In-Orbit On demand AI Model update for Geo Imaging platforms

In imaging satellites, the value of data erodes with time. The insights must be gathered where the data from the sensors is captured. This has given rise to the edge computing race amongst the space-faring nations known by the term Space-Based Earth Situational Awareness (SBESA). It enables reduced data downlink demands thus enabling faster response to dynamic events on earth. Conventional inference approach methods rely on static AI models which are uploaded prior to launch. This would limit adaptability to the various factors such as evolving environmental conditions, sensor degradation on board, and emerging time-to-time application needs.To address these limitations, mechanisms for updating AI models onboard satellites are essential. Provisioning for in-orbit AI model updates enhances the detection/classification accuracy, operational resilience, and mission flexibility while extending the utility of space assets viz. disaster monitoring, environmental surveillance, security applications etc.Leveraging on the persistent communication link in GEO platforms, in this paper, we present the design architecture for achieving On-demand AI model update onboard the Geo imaging satellite. The design encompasses hardware elements, provisioning the communication link and tailoring the CCSDS communication protocols. Here, we also present one simulation setup to achieve the functionality of model update. The simulation setup comprises the elements viz. UHF Radios with two separate channels for high and low data rate links, NVIDIA®Jetson AGX Orin as GPGPU edge computing board, and a ground based workstation for model updation. Finally, we present the results comprising the required uplink margin for onboard model update, algorithms to execute model update onboard, and time taken to update the AI model weights on the simulation setup.