ShipsEar Dataset Benchmark for Maritime Acoustic Recognition

• ShipsEar is an open-source collection of 90 passive sonar recordings capturing 12 target classes, including 11 vessel types and natural noise, under real-world conditions.
• It provides rich environmental metadata such as source–receiver range, water depth, and wind speed, enabling robust multi-task and adversarial learning methods.
• The dataset supports state-of-the-art benchmarking with a standard file-level train/test split, spectrogram-based features, and recommended preprocessing protocols.

The ShipsEar dataset is an open-source, real-world corpus of ship-radiated underwater acoustic signals collected along the Atlantic coast of Spain. It is designed as a benchmark for underwater acoustic target recognition, emphasizing both robust classification and the study of environmental and acquisition “influential factors.” Its structure, labeling, and metadata make it the most widely referenced public acoustic ship classification resource with in-situ annotations of conditions such as source–receiver range, water column depth, and wind speed, supporting adversarial and multi-task learning approaches for robust maritime signal processing (Xie et al., 2024).

1. Dataset Composition and Class Structure

ShipsEar comprises 90 single-channel passive sonar recordings (“files”), each corresponding to one of 12 target classes: 11 vessel types and a single “natural noise” category. The class breakdown and number of recordings per type are:

Class	Recordings
Dredger	5
Fish boat	4
Motorboat	13
Mussel boat	5
Natural noise	12
Ocean liner	7
Passenger ship	30
Pilot ship	2
RO-RO ship	5
Sailboat	4
Trawler	1
Tugboat	2
Total	90

Recording durations range from 15 seconds to 10 minutes, totaling approximately 3 hours of audio. The nominal sampling rate is 52,734 Hz, though most experimental protocols resample to 44,100 Hz. Each recording file contains not only the acoustic waveform but also rich metadata for environmental conditions at acquisition.

2. Environmental Annotations and Influential Factors

For each recording, ShipsEar provides metadata encompassing environmental and acquisition parameters:

• Source–receiver range (distance from vessel to hydrophone): 0–350 m, grouped as “close” (0–50 m, 65 recordings) and “medium” (50–350 m, 25 recordings).
• Water column depth: categorized into 0–6 m (“coastal wetland,” 33 recordings), 6–12 m (“coastal zone,” 33), and 12–20 m (“deeper,” 24).
• Wind speed at recording: classified as 0 km/h (“calm,” 23), 0–11 km/h (“light breeze,” 29), and 11–18 km/h (“gentle breeze,” 25); data is missing for 13 files.
• Additional contextual metadata: hydrophone depth and gain, GPS coordinates, date/time, and oceanographic data.

This rich collection enables research on environmental robustness and covariate bias in marine signal processing. Influential factors are stored as per-file metadata and, for auxiliary task learning, are discretized into categorical labels.

3. Data Acquisition, Signal Processing, and Feature Extraction

ShipsEar was assembled using single- and multi-element hydrophone arrays. Multi-hydrophone sequences use only the loudest channel to ensure source clarity. All data were collected under realistic field conditions, primarily nearshore and at shallow depths (≤20 m).

The standard preprocessing chain consists of:

1. Resampling to 44,100 Hz.
2. Bandpass filtering (10–22,050 Hz optimal for ship-acoustic energy distribution).
3. Mean–variance normalization.
4. Segmentation into 30-second windows, with 15-second overlap, yielding hundreds of segments per class for deep learning.

Spectrogram-based features are extracted after windowing: linear spectrogram, mel spectrogram (400 filters), and constant-Q transform (CQT). CQT-spectrograms in the 10–22,050 Hz band, paired with a ResNet-18 backbone, demonstrate superior baseline performance.

4. Labeling, Task Protocols, and Splits

Each 30-second segment inherits the parent file's class and influential factor labels. Auxiliary tasks for multi-task learning are constructed by grouping continuous annotations into classes, enabling robust modeling of not only the vessel category but also environment-induced variability.

Train/test splits are defined at the file level—never segment level—to preclude leakage. The canonical split assigns 65 files (541 segments) to training and 26 files (84 segments) to testing, maintaining strict separation of recording scenarios.

No dedicated validation split is defined; model selection is based on segment-level accuracy on test segments, averaged across multiple random seeds.

5. Benchmarking Methods and Performance

ShipsEar supports the evaluation of standard and state-of-the-art models for underwater acoustic classification, with particular emphasis on adversarial multi-task learning frameworks. The benchmark protocol reflects emerging best practices:

• Train adversarial multi-task networks (e.g., AMTNet) using vessel class as the primary task and source range, water depth, and wind speed as auxiliary tasks.
• Adopt Local Masking and Replicating (LMR) augmentation for generalization.
• Use adversarial training (with staged/alternate optimization and lower per-stage learning rate) to prevent shortcut learning or factor-induced bias.

State-of-the-art adversarial multi-task models achieve ≈81% twelve-way classification accuracy on the ShipsEar benchmark (Xie et al., 2024).

6. Limitations and Best Practices

ShipsEar is concentrated on shallow, nearshore, close-range (<350 m) acoustic conditions, limiting direct generalization to deep water or distant propagation regimes. Several classes have very few files (e.g., trawler and pilot ship), leading to potential overfitting or class imbalance bias.

Wind-speed labels are missing for 13 recordings, which must be masked from wind-speed auxiliary tasks. Best practice mandates splitting by file and using multi-task and adversarial learning to avoid overfitting on “shortcut” environmental signatures.

Recommended methodological protocol:

1. File-level splitting for all model development and evaluation.
2. Incorporation of environmental labels as auxiliary tasks during training.
3. Use of augmentations and adversarial objectives to enhance invariance to non-target factors.
4. Consideration of distributional shift and small-sample bias in interpretation.

7. Research Significance and Applications

ShipsEar serves as the de facto benchmark for underwater acoustic ship classification under real-world, inhomogeneous conditions, supporting research in both robust signal-based recognition and environmental bias mitigation. Its detailed metadata, class structure, and environmental annotations make it uniquely suitable for evaluating algorithms under domain shift, adversarial training, and multi-task learning settings (Xie et al., 2024).

A plausible implication is that ShipsEar’s modeling framework and labeling conventions can be adapted for transfer learning, self-supervised, or unsupervised research where labeled sonar data are rare or expensive to acquire. Its emphasis on environmental metadata, label grouping, and adversarial protocols reflects broader trends in robust acoustic sensing and recognition under realistic deployment conditions.