A multivariate functional-data mixture model for spatio-temporal data: inference and cokriging
In this paper, we introduce a model for multivariate, spatio-temporal functional data. Specifically, this work proposes a mixture model that is used to perform spatio-temporal prediction (cokriging) when both the response and the additional covariates are functional data. The estimation of such models in the context of expansive data poses many methodological and computational challenges. We propose a new Monte Carlo Expectation Maximization algorithm based on importance sampling to estimate model parameters. We validate our methodology using simulation studies and provide a comparison to previously proposed functional clustering methodologies. To tackle computational challenges, we describe a variety of advances that enable application on large spatio-temporal datasets. The methodology is applied on Argo oceanographic data in the Southern Ocean to predict oxygen concentration, which is critical to ocean biodiversity and reflects fundamental aspects of the carbon cycle. Our model and implementation comprehensively provide oxygen predictions and their estimated uncertainty as well as recover established oceanographic fronts.
READ FULL TEXT