GSA Connects 2021 in Portland, Oregon

Paper No. 43-4
Presentation Time: 2:20 PM

MACHINE LEARNING APPROACHES TO ASTEROID TAXONOMY USING METEORITE SPECTRA


WALLACE, Sydney, Planetary Science Institute, 1700 East Fort Lowell, Suite 106, Tuscon, AZ 85719-2395, DYAR, M. Darby, Planetary Science Institute, 1700 East Fort Lowell, Suite 106, Tucson, AZ 85719-2395, SHELDON, Daniel, College of Information and Computer Sciences, University of Massachusetts Amherst, Amherst, MA 01003 and BURBINE, Thomas H., Dept. of Astronomy, Mount Holyoke College, South Hadley, MA 01075

Currently, asteroids and meteorites are classified under separate taxonomies with no direct correspondence. Meteorites are classified based on mineralogy and texture, while asteroids are classified based on the shape of their spectra into 26 main classes defined by the Bus-DeMeo (B-DM) taxonomy. An ideal taxonomy would link meteorite spectra to their asteroid parent bodies. This study produces a new quantitative taxonomy for both meteorites and asteroids by combining mineralogy, spectroscopy, and machine learning (ML).

For this study, we acquired roughly 350 meteorites and prepared them as powders for spectral acquisition at the Reflectance Experiment Laboratory (RELAB). We then merged this data set with existing data on meteorite powders, slabs, and chips from RELAB and the University of Winnipeg. All spectra were resampled to cover the range from 0.3 to 2.5 µm at 0.01 µm resolution; data are posted on nemo.mtholyoke.edu. ML models were trained on 1,422 RELAB spectra spanning 21 meteorite classes as given in the Meteoritical Bulletin Database. Because similarities in mineralogy make classes with different textures spectrally indistinguishable, we created nine distinct classes for training: CM/C2/CR, CO/CV, CK/R/brachinites, L/H/LL/ureilites, EH/EL, Acapulcoites/lodranites, IAB/IIAB, aubrites, and howardites/eucrites/diogenites. Cross-validation was used for hyperparameter tuning and model selection. We tested the accuracy of classification models with as-acquired data versus data normalized to the intensity at a single channel. Then we chose the best model by testing all channels in the wavelength range at 0.05 μm increments. Normalization is needed to improve model generalization performance when classifying spectra from varying laboratories and telescopes. ML models including Kernel Fisher Discriminant Analysis (KFDA), Kernel Support Vector Machine (SVM), Logistic Regression, and Quadratic Discriminant Analysis (QDA) achieved test classification accuracies ranging between 85-93% within the RELAB dataset. Models were tested on the meteorite data from the University of Winnipeg with comparable accuracies. The best model was then used to classify asteroids into the nine new classes to better understand their compositional and spatial diversity and their relationship to meteorites.