Please join us for the following AOS Seminar on Thursday, May 5 at 4:00PM in Spiess Hall Room 330:
Michael Bianco, Scripps Institute of Oceanography
"Dictionary learning-based techniques for the inversion of geophysical data"
Ocean acoustic and seismic signals contain substantial environmental information from the ocean and the interior of the earth. However, the estimation of physical models from these data, a.k.a. inversion, is challenging due to limited measurements and the ill-conditioned nature of such problems. In this talk, sparse signal processing and machine learning approaches to the inverse problem are considered. It is shown that dictionary learning, a class of methods for automatically adapting basis functions to given data for sparse representation, provides a flexible means of incorporating prior information into the inverse solution. Recently, dictionary learning has been used to construct dictionaries of basis function describing ocean sound speed profiles (SSPs) from thermistor string data. Each of the resulting functions in the dictionary are informative of SSP variability, driven for example by tides and internal waves. It is shown that this property enables a significant reduction in the size of a parameter search, which may improve the fidelity of SSP estimates from acoustic data. Further, an adaptive approach is developed for 2D seismic travel time tomography, based on dictionary learning and the concept of "patch-level" sparsity from image processing.