README.md

---
license: cc-by-4.0
tasks:
  - underwater_acoustic_target_recognition # Corresponding to the underwater acoustic target recognition task
  - underwater_target_localization # Corresponding to the underwater target localization task
  - multi_task_learning # Corresponding to multi - task learning scenarios
frameworks: # Supported deep learning frameworks, not explicitly mentioned in the document, here are general examples
  - pytorch
  - tensorflow
language: # The dataset is audio data, with no specific language, left blank or marked as none here
  - en
tags:
  - underwater_acoustics
  - data_augmentation
  - ray_theory
datasets:
  train:
    - ShipsEar
    - DS3500
  test:
    - ShipsEar
    - DS3500
  evaluation:
    - ShipsEar
    - DS3500
metrics:
  - accuracy
  - range_localization_error
  - depth_localization_error
base_model:
  - MEG (multi - task, multi - expert, multi - gate) framework
indexing:
  results:
    - task:
        name: Underwater Target Recognition
      dataset:
        name: DS3500
        type: audio
        args: default
      metrics:
        - type: accuracy
          value: 95.93%
          description: recognition accuracy on DS3500 dataset
          args: default
    - task:
        name: Underwater Target Localization
      dataset:
        name: DS3500
        type: audio
        args: default
      metrics:
        - type: range_localization_error
          value: 0.2011 km
          description: range localization error on DS3500 dataset
          args: default
        - type: depth_localization_error
          value: 20.61 m
          description: depth localization error on DS3500 dataset
          args: default
domain:
  - audio # The dataset is of audio type, belonging to the acoustic field
---
- [English](README.md)
- [中文](README-zh.md)

## I. Basic Information of the Dataset

- **Dataset Name**: Underwater Acoustic Target Radiated Noise Dataset (including the original ShipsEar dataset and the enhanced DS3500 dataset)
- **Dataset Version**: V1.0
- **Release Date**: July 2025 (based on the paper submission date)
- **Update Records**: First release, no updates yet
- **Source and Contributors**:
  - Original ShipsEar dataset: Collected along the Atlantic coast of Spain from 2012 to 2013
  - Enhanced DS3500 dataset: Generated by institutions such as the School of Marine Engineering and Technology, Sun Yat - sen University based on ray theory (Contributors: Peng Qian, Jingyi Wang, etc., Affiliations: Sun Yat - sen University, Shanghai Marine Electronic Equipment Research Institute, etc.)
  - Contact information: qianp@mail2.sysu.edu.cn
- **Related Paper**: *Multi - Task Mixture - of - Experts Model for Underwater Target Localization and Recognition* (DOI: 10.3390/1010000)

## II. Description of Dataset Content

### 1. Data Scale and Distribution

| Dataset             | Categories                                                               | Number of Samples (5 - second segments)          | Category Distribution                                                                                                           | Data Format                          |
| ------------------- | ------------------------------------------------------------------------ | ------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------ |
| ShipsEar (original) | A, B, C, D, E (including environmental noise) numbered 0 - 4 in sequence | 1948 (A:345/B:235/C:785/D:395/E:188)             | Class A accounts for 17.7%, Class B for 12.1%, Class C for 40.3%, Class D for 20.3%, and Class E (environmental noise) for 9.6% | WAV audio (sampling frequency 16kHz) |
| DS3500 (enhanced)   | Same as ShipsEar (A - E, numbered 0 - 4 in sequence)                     | 1948 (consistent with the original dataset size) | Same as the original dataset                                                                                                    | WAV audio (sampling frequency 16kHz) |

### 2. Data Sources and Scenarios

- **ShipsEar (original)**: Actually collected ship radiated noise, covering 11 types of ships (such as motorboats, fishing boats, tugboats, etc.), with a sampling frequency of 52734Hz, preprocessed and segmented into 5 - second segments.
- **DS3500 (enhanced)**: Synthetic data of deep - sea environment (3500 - meter water depth) generated based on ray theory and BELLHOP sound field model. The simulation scenarios are as follows:
  - Geographical location: Deep - sea area north of the Zhongsha Islands (17.17°N, 114.22°E)
  - Target parameters: Distance from the sonar is 1 - 11km (interval 2km), depth is 100 - 1100m (interval 200m), with a total of 36 simulated positions
  - Sound field environment: The sound speed profile is calculated based on the temperature data from the WOA18 World Ocean Database. The seabed parameters are: sound speed 1601.9m/s, density 1.7g/cm³, attenuation coefficient 0.39f¹·⁷¹ dB/m (f is frequency in kHz)

## III. Data Preprocessing and Augmentation

- **ShipsEar preprocessing**:
  - Remove blank segments and split into 5 - second short segments to expand the data volume
  - No additional denoising is performed (retaining original noise characteristics)
- **DS3500 augmentation method**:
  - Simulate the marine acoustic channel based on ray theory and BELLHOP model
  - Perform channel transmission simulation on the 5 - second segments of ShipsEar to generate received signals including direct and shadow zones
  - Maintain the same sample size as the original dataset (to avoid a decrease in training efficiency)

## IV. Data Annotation Information

### 1. Annotation Content

- **Core labels**: Including classification labels, distance labels, and depth labels, as follows:
  - Classification labels: Corresponding to the 5 categories (A - E) of the original ShipsEar (represented by numbers such as "0" in the example for category encoding)
  - Distance labels: Horizontal distance between the target and the sonar (1.000 - 11.000km, accurate to 0.001km)
  - Depth labels: Deployment depth of the sonar (0.100 - 1.100km, accurate to 0.001km)

### 2. Annotation Example (File Path and Label Format)

train_list.txt

```
Path                                                                                                    Classification        Distance(km)        Depth(km)
E:\MTQP\wjy_codes\shipsear_5s_16k_ocnwav_Pos\0_0_2.wav                                               	0	       3.000	       0.100
E:\MTQP\wjy_codes\shipsear_5s_16k_ocnwav_Pos\0_0_3.wav                                               	0	       5.000	       0.100
```

## V. Dataset Uses and Applicable Scenarios

- **Main uses**:
  - Training and evaluation of underwater acoustic target recognition models
  - Development of underwater target localization (distance, depth) models
  - Verification of multi - task learning (simultaneously achieving recognition and localization) algorithms
- **Applicable scenarios**:
  - Marine monitoring in deep - sea and open - sea environments
  - Underwater defense and target detection
  - Performance testing of multi - task learning frameworks (such as MEG) (In the paper, the MEG framework achieved 95.93% recognition accuracy, 0.2011km distance error, and 20.61m depth error on this dataset)

## VI. Dataset Usage and Evaluation

- **Data partitioning**: A 5 - fold cross - validation strategy (sequential sampling) is adopted. For each type of data, 1 out of every 4 samples is selected as the test set, and the process is repeated 5 times (starting with samples 1 - 5 respectively) to ensure the balance of data order and category distribution.
- **Applicable models**: Supports deep learning models (such as CNN, Transformer) and multi - task frameworks (such as MoE, MEG), and is especially suitable for underwater multi - task models that need to integrate position information.

## VII. License Agreement

It follows the Creative Commons Attribution (CC BY) license agreement, allowing commercial use, modification, and distribution, with the need to indicate the original author and source.

## VIII. Related Resources

- Related paper: *Multi - Task Mixture - of - Experts Model for Underwater Target Localization and Recognition* (Authors: Peng Qian et al., Sun Yat - sen University)
- Recommended model: MEG (multi - task, multi - expert, multi - gate) framework (suitable for target recognition and localization tasks of this dataset)
- Community support: Technical support can be obtained by contacting the author (qianp@mail2.sysu.edu.cn)

## IX. Dataset Download and Update

- Download address: [ModelScope Dataset Repository](https://modelscope.cn/datasets) (Search for "Deep - sea Direct Zone - Acoustic Shadow Zone DS3500 Ship Radiated Noise Dataset (DS3500)")
- Update plan: There is no clear update plan yet. If there is a new version, simulated data of different deep - sea environments (such as different water depths and sea conditions) will be added.