A practical classifier for biomimetic underwater covert acoustic signals under unobserved modulation conditions using genetic-algorithm feature optimization

Biomimetic underwater acoustic communication disguises information-bearing signals as marine-mammal vocalizations, creating a challenging classification problem for underwater acoustic monitoring. This study evaluates classifier performance using pre-segmented whistle clips. It aims to overcome the dependence of conventional classifiers on specific modulation methods and parameters. This limitation means that these classifiers cannot guarantee performance for modulation conditions not covered during the fitting process. We propose a genetic-algorithm (GA)-based classifier-training framework that constructs proxy-unseen conditions from an available biomimetic dataset. Each chromosome jointly encodes the modulation–parameter conditions used for classifier fitting, the signal-feature subset, and the classifier configuration. The objective function combines mean and minimum area under the curve (AUC), condition-dependent AUC dispersion, and sparsity terms, thereby discouraging configurations that are specialized to particular fitting conditions or use unnecessary features and samples. The dataset for the framework was constructed using five whistle-mimicking modulation methods, nine parameter configurations per method, 11 SNR levels, and two fixed multipath channel profiles. Training with the proposed GA objective function effectively reduced localized performance degradation. Furthermore, this process ensured that the GA-selected features served as the primary driver of the overall performance gain. In the numerical ROC comparison, the final classifier achieved an inter-method AUC range of 0.0167, compared with 0.0784, 0.1176, and 0.0808 for three method-specific classifiers. The fixed classifier was then evaluated using sea-trial recordings obtained over a 1.5-km distance between Tx and Rx. Its modulation-wise AUCs varied from 0.8366 to 0.9067, yielding a range of 0.0701, whereas the largest range among the conventional classifiers was 0.3387. These results show that explicitly optimizing proxy-unseen performance produces a compact classifier that is less dependent on the evaluated modulation conditions and retains its relative advantage in clip-level validation under the controlled sea-trial conditions.

Authors

Institutions

Publication Details

Journal
Applied Ocean Research
Published
2026-11-18
DOI
https://doi.org/10.1016/j.apor.2026.105355
Primary Topic
Underwater Vehicles and Communication Systems
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

A practical classifier for biomimetic underwater covert acoustic signals under unobserved modulation conditions using genetic-algorithm feature optimization

Geun‐Ho Park, Hyung-Moon Kim, Donghun Lee, Wan‐Jin Kim et al.
1 links
Applied Ocean Research
Underwater Vehicles and Communication Systems
article

A practical classifier for biomimetic underwater covert acoustic signals under unobserved modulation conditions using genetic-algorithm feature optimization

Geun‐Ho Park, Hyung-Moon Kim, Donghun Lee, Wan‐Jin Kim, Jongmin Ahn, Sanguk Lee
article en

Abstract

Biomimetic underwater acoustic communication disguises information-bearing signals as marine-mammal vocalizations, creating a challenging classification problem for underwater acoustic monitoring. This study evaluates classifier performance using pre-segmented whistle clips. It aims to overcome the dependence of conventional classifiers on specific modulation methods and parameters. This limitation means that these classifiers cannot guarantee performance for modulation conditions not covered during the fitting process. We propose a genetic-algorithm (GA)-based classifier-training framework that constructs proxy-unseen conditions from an available biomimetic dataset. Each chromosome jointly encodes the modulation–parameter conditions used for classifier fitting, the signal-feature subset, and the classifier configuration. The objective function combines mean and minimum area under the curve (AUC), condition-dependent AUC dispersion, and sparsity terms, thereby discouraging configurations that are specialized to particular fitting conditions or use unnecessary features and samples. The dataset for the framework was constructed using five whistle-mimicking modulation methods, nine parameter configurations per method, 11 SNR levels, and two fixed multipath channel profiles. Training with the proposed GA objective function effectively reduced localized performance degradation. Furthermore, this process ensured that the GA-selected features served as the primary driver of the overall performance gain. In the numerical ROC comparison, the final classifier achieved an inter-method AUC range of 0.0167, compared with 0.0784, 0.1176, and 0.0808 for three method-specific classifiers. The fixed classifier was then evaluated using sea-trial recordings obtained over a 1.5-km distance between Tx and Rx. Its modulation-wise AUCs varied from 0.8366 to 0.9067, yielding a range of 0.0701, whereas the largest range among the conventional classifiers was 0.3387. These results show that explicitly optimizing proxy-unseen performance produces a compact classifier that is less dependent on the evaluated modulation conditions and retains its relative advantage in clip-level validation under the controlled sea-trial conditions.

Applied Ocean ResearchVol. 177
Agency for Defense Development (KR)
Agency for Defense Development, Defense Acquisition Program Administration
Underwater Vehicles and Communication Systems
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon