Segregating Maneuvers of Non-Cooperative Space Objects Using Imbalanced Learning and SMOTE

Understanding the behavior of non-cooperative space objects is a critical requirement of space situational awareness. Identifying maneuvers of non-cooperative space objects helps in analyzing their behavior. Since all maneuvers of non-cooperative space objects may not be threatening in nature, it is essential to segregate routine maneuvers needed by a satellite to maintain its orbit from anomalous and abnormal maneuvers that may be perceived as threats. In this paper, we present an approach to filter out benign and regular pattern-of-life maneuvers from the orbital data of non-cooperative space objects through machine-learning-based classification of publicly available orbital data of cooperative civilian satellites with their maneuver history published in the open domain. The main challenge in this task was that the routine pattern-of-life maneuvers of satellites are events of interest but are infrequent, and hence the nonmaneuver class was observed to be far more numerous than the maneuver class label in the dataset. To handle the imbalance in the datasets available for classification, various strategies, like resampling, cost-sensitive learning, one-class learning, and ensemble classifiers, were evaluated and are presented in this paper. The Synthetic Minority Oversampling Techniques (SMOTE) were found to be the most suitable to address the problem at hand and are presented in greater detail. Use cases of different missions of cooperative civilian satellites in the low-Earth-orbit regime are evaluated and presented. The results of the relative comparison between the different SMOTE variants on different satellite datasets are also provided.