Predicting Time-Varying Flow Fields Using Sparse Observations with Self-Organizing Maps

In this paper we propose a solution to the problem of predicting ocean currents across a large region of interest using only a small set of discrete point observations of an ocean current made by an AUV or team of AUVs and historical context provided by historical model output. We use a self-organizing map (SOM) to fit a low-dimensional manifold to the historical model output. This SOM then acts as a state vector in a recursive Bayesian filter, which is updated based on in-situ observations of ocean currents made by one or more AUVs. We compare our method with two Gaussian process estimators using a 2 year dataset of model output from around Ram Head on St. John in the U.S. Virgin Islands.