Using Machine Learning in Detection of Space Debris

Abstract

In the last decade, the number of space object 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 work, this problem is addressed in two phases. The first phase aims at classifying a detected space object into three categories: satellite, rocket bodies, and debris. In addition, the size of each object is decided. In all cases, features are extracted from light curves for classification using three tools, FATS, feets and UPSILoN Python-based libraries. To achieve classification, a machine learning pipeline is adopted. Starting with preprocessing the raw light curves, to extracting features and classifying these features using several machine learning algorithms. The category-based classification resulted in 70% accuracy, while the size-based classification was much less accurate, around 40s%, indicating that the size cannot be learnt by the classifiers. In the second phase, we reduced the problem into a binary classification problem. That is, we used the same pipeline to classify the space objects into two categories: debris and non-debris. In fact, this is practically more sensible as the presence of debris in the space is usually more concerning due to the damages that it may cause to the space assets. Almost the same feature extraction and machine learning pipeline is adopted with some modification to address the problem of data imbalance and feature-space dimensionality. Similarly, the same objects are also classified based on their size. Results showed an overall accuracy in the range of 80-90% for debris vs non-debris classification. Improvements in the size-based classification are also noted.

Date of Award	Oct 2019
Original language	American English