Goal-Oriented Vessel Detection with Distributed Computing in a LEO Satellite Constellation

Earth Observation (EO) has traditionally involved the transmission of a large volume of raw data to map the Earth surface. This results in congestion to the satellite network and delays in the availability of the results, invalidating the approach for timing-sensitive applications. Instead, the computation resources at the satellites can be used as an edge layer for compressing the data and/or doing inferences. In this paper, we investigate satellite edge computing for vessel detection with a LEO satellite constellation. First, we distribute the computation and inference load among the neighbouring satellites of the one taking the images, based on the VHRShips data set and YOLOv8. This semantic and fragmented information is then routed to a remote ground monitor through the whole constellation. The average and peak Age of Information (AoI) are reformulated to measure the freshness of the aggregated information at the receiver in this image-capture scenario. We then dimension the network - number of orbital planes and satellites per orbital plane - for a given target age and covered area that quantify the level of achievement of the task. The results show that 20 orbital planes with 20 satellites are necessary to keep the peak AoI below 60 s with a compression ratio$>23000$, i.e., a size reduction of 99.996%, and for a$\sim100\%$probability of coverage.