Expectation-maximization analysis of spatial time series

Abstract

Expectation maximization (EM) is used to estimate the parameters of a Gaussian Mixture Model for spatial time series data. The method is presented as an alternative and complement to Empirical Orthogonal Function (EOF) analysis. The resulting weights, associating time points with component distributions, are used to distinguish physical regimes. The method is applied to equatorial Pacific sea surface temperature data from the TAO/TRITON mooring time series. Effectively, the EM algorithm partitions the time series into El Niño, La Niña and normal conditions. The EM method leads to a clearer interpretation of the variability associated with each regime than the basic EOF analysis.