Graduate Thesis Or Dissertation
 

Automatic Classification of Ultrasonic Harbor Porpoise Clicks in a Varying Noise Environment

Public Deposited

Downloadable Content

Download PDF
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/9k41zn95b

Descriptions

Attribute NameValues
Creator
Abstract
  • This study compares three approaches in the design of an autonomous machine listening agent that predicts harbor porpoise ultrasonic echolocation clicks in diverse noise environments. Considering the temporal variations of noisy coastal ocean soundscapes which the harbor porpoises inhabit, we propose a leave-one-day-out (LODO) cross-validation strategy in the training of a random forest classifier that successfully addressed the covariate shift present in our time-series data. To evaluate the efficacy of our approach to capture signals in this noisy environment, we compare three preprocessing approaches and two deep learning architectures on our harbor porpoise click data. We find that feature extraction strategies of mel frequency cepstral coefficients (MFCC) and short time Fourier transform (STFT) outperformed our novel approach, the heterodyned-Teager-Kaiser Energy Operator (TKEO), which shifts down the ultrasonic signal to a lower frequency in the time domain. Building on these results, we seek to improve the robustness of our porpoise click classifier for a real-world environment by implementing a second-stage stacked random forest ensemble on combinations of subsets of 42 deep learning base models that were trained from the folds of our LODO cross-validation and the three preprocessing approaches that were explored in this study. Our results demonstrate that experiments using the LODO cross-validation strategy reported a difference between the average fold accuracy and a held-out test accuracy of 6%, while training without cross-validation and the equal k-fold cross-validation reported a 28.7% and 30.4% difference, respectively. From the three preprocessing approaches we implement, the models trained on MFCC produced the highest accuracy of 95.6% on the held-out test set while those trained on STFT and heterodyned-TKEO produced accuracies on the same held-out test set of 88.7% and 85.0%, respectively. Results from our stacked random forest show the greatest improvement in accuracy of 5.6% in the heterodyned-TKEO models while the STFT and MFCC models reported 4.5% and 1.9% improvements in accuracy, respectively. Highly varying noise environments are common across coastal areas inhabited by harbor porpoises. This study, with our proposed ensemble of different feature and model architectures, emphasizes the need to overcome such shifts in noise to design a robust porpoise click classifier that is ready for real-time deployment and able to generalize to all real-world conditions.
License
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Academic Affiliation
Rights Statement
Publisher
Peer Reviewed
Language

Relationships

Parents:

This work has no parents.

In Collection:

Items

Thumbnail Title Date Uploaded Visibility Actions
file details DeleonDaniel2022.pdf 2022-06-11 Public Download