MACHINE LEARNING METHOD FOR METEORITE CLASSIFICATION BASED ON REFLECTANCE SPECTROSCOPY

Meteorite classification depends on mineralogy, petrology, chemistry, and O-isotopes. Asteroid taxonomy uses spectroscopic characteristics that arise from the first three of these factors. Lab spectra of meteorites in known classes are therefore directly relatable to remote spectra of asteroids with unknown compositions. If a lab spectra-based classification system for meteorites can be developed to successfully replicate traditional methods, then it can be extended to create a new asteroid taxonomy that directly relates meteorites to their parent bodies.

This project analyzed 2401 VIS-NIR meteorite spectra obtained from the Mineral Spectroscopy Lab at Mount Holyoke College, the Reflectance Experiment Laboratory at Brown, and the University of Winnipeg. All meteorites had been classified and reported in the Meteoritical Bulletin Database. All data were resampled from 0.3 to 2.5 µm at 0.5 nm resolution. Data were then pre-processed to ameliorate differences among laboratories, detectors, and experimental procedures using baseline removal, normalization, squashing, and smoothing methods. Optimal combinations of these four tools were determined by a grid search to minimize spectral variation within meteorite classes. Data were then analyzed using an in-house Python tool utilizing the SciKit-learn library and two classification algorithms. Logistic regression is a machine learning technique that uses binary values to predict the probability that an input value belongs to a default class. K-nearest neighbor classification is a non-parametric technique with no assumptions about the underlying data distribution. Initial results suggest that classification of most meteorite groups is possible with >80-90% accuracy based on spectroscopy alone. This suggests that linking meteorite and asteroid spectra using machine learning similarity methods has tremendous potential to lend new insights into our knowledge of parent body distribution and characteristics.