Times are displayed in (UTC-05:00) Central Time (US & Canada) Change

Using Machine Learning to Analyze the Reflectance Spectra of Asteroids

Session: Advancing Mineralogy and Spectroscopy Across the Solar System in Honor of MSA Roebling Medalist M. Darby Dyar

Presenting Author:

Thomas Burbine

Authors:

Burbine, Thomas H.¹, Saha, Sulagna², Parida, Tanish³, Patel, Nayan⁴

(1) Department of Physics & Astronomy, Mount Holyoke College, South Hadley, MA, USA, (2) Department of Computer Science, Mount Holyoke College, South Hadley, MA, USA, (3) Acton-Boxborough Regional High School, Acton, MA, USA, (4) Duke University, Durham, NC, USA,

Abstract:

One of the pioneers in using machine learning to analyze planetary spectra has been M. Darby Dyar. Darby typically has used machine learning to analyze Raman spectra. However, Darby has also applied machine learning to analyze the visible and near-infrared reflectance spectra of meteorites and asteroids. Building on her work, some students and I have applied a variety of machine learning techniques to classify asteroid spectra. One study used a Convolutional Neural Network (CNN) to classify asteroid spectra according to the Bus-DeMeo taxonomy. After training on a random set of main-belt, near-Earth, and Mars-crosser spectra, the CNN was able to achieve accuracies of over 80% on a testing dataset of asteroid spectra. In another study, we used machine learning to classify meteorites by their reflectance spectra with the ultimate goal of mineralogically classifying asteroid spectra. We initially preprocessed the meteorite spectra using ROCKET (Random Convolutional Kernel Transform). ROCKET generates new numerical summaries (called features) that capture local and global patterns within the spectra. We then applied a ridge classier to these feature vectors and were able to produce accuracies of over 90% on a testing set of spectra for twenty-two meteorite classes. Future work will apply this technique to asteroid spectra.