04
2024-03-04
Results Brief | Advances in Acoustic Survey of Marine Mammals Using Unmanned Surface Vehicles
Summary:
China has rich marine mammal biodiversity, including various species of whales, dolphins, seals, and sea lions, with a total of 39 recorded species. The rational development and utilization of marine mammal biological resources is of great significance for promoting sustainable socio-economic development, strengthening ecological environmental protection, accelerating the construction of a “maritime power,” and enhancing international discourse in ocean governance.
In response to the application prospects and practical requirements of marine mammal acoustic surveys, the Unmanned Surface Vehicle research team of Ocean University of China, in collaboration with HydrovoX, initiated key research on acoustic intelligent identification technologies in early 2021. A dedicated marine mammal acoustic payload for Unmanned Surface Vehicles was developed, enabling underwater acoustic target online identification, remote control, real-time communication, and self-contained data storage.
The core hardware circuit of the acoustic payload consists of three main modules: a posture control module, an underwater acoustic acquisition module, and a control and signal processing circuit module. The architecture of the core hardware circuit is shown in Figure 1.

Figure 1. Core hardware circuit of the acoustic payload.
A Raspberry Pi Compute Module 4 is selected as the core unit of the control and signal processing module to deploy control and signal processing software. Python-based multi-process and multi-thread programming is employed to enable concurrent execution and resource sharing, thereby improving operating system efficiency and resource utilization.
Multi-process programming provides strong fault tolerance, as each process operates independently; failures in a single process or subprocess do not affect the stability of the main program. Threads within the same process share the same address space, enabling efficient data sharing and communication between threads.
Leveraging next-generation artificial intelligence technologies represented by deep learning, the Unmanned Surface Vehicle system is comprehensively empowered to achieve disruptive technological advancement, forming an innovative paradigm of “AI for Unmanned Surface Vehicles” in marine observation and acoustic sensing applications. Accordingly, the research team has developed a lightweight, highly robust, and strongly generalizable deep neural network model for multi-scenario marine mammal acoustic survey tasks.
Marine mammal acoustic data typically exhibit a long-tail or heavy-tailed distribution. While abundant data are available for common species with large populations and wide distribution ranges, data for low-density and spatially limited species remain severely insufficient, with even no available training samples for some endangered species. This data imbalance and few-shot learning problem represent key challenges in acoustic intelligent recognition. For highly diverse marine ecosystems, hierarchical multi-label classification of large numbers of marine mammal species is also a critical issue. In addition, underwater acoustic signals are inherently characterized by attenuation, spatiotemporal variability, and low signal-to-noise ratio, making marine mammal acoustic recognition a highly challenging task.
To address these challenges, the research team adopts different modeling strategies and categorizes marine mammal acoustic survey tasks into distinct deep learning paradigms suitable for different application scenarios, mainly including Underwater Acoustic Target Recognition and Acoustic Anomaly Detection.
(1) Underwater Acoustic Target Recognition
Underwater acoustic target recognition models marine mammal acoustic targets and analyzes large-scale acoustic datasets. This approach is mainly applicable to scenarios where sufficient labeled acoustic data are available and the species composition in a fixed sea area is known. A deep neural network is constructed for underwater acoustic target recognition.
The Unmanned Surface Vehicle research team designs an innovative recognition framework based on deep convolutional neural networks and attention mechanisms. Large-scale pre-training is performed using non-underwater acoustic datasets such as AudioSet, followed by fine-tuning on underwater acoustic datasets through transfer learning to enhance robustness and generalization ability.
In addition, model compression techniques such as pruning, quantization, and knowledge distillation are adopted to achieve lightweight network design. These methods significantly reduce model parameters with minimal or no loss in accuracy, making the model suitable for deployment on resource-constrained mobile marine platforms such as Unmanned Surface Vehicles. The overall workflow is illustrated in Figure 2.

(2) Acoustic Anomaly Detection
Acoustic anomaly detection models ocean ambient noise and performs data analysis under different sea states and meteorological conditions. It is primarily designed for scenarios where acoustic data of rare species are missing or where marine mammal categories are unknown. This approach addresses the limitation of conventional underwater acoustic target recognition methods, which often suffer from significant performance degradation when encountering unseen classes during inference. It is therefore of great importance for detecting previously unobserved marine mammal species.
The Unmanned Surface Vehicle research team proposes a deep learning algorithm based on a Deep Autoencoder, trained using an unsupervised learning approach for underwater acoustic signal anomaly detection. The overall model architecture is shown in Figure 3. The framework consists of an Encoder and a Decoder. The Encoder is responsible for feature extraction.
The Decoder adopts a dual-branch structure, consisting of a reconstruction branch and a prediction branch. The reconstruction branch performs low-dimensional acoustic representation reconstruction, while the prediction branch forecasts the current frame based on previous acoustic frames. When the reconstructed signal and the predicted signal show high similarity, the resulting confidence is high, indicating a normal acoustic event.
Since the training dataset contains only ocean ambient noise (i.e., normal acoustic events), the model may fail to properly reconstruct or predict unseen acoustic events during testing. In such cases, a significant discrepancy between reconstruction and prediction results occurs. If the discrepancy exceeds a predefined threshold, the event is classified as an acoustic anomaly, potentially corresponding to marine mammal acoustic targets or other types of underwater sources.
The corresponding acoustic information is then transmitted back to the shore-based control system via satellite communication, where further analysis and validation are conducted by operators.

Figure 3. Framework of the acoustic anomaly detection model.
The research team has conducted dozens of routine sea trial campaigns in the South China Sea and the Western Pacific Ocean. Taking the marine mammal survey conducted from June to November 2023 in the South China Sea as an example, selected self-contained marine mammal acoustic datasets were analyzed using LOFAR spectrograms and Welch power spectral density (PSD) estimation, as shown in Figure 4.
The spectral characteristics at different time periods reflect the entire process of marine mammals approaching and then moving away from the Unmanned Surface Vehicle. It can be observed that the acoustic spectrograms of marine mammals exhibit distinct harmonic structures, primarily concentrated in the mid-to-high frequency range, and display a clear “V-shaped” pattern.
When marine mammals approach the Unmanned Surface Vehicle, a significant energy increase is observed in the 5 kHz–9 kHz frequency band based on the Welch PSD results. As a novel unmanned surface platform, the Unmanned Surface Vehicle enables long-duration and wide-area acoustic observation of marine mammals in marine environments.

Figure 4. Spectral characteristics of acoustic detection in the South China Sea by Unmanned Surface Vehicles at different time periods.
Currently, the Unmanned Surface Vehicle-based marine mammal survey program is being carried out in an orderly manner. The development of underwater acoustic target recognition and acoustic anomaly detection methods based on deep learning is progressing steadily, with ongoing model training and sea trial experiments, accelerating the transition from theoretical research to engineering applications.
The Unmanned Surface Vehicle research team, in collaboration with multiple marine mammal research institutions, industry associations, and non-governmental organizations, plans to conduct a five-year (July 2023–June 2028) acoustic survey of marine mammals in the South China Sea and the Western Pacific Ocean. This program aims to provide high-quality real-time in situ datasets to support marine mammal research and conservation in China.
Key words:
Related News
How GNSS Positioning Solutions are Transforming Marine Surveying Techniques
How GNSS Positioning Solutions are Transforming Marine Surveying Techniques In the rapidly evolving field of marine surveying, the integration of GNSS (Global Navigation Satellite System) positioning solutions has revolutionized traditional methods of data collection and analysis. This article delves into how GNSS technology enhances precision, reliability, and safety in marine surveying, paving t
Nov 15,2026
Enhancing Deep Sea Navigation: The Role of Underwater Beacons in Positioning
The advent of underwater beacons has significantly revolutionized deep-sea positioning, a critical aspect of various marine operations, including scientific research, underwater construction, and search-and-rescue missions. These devices, often deployed to assist in locating submersible vehicles or underwater assets, utilize acoustic signals to provide real-time positioning data, which is vital fo
Nov 10,2026
A Comprehensive Guide to Selecting the Right Hydrophone Buoy for Marine Applications
A Comprehensive Guide to Selecting the Right Hydrophone Buoy for Marine Applications Table of Contents 1. Introduction to Hydrophone Buoys 2. What is a Hydrophone Buoy? 3. Importance of Hydrophone Buoys in Marine Research 4. Types of Hydrophone Buoys 5. Key Features to Consider When Selecting a Hydrophone Buoy 6. Environmental Factors Impacting Hydrophone Performance 7
Oct 21,2026
Understanding Shallow Water AIS Beacons: Enhancing Marine Safety and Navigation
Shallow water AIS beacons are specialized devices designed to aid navigation in shallow or restricted waters. AIS stands for Automatic Identification System, a vital communication tool for maritime safety that enables vessels to exchange information about their position, course, and speed in real time. These beacons serve as an important navigational aid, especially in areas where traditional navi
Oct 16,2026
Essential Marine AIS Beacon Transponder Maintenance Tips for Longevity
Essential Marine AIS Beacon Transponder Maintenance Tips for Longevity Table of Contents Understanding Marine AIS Beacon Transponders Importance of Regular Maintenance Routine Inspection Procedures Cleaning Your AIS Transponder Battery Maintenance and Replacement Software Updates and Firmware Management Troubleshooting Common Issues Professional Servicing: When to
Oct 11,2026
The Future of Wind Powered Unmanned Vessels: Revolutionizing Maritime Operations
In recent years, wind powered unmanned vessels have gained significant attention as a sustainable alternative in maritime operations. These vessels utilize wind energy as their primary propulsion source, which not only reduces reliance on fossil fuels but also minimizes greenhouse gas emissions. This environmentally conscious approach aligns perfectly with the global push for greener technologies
Oct 06,2026