Once-for-All Sub-Network Selection Methodology for Rapid Orbital Deployment

Demand for orbital image data is increasing at a pace much faster than down-link capacity to move this data to ground stations for processing. Space Edge Computing for DL-based analysis of orbital image data (categorization/change detection) offers a promising solution. Dynamic deployment of DL models to space edge computing devices is desirable but significantly constrained by uplink bottlenecks, hardware limitations and power budgets. This paper proposes a selection methodology for dynamic deployment of DL models for an on-orbit context. Making use of the Once-for-all (OFA) framework, our proposed solution considers required ML accuracy performance, upload availability and hardware limitations for time-critical, earth observation scenarios with the goal of reducing the number of orbital periods required. Groundstation aware orbital simulations are performed to determine maximum transmission size for a given time window to determines the maximum network size. This combined with space edge computing hardware limitations are used as input for a suitable OFA sub-network. Extensive experimental evaluation using space, ground scenarios that consider orbit position are used to evaluate the efficacy of the proposed selection methodology.