Space Analytics Engine for On-Board Machine Learning and Multimodal Data Fusion

Current ISR satellites typically play a passive data collection and dissemination role, with processing typically done further downstream at ground stations. This “offline” processing approach introduces significant delays in converting data to actionable insights, prevents low-latency coordination between end-users (e.g., field operatives) and space-based assets, and precludes more intelligent sensing capabilities; for example, adaptive tasking of the sensor suite based on real-time data enhancement and analytics to improve intelligence gathering.

The project builds upon existing work at BAE Systems and the partner academic institutions (UoA and UNSW) on machine learning for ISR applications. The project will develop novel algorithms and workflows to enable machine learning on nanosatellites for space based ISR from multi-modal sensors. Note that the capacity of current edge computing hardware (e.g., Nvidia Jetson series) is still significantly smaller than standard hardware, thus necessitating algorithms for model pruning and data pre-processing to perform on-board machine learning. Operating in space also presents unique obstacles to updating the pre-trained on-board models, in terms of procuring data and supervisory labels for retraining, and bandwidth constraints in updating models. The project will tackle the above challenges to produce a novel space analytics engine that is reconfigurable after launch, which significantly increases the value proposition of on-board processing.


Project Leader:
Professor Tat-Jun Chin, University of Adelaide