Space object classification using deep neural networks

Space object classification is desired for space situational awareness to be able to discern resident space object (RSO) characteristics, behaviors, and perspective changes. Due to the limited sensing resources and observations, it is challenging for space object classification to be responsive to unfolding and unexpected events. Many machine learning algorithms are already used to classify space objects based on various sensor observations from radar and telescope. In this paper, the use of deep neural networks (DNN) is proposed to classify space objects due to DNN robust performance in many classification tasks, such as face recognition and object recognition. This paper explores DNN using light curve data. Conventional classification algorithms, such as k nearest neighbor (k-NN), are implemented and compared to the proposed DNN based classification algorithms, including the popular convolutional neural network (CNN) and the recurrent neural network (RNN), in terms of accuracy. Inherent advantages and disadvantages of the deep neural network based classification algorithms are summarized and the potential for future space object classification tasks is analyzed and postulated.

1. Introduction

Space situational awareness includes monitoring of the space environment and resident space objects [1]. One of the critical tasks is to classify space objects according to their properties. Unfortunately, the information of space objects available is often limited. Typically, the visual magnitude and the radar cross section (RCS) of space objects can be obtained via optical and radar sensor, respectively. The light curve data, which includes a sequence of visual magnitude, and the RCS hence provide a way to classify space objects.