Representing Underwater Sound Speed Profiles via Parameterized Curve-Fitting and Model Selection

The underwater sound speed measured as a function of depth, known as the sound speed profile (SSP), provides tremendous insight into acoustic transmission properties in a geographic region. With the recent increase in SSP dataset size and fidelity, there is a growing need to accurately represent large datasets with a small number of parameters so that SSPs can be transmitted effectively across low bandwidth underwater communication channels. Here we present a novel parametric curve-fitting model with the following properties: (1) it achieves a high goodness of fit to SSP data; (2) it requires few model parameters; and (3) it preserves important physically meaningful features while reducing the appearance of non-physical artifacts. By designing our model as the weighted sum of skewed and non-skewed Gaussian distribution functions and employing a model selection procedure to prevent overfitting, we accomplish the above goals. We show that for a particular dataset, SSPs can almost always be reconstructed with a maximum mean absolute error less than 1 m/s while achieving a data compression rate that equates to greater than 90% bandwidth savings.