Roberto Furfaro, University of Arizona; Richard Linares, Massachusetts Institute of Technology; Vishnu Reddy, University of Arizona
Keywords: Deep learning, Convolutional Neural Networks, Space Object Classification
Abstract:
Developing a detailed understanding of the Space Object (SO) population is a fundamental goal of Space Situational Awareness (SSA). The current SO catalog includes simplified characteristic for the observed space objects, mainly the solar radiation pressure and/or drag ballistic coefficients. Such simplified description limits the dynamic propagation model used for predicting the state of motion of SO to models that assume cannon ball shapes and generic surface properties. The future SO catalog and SSA systems will have to be capable of building a detailed picture of SO characteristics. Traditional measurement sources for SO tracking, such as radar and optical, provide information on SO characteristics. These measurements have been shown to be sensitive to shape, attitude, angular velocity, and surface parameters. State-of-the-art in the literature has been advanced over the past decades and in recent years seen the development of multiple models, nonlinear state estimation, and full Bayesian inversion approaches for SO characterization. The key shortcoming of approaches in literature is their overall computational cost and the limited flexibility to deal with a larger and larger amount of data.
In this paper, we present a data-driven approach to classification of SO based on a deep learning approach that takes advantage of the representational power of deep neural networks as well as on the ability to transfer learned features across deep architectures. Indeed, recent advancements in deep learning have demonstrated ground-breaking results across a number of domains. Deep learning approaches mimic the function of the brain by learning nonlinear hierarchical features from data that build in abstraction. Here, we design, train and validate a Convolutional Neural Network (CNN) capable of learning to classify SOs from collected light-curve measurements. The proposed methodology relies a physically-based model capable of accurately representing RSO reflected light as function of time, size shape and state of motion. The model generates hundreds of thousands of light-curves per selected class of SO and such data are employed to pre-train a deep CNNs that autonomously select the best features proper object discrimination and classification. The model-based CNN is the basis for transfer learning in SSA, where the fundamental features captured in the hidden layers are employed to train a new classifiers where the upper layer is tuned using light curves obtained by real measurements. It is expected that after model-based pre-training , the new CNN can be fine-tuned with a very limited amount of real data, yet preserving the generalization capabilities necessary to accurately discriminate between a large variety of SOs.
Date of Conference: September 11-14, 2018
Track: Non-Resolved Object Characterization