Gaussian Process Latent Variable Models Applied to Study Maritime Traffic Patterns from VIIRS Data

Gaussian process latent variable models are used as a data dimensionality reduction technique and applied to analyze long spatio-temporal series of ship traffic patterns measured from data acquired by the Visible Infrared Imaging Radiometer Suite nighttime sensor on board the NOAA-Suomi National Polar-Orbiting Partnership spacecraft. The results show that these techniques are able to model traffic pattern with a number of variables much lower than the number of cells of the time-spatial grid supporting the input data. The use of a Bayesian formulation allows the introduction of spatio-temporal prior constraints that clearly improve the visualization of the time series in the reduced dimensionality space with respect to the classical principal component analysis.