WasuratS/ocean_soundscape

Dataset Card for Koh Man Marine Soundscape (Raw Audio)

Dataset Summary

The Koh Man Marine Soundscape dataset is an unannotated, raw underwater acoustic dataset collected via Passive Acoustic Monitoring (PAM) around the Man Islands, Rayong Province, Thailand. The primary objective is to study and monitor coral reef ecosystem health by comparing a relatively pristine reef area (Ao Ton Liab / MN-GOOD) with a degraded reef area (Na Ban Beach / MN-DEGRADED).

This repository contains pure raw .wav data continuously recorded over an 8-week period using high-fidelity underwater audio recorders (SoundTrap ST600 HF from OceanInstruments NZ). The data is currently unannotated and serves as a foundational baseline for researchers and data scientists looking to develop unsupervised learning pipelines, perform ecological acoustic index calculations (ACI, ADI, AEI), or build custom labeling workflows for marine bioacoustics.

Supported Tasks and Leaderboards

• unsupervised-learning: Clustering and anomaly detection on raw marine soundscapes to discover novel biological or anthropogenic sound patterns.
• feature-extraction: Generating spectrograms, Long-Term Spectral Averages (LTSA), and computing traditional ecological acoustic indices.
• audio-classification: (Future/Intended) Serving as the base data for classifying underwater sound events, such as fish choruses, snapping shrimp activity, and vessel noise, once downstream annotations are applied.

Languages

The metadata is documented in English. The primary data modality is biological and environmental audio (Bioacoustics).

Dataset Structure

Data Instances

The dataset consists exclusively of raw, continuous .wav audio files directly from the recording devices, accompanied by basic field metadata.

An example of a raw file name structure (following the native SoundTrap output format):

10168.260805004645.wav

(Note: 10168 represents the device serial number, and 260805004645 represents the UTC timestamp in YYMMDDHHMMSS format).

Data Fields

The minimum metadata schema accompanying the raw audio includes the following tables:

• sites.csv: site_id, site_name, reef_condition, latitude, longitude, depth_range_m, habitat_note
• deployments.csv: Hardware deployment details such as mounting_type, depth_m, and recording schedules.
• recording_files.csv: File-level data including file_path, start_time, duration_sec, sample_rate, and checksum_sha256.
• maintenance_logs.csv: Field notes detailing deployment checks, biofouling observations, and retrieval logs.

(Note: No label or annotation files are included in this raw data release.)

Dataset Creation

Curation Rationale

Coral reefs are highly vulnerable ecosystems, and traditional visual surveys are often constrained by time, logistics, and weather. This raw dataset was curated to establish a continuous acoustic baseline for soundscape analysis, aiding in the evaluation of coral reef restoration efforts and the assessment of anthropogenic impacts (e.g., marine traffic). Providing the raw data allows the research community to apply novel signal processing and AI techniques without the bias of pre-defined segments or filters.

Source Data

Data Collection and Processing

• Hardware: Exclusively recorded using the SoundTrap ST600 HF (OceanInstruments NZ).
• Deployment: Recorders were mounted on custom gravity-friction anchors, elevated at least 60 cm above the seafloor (to reduce acoustic reflection and increase SNR) at an average depth of 3 meters.
• Data Processing: The data is provided in its pure raw format (.wav). No filtering, downsampling, or segmenting has been applied.

Personal and Sensitive Information

This dataset consists entirely of underwater environmental and biological audio recordings. It contains no sensitive personal information.

Considerations for Using the Data

Social Impact of Dataset

The insights derived from this dataset can inform marine conservation policies, such as establishing vessel speed limits or rerouting marine traffic to mitigate noise pollution in sensitive reef habitats.

Discussion of Biases

• Environmental Masking: Extreme weather events (heavy rain, strong waves) or tidal currents can mask biological sounds. Users are advised to review the accompanying weather or deployment logs to account for environmental noise during their analysis.
• Deployment Logistics: Hardware maintenance, such as SD card/battery replacement or diver presence near the hydrophones, introduces temporary anthropogenic noise (handling noise/diver noise).

Other Known Limitations

• The recording period spans 8 weeks, which may not capture full seasonal or annual variations in the marine soundscape.
• The data lacks ground-truth labels for specific biological calls. Researchers will need to apply their own detection algorithms or manual annotation protocols for supervised tasks.
• The accompanying biological metadata (coral cover, fish abundance, macrobenthos) is based on point-in-time visual line-transect surveys, representing a snapshot rather than continuous biological tracking.

Additional Information

Dataset Curators

The dataset was curated by the data science and marine biology research team conducting the soundscape study at Koh Man, Rayong Province, Thailand.

Licensing Information

All Rights Reserved. Data are owned by the Department of Marine and Coastal Resources (DMCR), Thailand. For inquiries regarding data usage, commercial applications, or academic collaboration, please contact the dataset curators or DMCR directly.

Citation Information

@misc{koh_man_soundscape_raw_2026,
author = {Soontronchai, Wasurat and Sukpholtham, Sitthichat and others},
title = {Koh Man Marine Soundscape Dataset (Raw Audio)},
year = {2026},
publisher = {Hugging Face},
note = {Data owned by Department of Marine and Coastal Resources, all rights reserved.},
}