Asteroid Spectra and Machine Learning

Abstract

The aim of this thesis is to explore applications of machine learning to the study of asteroid spectra, and as such, its research question can be summarized as: How can asteroid spectra be analyzed using machine learning? The question is explored through evaluation of the obtained solutions to two tasks: the optimal locations of spectrophotometric filters for asteroid classification success and the formation of an asteroid taxonomy through unsupervised clustering.

First, background theory for asteroids and particularly spectroscopy of asteroids is presented. Next, the theory of machine learning is briefly discussed, including a focus on the method utilized to solve the first task: neural networks. The first task is executed by developing an optimization algorithm that has access to a neural network that can determine the classification success rate of data samples that would be obtained using spectrophotometric filters at specific locations within the possible wavelength range. The second task, on the other hand, is evaluated through determining the optimal number of clusters for the given dataset and then developing taxonomies with the clustering algorithm k-means.

The obtained results for the first task involving the optimal locations of filters for spectrophotometry seem reliable, and correlate relatively well with well-known mineralogical features on asteroid surfaces. The taxonomic systems developed by the unsupervised clustering also succeeded rather well, as many of the formed clusters seem to be meaningful and follow the trends in other asteroid taxonomies. Therefore, it seems that based on the two investigated tasks, machine learning can be applied well to asteroid spectroscopy. For future studies, larger datasets would be required for improving the overall reliability of the results.