Date of Award

January 2019

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Space Studies

First Advisor

Michael J. Gaffey

Abstract

Taxonomic studies of asteroids have been ongoing for more than fifty years without a clear understanding of the class parameters. The current method of Principal Component Analysis is computationally expensive and leaves ambiguous results. In this study, I selected the machine learning algorithm, k-Nearest Neighbor in combination with the current Bus-DeMeo (DeMeo, et al. 2009) taxonomic classification schema to test if machine learning can take the place of Principal Component Analysis. Using a dataset of spectrophotometric color indices derived from combined visible and near-infrared (NIR) observations and paired with Bus-DeMeo taxonomic class, I created a training dataset for the model to learn. The results support the visible wavelength region as more diagnostic of spectral slope and the NIR wavelength region as more diagnostic for surface mineralogy. The overall accuracy scores (>80%) of the machine learning test dataset validate the methodology, but fall short of the threshold necessary to replace current methods of classification (>95%). The overall robustness of the Bus-DeMeo taxonomy is corroborated through the relatively similar grouping structure between the C-, S-, and X-complexes in both wavelength ranges, suggesting an overall relationship between slope and qualities present across multiple wavelength regimes. This is possibly due to spectral features being closely tied to surface mineralogy and spectral reddening of the slope believed to be tied to the effects of space weathering.

Share

COinS