Cognitive Satellite Sensor Controller (CS2C)

As the complexity of what we want to track grows, we increasingly need an end-to-end solution where we task satellites with what to track and they pass down tracks. This transition to an end-to-solution is analogous to the path taken by machine learning in both computer vision and natural language processing. However, satellites are a uniquely challenging platform to employ end-to-end learning due to their power and computation constraints. Also, it is difficult to get enough training data for machine learning algorithms. To address these issues to bring the benefits of end-to-end learning to satellite sensor systems, DeUmbra proposes Cognitive Satellite Sensor Controller (CS2C). The ōcognitiveö in CS2C is because CS2C learns a forward model of how changes in the acquisition parameters change image the image. CS2C acquires this model using self-supervised learning. This self-supervised learning also enables CS2C to build synthetic training examples, which are used both for the newness of the domain and for the skewed nature of the data. Approved for Public Release | 23-MDA-11635 (9 Nov 23)