Machine-Learned Data Compression for Outer Solar System Missions

It is easy to take for granted our ability to draw scientific information from distant interplanetary spacecraft, particularly those in outer solar system missions. The Cassini Imaging Science Subsystem (ISS), for example. has returned a wealth of high-resolution images of Saturn and its moons, enabling detailed scientific retrievals of anomalies in Saturn’s rings and the atmosphere of Titan. But in reality, all data received from an outer solar system mission must overcome severe constraints on a spacecraft’s available power, antenna size, and bandwidth for transmitting data back to Earth. Given the enormous cost of these missions, the pressure is high on scientists to have not only a high return of data, but a pristine quality thereof. As such, many science teams have an aversion to employing lossy compression techniques, despite their promise to allow for a substantially higher data return, since many are known to introduce artifacts into data. Such was the case for the Cassini ISS, whose hardware-implemented JPEG compressor introduced large blocking artifacts in camera images, necessitating the need for overly conservative compression. It was demonstrated in [1] that off-the-shelf compressors could be fine-tuned to effectively cater to specific data sets without compromising their scientific integrity. In this paper, we take this a step further by incorporating machine learning to automate the selection of the compressor and its parameters. In particular, we use neural networks to generalize the compression models which were found to perform well in the previous study, and incorporate aspects of the scientific retrievals into both the training process and the structure of the networks. We show how, in some cases, machine learning can supplement the previous algorithms, and demonstrate the performance on detached haze retrievals of Titan from Cassini images. We then present a novel neural network model and algorithm which implements machine-learned transform coding, a complete generalization of the algorithm from [1]. We demonstrate its performance on two classes of data from the Cassini mission: 1) tie-point matching in flyby sequences of Tethys, and 2) measuring propeller structures in the rings of Saturn. We also discuss modifications of our original algorithm to incorporate machine-learned estimation of optical flow between sequential images, as well as the use of discriminator networks trained adversarially with our compression/decompression networks to distinguish compression error from noise. Our system is data-driven, achieving higher compressibility when trained on more particular datasets, and allows for trading off between incorporating prior knowledge of the ambient structure and noise in the data versus allowing the networks to learn this for themselves. Furthermore, our algorithms are portable, and we demonstrate high performance from lightweight networks, which can be quickly trained and easily uploaded.