Stochastic Models to Generate Sound Speed Ensembles and Calculate Transmission Loss Ensembles at the New England Shelf Break Using Empirical Orthogonal Functions and Sparse Dictionary Learning
Open Access
- Author:
- Benit, Simeon
- Graduate Program:
- Acoustics (MS)
- Degree:
- Master of Science
- Document Type:
- Master Thesis
- Date of Defense:
- April 17, 2023
- Committee Members:
- Andrew Barnard, Program Head/Chair
Ying-Tsong Lin, Committee Member
Daniel C. Brown, Committee Member
James Miller, Special Signatory
Andrew Barnard, Thesis Advisor/Co-Advisor
Daniel C. Brown, Thesis Advisor/Co-Advisor - Keywords:
- Stochastic Model
Empirical Orthogonal Functions
Machine Learning
Acoustics
Underwater Acoustics
Principal Component Analysis
PCA
EOF
Dictionary Learning
Sparse Dictionary Learning
K-SVD
Orthogonal Matching Pursuit
Sound Speed Profiles
Range-dependent Acoustic Model
Transmission Loss
New England Shelf Break - Abstract:
- The New England Shelf Break’s (NESB) bathymetry and proximity to the Gulf Stream current make it one of the most complex ocean environments in the world. This makes it a valuable test case for algorithms designed to recognize patterns in large datasets. Empirical Orthogonal Function (EOF) analysis, or Principal Component Analysis (PCA), has historically been the standard statistical method for detecting spatiotemporal patterns in data. These patterns are used to reconstruct datasets accurately and efficiently. Recently, sparse dictionary learning methods like the K-SVD have begun to supplant EOF analysis for this application. Results from these techniques are valuable to underwater acoustic propagation models that require realistic environmental data input to produce useful approximations of the underwater soundscape. Using World Ocean Database (WOD) in situ Conductivity-Temperature-Depth (CTD) measurements taken at the NESB, two stochastic models were developed: one based on the K-SVD, the other based on EOF analysis. Additionally, a new approach to representing sparse coefficients from the K-SVD as a normal random variable was developed. Both models were used to produce sound speed profile (SSP) ensembles. These SSP ensembles were used to calculate Transmission Loss (TL) ensembles from the Range-dependent Acoustic Model (RAM). A probability density function (PDF) of TL ensembles was produced to show the relative likelihood that a TL ensemble would fall within a certain range of TL. The EOF stochastic model produced SSP ensembles that reliably covered SSP spatial variability at the shelf. While the EOF model did not cover variability at the shelf break, the resulting ensemble envelope was consistent. The K-SVD ensemble envelope was nondeterministic and did not reliably cover SSP spatial variability at the shelf or the shelf break. The EOF stochastic model required that missing values be filled with the depth-wise mean to function, while the K-SVD flexibly handled the imputation of zeros to compensate for missing values. TL ensembles were successfully calculated using K-SVD SSP ensembles. Ultimately, EOF analysis showed more potential as the basis for a stochastic model to inform an underwater acoustic propagation model.