Evaluation of Oversampling Strategies in Machine Learning for Space Debris Detection

In recent years, the number of resident space objects has increased dramatically. The chances of space objects colliding with each other are increasing, thus posing a threat to active satellites and future space missions. Identifying and detecting space debris is essential in ensuring the security of space assets. In this contribution, we investigate the effectiveness of several feature extraction and oversampling techniques by attempting classification of real-world light curves of space objects using eight machine learning methods. Three feature extraction tools are utilized to represent the light curves as sets of features, i.e., FATS (Feature Analysis for Time Series), feets (feATURE eXTRACTOR FOR tIME sERIES) and UPSILoN (AUtomated Classification for Periodic Variable Stars using MachIne LearNing) public tools. To address the problem of class imbalance, four oversampling techniques are applied, i.e., ADaptive SYNthetic Sampling Approach (ADASYN), Synthetic Minority Oversampling TEchnique (SMOTE), and two modifications of SMOTE, specifically, Borderline-SMOTE and Support Vector Machine (SVM)-SMOTE. Results show that the features extracted using the FATS tool lead to a better performance, and therefore, they appear to represent light curves in a more informative manner, compared to feets and UPSILoN. Moreover, the use of SVM-SMOTE technique improves the performance of the utilized classifiers more than other oversampling techniques.