1 Image segmentation for dust detection using unsupervised machine learning CyberTraining 2020: Big Data + High-Performance Computing + Atmospheric Sciences Julie Bessac 1 , Ling Xu 2 , Manzhu Yu 3 Faculty mentor: Aryya Gangopadhyay 4 ; External mentor: Yingxi Shi 5 ; Research Assistant: Pei Guo 4 1 Mathematics and Computer Science Division, Argonne National Laboratory; 2 Department of Mathematics, North Carolina A&T State University; 3 Department of Geography, Pennsylvania State University; 4 Department of Information Systems, University of Maryland Baltimore County; 5 NASA GSFC Climate and Radiation Laboratory Technical Report HPCF–2020–17, hpcf.umbc.edu > Publications Abstract Dust and sandstorms originating from Earth’s major arid and semi-arid desert areas can significantly affect the climate system and health. Many existing methods use heuristic rules to classify on a pixel-level regarding dust or dust-free. However, these heuristic rules are limited in applicability when the study area or the study period has changed. Based on a multi- sensor collocation dataset, we sought to utilize unsupervised machine learning techniques to detect and segment dust in multispectral satellite imagery. In this report, we describe the datasets used, discuss our methodology, and provide preliminary validation results. 1 Introduction Dust events are common meteorological phenomena in arid and semi-arid regions, often arising when strong winds uplift fine-grained dust particles from the surface of the Earth. Atmospheric dust plays a positive role in absorbing light radiation and the formation of clouds (Prospero, 1999). On the other hand, dust storms are usually damaging. Due to climate change, the dynamics of dust storms at a local scale have changed drastically along with climate and weather variables, such as total precipitation and average wind speed (Middleton, 2019). Frequencies and intensities of local dust storms are observed to be increasing, bringing higher impacts on wildlife, human beings, and bio-community (Taylor et al., 2017). A highly accurate and efficient method for dust detection is desired, which can predict the occurrence and intensity of dust events from high-quality dust observations in a timely fashion, and at the same time mitigate the adverse effects of dust storms. One such method of dust detection is to use various satellites and their observations. For example, the satellites that range from polar to geostationary orbiting, and the various spectral bands ranging from thermal infrared to LIDAR. It is noted that the polar-orbiting satellites generally have a high spatial resolution but limited temporal resolution. Examples include the Visible Infrared Imaging Radiometer Suite (VIIRS) on the Suomi National Polar-orbiting Partnership (Suomi NPP) satellite, the Ozone Mapping Profiling Suite (OMPS) on Suomi NPP, and Cloud‐Aerosol Lidar with Orthogonal Polarization (CALIOP) onboard the Cloud‐Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO).