Reliable ML-Based Image Processing and Compression for an Onboard Imaging Pipeline

Sensing and onboard-processing capabilities of next-generation spacecraft continue to evolve. Enabled by advances in avionic systems, large amounts of data can be collected and stored on orbit. Nevertheless, loss of signal, communication delays, and limited downlink rates remain a bottleneck for delivering data to ground stations or between satellites. This research investigates a multistage image-processing pipeline and demonstrates rapid collection, detection, and transmission of data using the Space Test Program - Houston 7 - Configurable and Autonomous Sensor Processing Research (STP-H7-CASPR) experiment aboard the International Space Station as a case study. Machine-learning (ML) models are leveraged to perform intelligent processing and compression of data prior to downlink to maximize available bandwidth. Moreover, to ensure accuracy and preserve data integrity, a fault-tolerant ML framework is employed to increase pipeline reliability. This pipeline fuses the fault-tolerant Resilient TensorFlow framework with ML-based tile classification and the CNNJPEG compression algorithm. This research shows that the imaging pipeline is able to alleviate the impact of limited communication bandwidth by using reliable, autonomous data processing and compression techniques to achieve reduced transfer sizes of essential data. The results highlight the benefits provided by resilient classification and compression including minimized storage use and reduced downlink time. Furthermore, the findings of this research are used to assess the feasibility of such a system for future space missions. The combination of these approaches enables the system to achieve up to a 98.67 % reduction in data size and downlink time as well as the capacity to capture imagery up to $75.19\times$ longer time period for a given storage size, respectively, while maintaining reconstruction quality and data intearity.