Date of Award

Fall 2025

Project Type

Dissertation

Program or Major

Electrical and Computer Engineering

Degree Name

Doctor of Philosophy

First Advisor

Nicholas Kirsch

Second Advisor

Wayne Smith

Third Advisor

Thomas Blanford

Abstract

Sound is a primary way for many animal species to communicate, forage, navigate, and assess their environment. As a result, passive acoustic monitoring (PAM), the recording and analysis of animal vocalizations, has become a valuable, non-invasive alternative to visual surveys. However, the massive volume of acoustic data generated annually, often reaching hundreds of terabytes, presents a significant challenge for efficient analysis. Since manual inspection of every frame in such data is infeasible, it is essential to develop automated detection and classification tools.

This thesis investigates and develops several statistical signal processing and machine learning techniques to address this challenge. The first contribution is a novel classification system for distinguishing between two acoustically similar dolphin species—Risso’s dolphins (Grampus griseus) and Pacific white-sided dolphins (Lagenorhynchus obliquidens). A Bayesian classification framework based on variational mode decomposition (VMD) was developed to extract key spectral features from pulsed vocalizations. This method, combined with a visualization tool called the VMD-gram, achieved over 80\% classification accuracy even with low signal-to-noise ratio recordings, demonstrating the method's robustness and practical utility in sparse acoustic datasets.

The second contribution introduces a novel weighted spectral entropy (WSE) technique for detecting several types of marine mammal tonal calls. An adaptive band-pass filter and a Bayesian classification-based thresholding mechanism were integrated into a complete automatic detection system. The WSE method improved the traditional spectral entropy technique by incorporating background noise estimation, significantly enhancing robustness in noisy environments. The detector was evaluated using both synthetic and real-world acoustic datasets and was shown to outperform existing state-of-the-art methods in precision, recall, and noise resilience.

Subsequent work in this thesis focuses on detecting and analyzing vocalizations from other species. Statistical signal processing techniques and machine learning methods were used to detect and analyze vocalizations from American bullfrogs and gray bats. In the bullfrog study, the WSE method was used to detect advertisement calls from American bullfrogs, enabling precise temporal localization. This was followed by the application of an unsupervised machine learning method, spectral clustering, to estimate the number of vocal bullfrogs by grouping calls from them. With each group representing an individual bullfrog, the number of vocal bullfrogs at a monitoring site can be estimated. For gray bats, acoustic indices such as Acoustic Complexity Index (ACI), spectral entropy (SE), and power spectrum were extracted from echolocation calls to estimate their abundance.

Collectively, this thesis contributes to the development of automated detection and classification tools for passive acoustic monitoring, enhancing our ability to study biodiversity and assess ecological impacts in a scalable and efficient manner. The studies demonstrate the versatility and effectiveness of combining statistical signal processing techniques with machine learning techniques for acoustic analysis across a range of species.

Download

Share

COinS