AI Learns to Identify Ships From Sound With Only a Handful of Examples

Under the surface of the world’s oceans, every vessel leaves a signature. The drone of a cargo ship’s propeller, the rhythmic pulse of a fishing trawler’s engine, and the faint hum of an autonomous underwater glider all propagate through seawater as distinctive acoustic patterns. For years, researchers have tried to teach artificial intelligence systems to recognize these signatures automatically, a task known as underwater acoustic target recognition. The technology promises major advances in ocean monitoring, port security, fisheries management, and environmental protection, but it has long been hampered by a stubborn bottleneck: deep learning models typically demand enormous quantities of labeled audio recordings, and annotated underwater acoustic data is scarce, expensive, and difficult to collect.

A new study published in Earth Science Informatics by Gang Hu of Anshan Normal University, Yifei Song of Universiti Utara Malaysia, and Mohamad Farhan Mohamad Mohsin of the same institution’s School of Computing tackles this problem head-on. The researchers have developed a model called AMCUATR, short for Attention-based Meta-learning and Class-Incremental Underwater Acoustic Target Recognition. The framework is designed to do two things that conventional deep learning systems struggle to combine: learn to recognize new vessel classes from only a handful of labeled examples, and absorb those new classes without erasing what the model has already learned about previous ones. According to the paper, the model achieved a recognition accuracy of 80.01 percent under a demanding 4-way 5-shot testing scenario on the widely used ShipsEar dataset, meaning it correctly identified ship types after seeing just five examples of each category.

The challenge the researchers address is twofold. First, there is the few-shot learning problem: in real deployments, oceanographers and naval operators constantly encounter vessel types that were never represented in the training data, and gathering hundreds or thousands of labeled recordings for each new class is rarely feasible. Second, there is the problem of catastrophic forgetting, a well-documented failure mode in neural networks in which a model trained on new categories abruptly degrades on the categories it previously mastered. Standard fine-tuning approaches, when applied to a trained network with new data, tend to overwrite the internal weight configurations that encoded the old knowledge, effectively wiping the model’s memory of earlier targets.

To overcome these obstacles, AMCUATR combines three interlocking technical components. The first is a dual-branch time-frequency feature extraction module with adaptive channel calibration. Underwater acoustic signals carry information both in how their energy evolves over time and in how it is distributed across frequencies, so the model processes two complementary representations in parallel. One branch focuses on temporal structure while the other captures spectral characteristics, and a channel attention mechanism then learns to weight the most informative feature channels more heavily. This adaptive calibration allows the network to suppress noisy or redundant dimensions and amplify the subtle cues that distinguish one vessel’s radiated noise from another’s, a critical capability when working with the low signal-to-noise recordings typical of real ocean environments.

The second component is a hierarchically embedded Swin Transformer, a vision architecture adapted here for acoustic data. Unlike conventional convolutional networks that process local neighborhoods of pixels or spectrogram cells, transformer architectures use self-attention to model relationships across an entire input. The Swin variant computes attention within hierarchical windows and shifts those windows between layers, allowing the model to capture dependencies at multiple scales without the prohibitive computational cost of full global attention. In AMCUATR, this module serves to build a richer global context around the features extracted by the dual-branch front end, improving the discrimination of the embedding space in which different vessel classes must be separated. The authors report that this hierarchical global modeling improves feature quality precisely where few-shot classification needs it most: when only a few examples define each class, the geometry of the embedding space determines whether those examples cluster cleanly or blur together.

The third component, and the conceptual heart of the system, is a prototype-based class-incremental learning mechanism. The approach draws on ideas from prototypical networks, a meta-learning method in which each class is represented by a prototype vector, typically the average of the embedding vectors of its few labeled examples. Classification then reduces to measuring which prototype an unknown sample lies closest to. Because class knowledge is stored in these prototype summaries rather than only in the network’s weights, new categories can be added by computing new prototypes from a handful of samples, while the prototypes of previously learned classes remain available to anchor the model’s old knowledge. Combined with meta-learning, in which the model is trained across many simulated few-shot tasks so that it internalizes how to adapt quickly rather than memorizing specific classes, this design gives AMCUATR its ability to keep learning over time.

The experimental evaluation was conducted on two public benchmark datasets that have become standard proving grounds for this field. ShipsEar, first described in Applied Acoustics in 2016, is a database of underwater vessel noise recordings covering multiple ship categories along with environmental background noise. DeepShip, introduced in Expert Systems with Applications in 2021, offers a larger collection of ship-radiated noise organized by vessel type and operating condition. Testing on both datasets allowed the researchers to assess how well the framework generalizes beyond a single recording campaign, an important consideration given that acoustic conditions vary with sea state, water depth, and sensor placement.

The incremental learning results are particularly notable. In the few-shot class-incremental setting, where the model had to learn novel vessel categories from only five labeled examples each, AMCUATR achieved a new-class recognition accuracy of 83.14 percent. More importantly for the forgetting problem, it maintained an old-class retention accuracy of 79.05 percent with a forgetting rate of just 1.20 percent, yielding an average incremental accuracy of 81.10 percent across the evaluated setting. Those numbers indicate that the model acquired new knowledge at only a modest cost to its existing capabilities, which is precisely the balance that practical, continuously deployed recognition systems require. A monitoring network that must be updated each time a new vessel class appears cannot afford to retrain from scratch, nor can it tolerate a model that forgets yesterday’s targets every time it learns today’s.

The significance of this work extends beyond the specific accuracy figures. Underwater acoustic monitoring is becoming central to a range of scientific and operational missions, from tracking the impact of shipping noise on marine mammals to detecting natural gas seeps with autonomous underwater vehicles, a task explored in recent control engineering research. Passive acoustic sensors are relatively inexpensive and can operate continuously in remote locations, but their value depends on automated analysis, since human analysts cannot listen to years of recordings. Systems like AMCUATR point toward recognition platforms that can be deployed with modest initial training data and then grow their knowledge base incrementally as they encounter new sounds in the field, whether novel vessel designs, unfamiliar marine species, or anomalous signals worth flagging for investigation.

The study also situates itself within a rapidly evolving research landscape. Recent years have seen transformer-based networks, self-supervised pretraining with masked hierarchical tokens, graph embeddings built on mel-spectrograms, and physics-informed Bayesian graph networks all applied to underwater acoustic recognition, while other groups have pursued data augmentation through generative adversarial networks and diffusion models to stretch limited datasets further. The contribution of Hu and colleagues is to fuse attention-based feature extraction, meta-learning, and prototype-driven incremental learning into a single coherent pipeline and to demonstrate quantitatively that the combination holds up under the joint pressures of scarce labels and evolving class sets. The authors note that the datasets analyzed in the study are publicly available, and the research received no external funding. As ocean observation networks expand and the demand for autonomous acoustic intelligence grows, frameworks that can learn continuously from little data may prove to be not just a convenience but a necessity for keeping pace with the ever-changing soundscape beneath the waves.

Subject of Research: Few-shot class-incremental deep learning for underwater acoustic target recognition of ship-radiated noise

Article Title: A few-shot UATR Model integrating attention mechanism-based meta-learning with class incremental learning approaches

Article References: A few-shot UATR Model integrating attention mechanism-based meta-learning with class incremental learning approaches. (n.d.). https://doi.org/10.1007/s12145-026-02247-y

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02247-y

Keywords: underwater acoustics, deep learning, meta-learning, few-shot learning, class-incremental learning, attention mechanism, Swin Transformer, ship-radiated noise, ShipsEar, DeepShip, prototype networks, catastrophic forgetting