Spectroradiometer (MODIS) have been used as a base to draw polygons over HLD events. The pixel-based neural network is trained using 80% of the data and validated with the rest of 20%. Preliminary results will be presented to evaluate the utility of machine learning for automated HLD detection over traditional methods.