View in EDS

Physics-driven environmental augmentation for underwater acoustic target recognition.

Saved in:
Bibliographic Details
Title: Physics-driven environmental augmentation for underwater acoustic target recognition.
Authors: Lu, Yiyang1 (AUTHOR) luyy59@mail2.sysu.edu.cn, Xu, Lingji1,2,3 (AUTHOR) xulj26@mail.sysu.edu.cn, Li, Zhenglin1,2,3 (AUTHOR) lizhlin29@mail.sysu.edu.cn
Source: Signal Processing. Jul2026, Vol. 244, pN.PAG-N.PAG. 1p.
Subjects: Data augmentation, Acoustic models, Computer simulation, Spectrograms, Machine learning, Generalization
Abstract: • A physics-driven environmental augmentation framework is proposed for integrating underwater acoustic broadband modeling with real-world oceanographic data to synthesize acoustically authentic training data across diverse underwater environments. • An MTR-MoE architecture is introduced that was specifically designed for underwater acoustic spectrograms. Mamba, Transformer, and ResNet experts were integrated through dynamic gating mechanisms for comprehensive feature extraction. • It was demonstrated that PEA significantly outperformed traditional augmentation methods and GANs, while multi-environment training achieved superior cross-domain generalization, revealing environmental diversity as the primary determinant of generalization performance. Underwater acoustic target recognition (UATR) faces the fundamental challenges of data scarcity and cross-environment generalization, and these challenges limit practical deployment. Existing augmentation methods apply environment-agnostic transformations that fail to capture the physics-governed acoustic propagation variability across diverse underwater environments. This paper proposes a physics-driven environmental augmentation (PEA) framework that synthesizes training data by integrating underwater acoustic broadband modeling with realistic waveguide environment modeling across shallow, transitional, and deep water regimes. The framework explicitly models multipath interference and frequency-dependent absorption through physics-based impulse responses that are derived from real-world oceanographic data. We introduced a Mamba-Transformer-ResNet Mixture of Experts (MTR-MoE) architecture that integrated three specialized experts through dynamic gating mechanisms for comprehensive feature extraction across sequential, global, and local relationships within multi-channel spectrograms. Experimental validation on the ShipsEar dataset demonstrated that PEA combined with MTR-MoE achieved 89.74% accuracy, outperforming MTR-MoE trained on original data by 15.59 percentage points. Cross-environment generalization experiments revealed that multi-environment training achieved higher accuracy across unseen locations, significantly outperforming single-environment specialists. These results established that physics-driven environmental modeling provided a generalizable paradigm for data augmentation in physics-governed recognition tasks, offering an important methodological reference for practical UATR deployment. [ABSTRACT FROM AUTHOR]
Copyright of Signal Processing is the property of Elsevier B.V. and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:• A physics-driven environmental augmentation framework is proposed for integrating underwater acoustic broadband modeling with real-world oceanographic data to synthesize acoustically authentic training data across diverse underwater environments. • An MTR-MoE architecture is introduced that was specifically designed for underwater acoustic spectrograms. Mamba, Transformer, and ResNet experts were integrated through dynamic gating mechanisms for comprehensive feature extraction. • It was demonstrated that PEA significantly outperformed traditional augmentation methods and GANs, while multi-environment training achieved superior cross-domain generalization, revealing environmental diversity as the primary determinant of generalization performance. Underwater acoustic target recognition (UATR) faces the fundamental challenges of data scarcity and cross-environment generalization, and these challenges limit practical deployment. Existing augmentation methods apply environment-agnostic transformations that fail to capture the physics-governed acoustic propagation variability across diverse underwater environments. This paper proposes a physics-driven environmental augmentation (PEA) framework that synthesizes training data by integrating underwater acoustic broadband modeling with realistic waveguide environment modeling across shallow, transitional, and deep water regimes. The framework explicitly models multipath interference and frequency-dependent absorption through physics-based impulse responses that are derived from real-world oceanographic data. We introduced a Mamba-Transformer-ResNet Mixture of Experts (MTR-MoE) architecture that integrated three specialized experts through dynamic gating mechanisms for comprehensive feature extraction across sequential, global, and local relationships within multi-channel spectrograms. Experimental validation on the ShipsEar dataset demonstrated that PEA combined with MTR-MoE achieved 89.74% accuracy, outperforming MTR-MoE trained on original data by 15.59 percentage points. Cross-environment generalization experiments revealed that multi-environment training achieved higher accuracy across unseen locations, significantly outperforming single-environment specialists. These results established that physics-driven environmental modeling provided a generalizable paradigm for data augmentation in physics-governed recognition tasks, offering an important methodological reference for practical UATR deployment. [ABSTRACT FROM AUTHOR]
ISSN:01651684
DOI:10.1016/j.sigpro.2026.110545