Please join us for the following AOS Seminar on Thursday, June 2 at 4:00PM in Spiess Hall Room 330:
Speaker: Mike Bianco, Scripps Institution of Oceanography
Title: Dictionary Learning of Ocean Acoustic Sound Speed Profiles
Acoustic sound speed profiles (SSPs) in the ocean are often highly variable, with fine-scale fluctuations. Remote estimation of SSPs using geoacoustic inversion is an ill-posed problem which requires significant regularization to obtain physically plausible solutions. Traditional regularization methods, which minimize the energy of best-fit solutions, require undersampling of true SSPs or forcing the SSPs to conform to few, leading order empirical orthogonal functions (EOFs). These methods often give low resolution SSP estimates and can affect the accuracy of other parameters in geoacoustic inversion. High resolution estimates can be obtained using sparse processing (e.g. compressive sensing or matching pursuit) provided a collection or 'dictionary' of shape functions can be designed which express much of the SSP signal variability. Using Dictionary Learning methods, it is shown that a dictionary of shape functions can be 'learned' from a set of SSP observations and that only a few of these shape functions are needed to represent much of the observed variability. It is further shown that these learned dictionaries outperform EOFs for SSP signal reconstruction.