Stochastic analysis of exit fluid temperature records from the active TAG hydrothermal mound (Mid-Atlantic Ridge, 26°N) : 2. Hidden Markov models of flow episodes

Abstract

I develop a stochastic signal model for episodic modes of variability in hydrothermal flow records using probabilistic functions of Markov processes (i.e., hidden Markov models, HMMs) and fit the model to exit fluid temperature time series data from diffuse flow sites on the active TAG hydrothermal mound. The flow states are modeled using Gamma densities to provide flexibility for application to a range of signal types. Between three and five flow states are needed to fit the diffuse flow temperature records from TAG, which correspond to models with between 10 and 28 degrees of freedom. The number of flow states required to fit a given record is related to the signal variance, with more variable records requiring a larger state space. HMMs thus provide an efficient signal model for episodic variability in hydrothermal flow records, suggesting that Markov processes may provide a means to generate stochastic subsurface flow models for deep-sea hydrothermal fields if the spatial flow correlations can be incorporated into a statistical framework. I also use the Viterbi algorithm to “decode” the time series data into best fitting state sequences, which can be used to classify the records into discrete flow episodes. This may provide an objective means to identify discrete events in a flow record if misclassification issues arising from nonepisodic variability (e.g., tidal forcing) can be addressed.