Machine Learning using Sparse Feature Representations for Non-Resolved Space-Based Space Sensing

Numerica is proposing a data-driven approach using modern state-of-the-art machine learning methods on sparse feature representations of non-resolved photometric data, for space object characterization, classification, and anomaly/change detection in support of threat identification and notification (TIN). In order to support the future needs of the SSA mission, including the maintenance of a growing catalog of RSO characterizations and classifications, improved algorithms capable of processing non-resolved optical data from new and existing optical sensors such as the SST, SBSS, and GEODSS will be required to more effectively characterize, classify, and identify such objects. To address these objectives, Numerica proposes to develop a modern machine-learning-based algorithmic suite by providing (i) algorithms for extracting sparse feature representations of photometric data; (ii) supervised learning algorithms for classification; (iii) unsupervised learning algorithms for anomaly/change detection and TIN; and (iv) model and feature selection algorithms. Numerica will also leverage its experience processing real and simulated data from the SSN and the SST.