Predicting Asteroid Types: Importance of Individual and Combined Features

Asteroid taxonomies provide a link to surface composition and mineralogy of those objects, although that connection is not fully unique. Currently, one of the most commonly used asteroid taxonomies is that of Bus-DeMeo. The spectral range covering 0.45-2.45 mu m is used to assign a taxonomic class in that scheme. Such observations are only available for a few hundreds of asteroids (out of over one million). On the other hand, a growing amount of space and ground-based surveys delivers multi-filter photometry, which is often used in predicting asteroid types. Those surveys are typically dedicated to studying other astronomical objects, and thus not optimized for asteroid taxonomic classifications. The goal of this study was to quantify the importance and performance of different asteroid spectral features, parameterizations, and methods in predicting the asteroid types. Furthermore, we aimed to identify the key spectral features that can be used to optimize future surveys toward asteroid characterization. Those broad surveys typically are restricted to a few bands; therefore, selecting those that best link them to asteroid taxonomy is crucial in light of maximizing the science output for solar system studies. First, we verified that with the increased number of asteroid spectra, the Bus-DeMeo procedure to create taxonomy still produces the same overall scheme. Second, we confirmed that machine learning methods such as naive Bayes, support vector machine (SVM), gradient boosting, and multilayer networks can reproduce that taxonomic classification at a high rate of over 81% balanced accuracy for types and 93% for complexes. We found that multilayer perceptron with three layers of 32 neurons and stochastic gradient descent solver, batch size of 32, and adaptive learning performed the best in the classification task. Furthermore, the top five features (spectral slope and reflectance at 1.05, 0.9, 0.65, and 1.1 mu m) are enough to obtain a balanced accuracy of 93% for the prediction of complexes and six features (spectral slope and reflectance at 1.4, 1.05, 0.9, 0.95, and 0.65 mu m) to obtain 81% balanced accuracy for taxonomic types. Thus, to optimize future surveys toward asteroid classification, we recommend using filters that cover those features.