Classification of Space Objects Using Machine Learning Methods

In the last decade, the number of space objects has skyrocketed. Collecting and analyzing data about these objects is essential in maintaining security of space assets. Classifying unknown objects into satellites, rocket bodies and debris represents a significant milestone in the analysis process. In this context, we investigate the effectiveness of several machine learning methods in classifying real-world light curves of space objects. The light curves are represented with a set of features extracted using the feets (feATURE eXTRACTOR FOR tIME sERIES) public tool. To address the problem of class imbalance, the synthetic minority over-sampling technique (SMOTE) is applied. We also investigate the use of Principal Component Analysis (PCA) in reducing the dimensionality of the feature space, prior to classification. In the case of the original feature set, the top performing classifier is the feedforward neural network with an accuracy of 73.6%. When SMOTE is used, an improvement in accuracy of approximately 15% is observed, with the use of SVM. However, PCA-based feature transformation leads to a slight degradation in performance of around 3%, in the case of the original feature representation, and a considerable degradation of 10%-30%, when SMOTE is used.