Dissertation
Regularized learning techniques for interpretable sonar target recognition systems
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Autumn 2025
Abstract
Active sonar target recognition (ATR) systems utilize backscattering measurements of emitted sound waves to identify and recognize objects in the ocean. These systems extract features encoded in backscattering signals, which result when transmitted sound waves reflect off the target of interest. However, a major challenge lies in accurately recovering target-specific features from the raw backscattering measurements, which are often highly distorted due to numerous oceanic variables. Time varying environmental conditions, such as ocean temperature and speed profiles, significantly affect sonar echo propagation. Additionally, scattering properties depend on the size, geometry and material composition of the target. The complexity is further compounded by environmental factors such as clutter, surface reflections, and inference from marine life. Collectively, these factors make it difficult to model the ocean dynamics accurately and extract the encoded target features.
ATR systems predominantly rely on data-driven methods to estimate target features, which can be divided into two main approaches. The first approach leverages statistical and signal processing techniques to extract features from data, often incorporating expert knowledge through carefully chosen signal transforms and probability distribution assumptions. The second approach uses deep learning techniques, circumventing the need for manually designed feature extraction methods. However, deep learning methods pose challenges for ATR systems. The "black-box" nature of deep neural networks conflicts with the need for interpretable decision systems. Deep neural networks are also often prone to overfitting due to the limited sample sizes of sonar datasets. Therefore, a balanced approach that combines elements from both extremes is desirable.
We propose novel machine learning techniques that bridge the gap between traditional statistical and signal processing techniques and the deep learning approaches. We explore incorporating various signal transforms, such as Fourier and Wavelet transforms, into the learning process to reduce the amount of learning required. Additionally, we incorporate regularization techniques, such as sparsity, into the statistical models to improve generalization. To further enhance the interpretability of the methods, we explore additional model constraints such as group-sparsity and low-rankness, which have physical relevance to the underlying problem. We propose efficient, tractable algorithms for solving these regularized methods, enabling their practical use in ATR systems.
Details
- Title: Subtitle
- Regularized learning techniques for interpretable sonar target recognition systems
- Creators
- Andrew Jonathan Kaltoft Christensen
- Contributors
- Ananya Sen Gupta (Advisor)Matthew Bays (Committee Member)Yang Liu (Committee Member)Tyler Bell (Committee Member)Holavanahalli Udaykumar (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Electrical and Computer Engineering
- Date degree season
- Autumn 2025
- Publisher
- University of Iowa
- Number of pages
- xix, 201 pages
- Copyright
- Copyright 2025 Andrew Jonathan Kaltoft Christensen
- Grant note
Lastly, I thank the Office of Naval Research (Grant Nos. N000142112420 and N000142312503) and Department of Defense Navy (NEEC) (Grant No. N00174201001) for funding this work.
- Language
- English
- Date submitted
- 12/07/2025
- Description illustrations
- illustrations (some color)
- Description bibliographic
- Includes bibliographical references (page 182-201).
- Public Abstract (ETD)
Active sonar target recognition uses short sound pulses to identify objects underwater by analyzing the unique characteristics of their reflected echoes. The difficulty here is that the ocean distorts the echoes as they travel, due to environmental variables (such as temperature) or noise sources (such as marine life). These distortions scramble the acoustic characteristics of a target, making it difficult to determine its identity.
Current methods for identification rely heavily on Artificial Intelligence (AI). While these AI methods are incredibly powerful tools, they suffer from two major flaws. First, their internal workings are incredibly complex, making it difficult to interpret, let alone understand, why they made a particular decision. This black-box nature makes it difficult for engineers and physicists to verify the results of these programs. Second, the AI models typically require vast amounts of data to generalize to new target types, which is difficult given the limited availability of real-world sonar target recognition data.
This work seeks to bridge this gap using innovative machine learning techniques. Our work uses a hybrid approach, combining classical methods from signal processing that incorporate prior knowledge about acoustics, and applying mathematical constraints on the AI models to control their expressivity and ensure their predictions are trustworthy and reliable. This research delivers practical, efficient algorithms that lead to more reliable and interpretable sonar target recognition systems.
- Academic Unit
- Electrical and Computer Engineering
- Record Identifier
- 9985135247202771
Metrics
1 Record Views