Machine Learning to Identify and Define Clusters of Aerosols According to Their Optical Properties

Zachary McQueen, Ryan Poland, GEOFFREY SMITH, University of Georgia

     Abstract Number: 144
     Working Group: Instrumentation and Methods

Abstract
It has been shown that aerosols can be classified according to their optical properties, including absorption Ångström exponent (AÅE), scattering Ångström exponent (SÅE), single scattering albedo (SSA) and others. Typical approaches for doing so rely on presumed knowledge of how these optical properties vary for different types of particles and result in strict, sometimes subjective, criteria for separation. Here, we take a different, data-driven approach by using unsupervised machine learning methods to identify natural clusters of observations that have similar optical properties; in essence, we allow the data to define the clusters.

We show how the geographic concept of topographic prominence can be extended beyond two dimensions to identify locations of clusters that vary in density. Then, we describe the use of a density-based clustering algorithm, K-DBSCAN (Kernel-based, Density-Based Spatial Clustering of Applications with Noise; Pla-Sacristán et al., 2019), to assign samples to the clusters without assumption of cluster shapes or boundaries. We also demonstrate how the clustering model helps us to identify the optical properties that contain the most information for clustering and how the model can be adapted if only a subset of optical properties is available. Finally, we build a classification model based on the clusters found allowing new samples to be classified optically.

Reference
Pla-Sacristán, E., González-Díaz, I., Martínez-Cortés, T. & Díaz-de-María, F. Finding landmarks within settled areas using hierarchical density-based clustering and meta-data from publicly available images. Expert Syst. Appl. 123, 315–327 (2019).