Speaker
Description
Recent advances in Earth observation, embedded computing and machine learning optimisation have led to an increasing number of demonstrations of “on-board AI”, showcasing such applications as cloud detection,ship tracking and disaster monitoring. However, there still exists a disconnect between the outputs of such applications and the real needs of end users.
In collaboration with Surrey Satellite Technology Ltd (SSTL) and the University of Surrey (UoS) under ESA InCubed funding, Craft Prospect is consolidating its software solutions for on-board data processing andinformation extraction into a modular and rapidly-configurable framework – the Astral Intelligence Toolbox (AITB). The AITB employs a component-based design that allows new applications to be configured quickly from customer use cases and requirements. A cloudmasking component can easily be re-configured for land cover classification tasks. Pre-processing tasks can be integrated into the application to handle radiometric calibration, band co-registration and georeferencing as required. Traditional and machine learning-powereddata products can be generated, harmonised and compressed to meet customer requirements on content, accuracy, latency and throughput. Each component is configured such its properties and behaviours work in concert with others to deliver a solution that meetsthe mission needs.
This approach also improves the ability to assure the performance and functionality of configured applications. Components are verified and assured on the unit level, allowing application testing to focus onintegration and end-to-end testing. Machine learning processes developed with the Assuring Autonomy International Programme ensure that the models powering information extraction tasks are developed with a focus on explainability, interpretability and robustness.
Under InCubed funding with SSTL and UoS, three applications – defined through customer engagement and configured using the AITB – will be deployed and demonstrated on SSTL’s Flexible & Intelligence Payload Chain(FIPC). The modular nature of the AITB also makes it portable; applications can be deployed on embedded Linux, SpaceCloud and other platforms.
In this talk, the philosophy, driving requirements and design of the AITB will be presented. Preliminary results from the InCubed activity will be presented and discussed in the context of customer use cases. Finally, a roadmap will bepresented on the AITB’s continued development to target further use cases and mission-critical on-board activities.
