Seismology and ocean acoustics are important remote sensing tools, enabling observation of environments that are difficult to access and directly measure. Seismic and ocean acoustic remote sensing are data-intensive tasks, and the proliferation of remote sensing systems has led to the generation of vast amounts of data. Meanwhile, advances in machine learning (ML) techniques and computational capacity have yielded state-of-the-art methodologies for processing and analyzing large seismic and acoustic data sets. This dissertation presents two ML-based paradigms for the characterization of environments using seismic and acoustic data.
First, unsupervised ML is demonstrated for automatically identifying dominant types of seismicity present data recorded from a 34-station broadband seismic array deployed on the Ross Ice Shelf (RIS), Antarctica from 2014 to 2017. The data set contains signals generated by glaciological processes that have been used to monitor the integrity and dynamics of ice shelves. Deep clustering automatically groups these signals into classes without the need for manual labeling, enabling comparison of potential source mechanisms with not only the spatial and temporal distributions of the signals but also their characteristics. The method learns the salient features of spectrograms and encodes them into a lower-dimensional latent representation using an autoencoder, a type of deep neural network. Two clustering methods are applied to the latent data and compared: a Gaussian mixture model (GMM) and deep-embedded clustering (DEC). Dominant types of seismic signals are identified and compared with environmental data such as temperature, wind speed, tides, and sea ice concentration. The highest seismicity occurred at the RIS front during the 2016 El Niño summer, and diurnally near grounding zones throughout the deployment.
The second paradigm presents Bayesian optimization (BO) as a method for efficiently estimating geoacoustic parameters within a fixed computational budget. An objective function is defined using the Bartlett processor, whose output measures the match between a received and predicted pressure field on a vertical line array. BO is a sequential framework that iteratively fits a Gaussian process surrogate model to the objective function and then uses a heuristic acquisition function to select the next point to evaluate. After each evaluation, the GP surrogate model is re-fit, and the optimization proceeds until the budget is expended. BO is demonstrated using both simulations and real data collected during an ocean acoustics experiment. Results indicate BO rapidly estimates the correct parameters and achieves better correlations between observed and predicted data.