• About
  • Advanced Search
  • Browse Proceedings
  • Access Policy
  • Sponsor
  • Contact
  • Gallery Index
  • Gaussian process for modelling the space environment from sparce data collected by massively distributed femtosatellite networks

    Paper ID

    87926

    DOI

    10.52202/078365-0159

    author

    • Christopher Teale
    • Pietro Colombo
    • Stephen Jun Villejo
    • Colin R. McInnes

    company

    School of Engineering, University of Glasgow; University of Glasgow

    country

    United Kingdom

    year

    2024

    abstract

    Massively distributed in-situ parallel sensing is an emerging femtosatellite mission architecture that can be deployed in low Earth orbit to collect environmental data with unprecedented spatial coverage, for any given instant in time. In previous work it was shown that this mission architecture can deconvolve time and space as independent variables in orbits that are densely populated by the femtosatellites, offering a unique opportunity for modelling highly localised and rapidly evolving changes in the space environment. However, such events are frequently associated with random interactions between the steady state orbital environment and external sources, such as the solar wind. Additionally, the data acquired is sparce in comparison to the spatial length-scales involved. This presents a challenge for traditional interpolation methods. Namely, the probabilistic behaviour of the observed information is hard to track using deterministic methods, i.e., linear interpolation, which might fail in characterizing the local behaviour at lower resolutions. Furthermore, such methods usually are unable to provide an evaluation of the quality of predictions made in unsampled regions. Gaussian process regression offers the end user an alternative for mapping environmental variables to unobserved locations for any given time. They not only accommodate variables with different degrees of spatial-temporal correlation, but also provide uncertainty estimates about predicted values. In this paper, the magnetic environment on low Earth orbit was reconstructed by means of Gaussian processes using data generated from a simulated femtosatellite mission, deployed in a massively distributed sensor network. The model was calibrated by selecting an appropriate kernel function, doing a Maximum Likelihood hyperparamater tuning, and performing cross validation in space and time. The potential extensions arising from this work are multiple, such as accounting for non-stationary isotropy, and considering multiple output scenarios, thus highlighting the relevance of femtosatellite networks as a powerful tool for space environment research.

    STATUS: waiting for script files to load