Categorizing Shallow Marine Soundscapes Using Explained Clusters

Clea Parcerisas, Irene T. Roca, Dick Botteldooren, Paul Devos, Elisabeth DebusschereJournal of Marine Science and Engineering 2023 · DOI

TL;DR

This paper introduces a novel semi-supervised method to automatically categorize shallow marine soundscapes, transforming raw acoustic data into meaningful ecological labels. By combining dimensionality reduction, clustering, and explainable AI, the approach provides ecologically interpretable soundscape categories, crucial for monitoring marine environments under increasing anthropogenic pressure.

What Problem Does This Paper Address?

Marine soundscapes, the complex acoustic environments of underwater habitats, are vital for understanding ecosystem health and biodiversity. However, these environments are increasingly threatened by anthropogenic noise from shipping, industry, and other human activities. Effectively monitoring these soundscapes is challenging due to the sheer volume of acoustic data and the lack of standardized methods for categorization. Traditional approaches often rely on manual annotation or eco-acoustic indices that can be inconsistent in marine settings. This paper addresses the need for an automated, robust, and ecologically interpretable method to classify and understand these complex underwater soundscapes, particularly in challenging shallow water environments.

Methods

The authors propose a multi-stage methodology:

  1. Acoustic Feature Extraction: Raw acoustic data is processed into one-third octave bands, a standard for environmental impact assessments, creating acoustic samples.
  2. Dimensionality Reduction: Uniform Manifold Approximation and Projection (UMAP) is used to map these acoustic samples into a 2D space, preserving local and global data structure. This step is crucial for visualizing and clustering the high-dimensional acoustic data.
  3. Clustering: Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is applied to the 2D UMAP representation to identify distinct groups of similar soundscapes.
  4. Artifact Removal: A semi-supervised approach is employed to identify and remove non-ecologically relevant artifacts (e.g., clipping, electronic noise) from the dataset. This involves manual labeling of some artifacts and then using DBSCAN to identify and remove artifact-heavy clusters.
  5. Soundscape Categorization: After artifact removal, UMAP and DBSCAN are applied again to the "clean" data to derive the final acoustic categories.
  6. Ecological Interpretation: A Random Forest (RF) classifier is trained to predict the acoustic categories using environmental variables (e.g., bathymetry, salinity, shipping density). Explainable AI (XAI) tools, specifically SHAP (SHapley Additive exPlanations), are then used to interpret the RF model's predictions, revealing which environmental parameters are most influential in defining each soundscape category. This step is key to providing ecological meaning to the identified clusters.

Datasets

The study utilizes acoustic data collected in the Belgian Part of the North Sea (BPNS).

  • Acoustic Data: 54 deployments were recorded between April 2020 and October 2020, and in June 2021, totaling approximately 40 hours of data. This data was collected using SoundTrap ST300HF and Bruel & Kjaer Nexus 2690 instruments.
  • Environmental Data: A comprehensive set of environmental variables was compiled, including season, bathymetry, moon phase, shipping density, substrate type, seabed habitat, benthic habitat, salinity, temperature, surface current speed, height above sea level, proximity to shipwrecks, and distance to coast.

Key Insights and Ideas

  • Automated Ecological Profiling: The core innovation is the ability to automatically generate ecologically meaningful labels for soundscapes, moving beyond simple acoustic similarity to understanding why certain sounds occur together.
  • UMAP + DBSCAN Synergy: The combination of UMAP for dimensionality reduction and DBSCAN for clustering proved effective in identifying distinct acoustic groups, even in complex, noisy marine environments.
  • XAI for Ecological Understanding: SHAP values are critical for translating abstract acoustic clusters into understandable ecological drivers, highlighting the importance of specific environmental factors for each soundscape type.
  • Shallow Water Complexity: The study highlights the unique acoustic challenges of shallow marine environments, where soundscapes can vary significantly over small spatial scales due to factors like bathymetry and tidal currents.
  • Artifacts as a Separate Class: The method effectively identifies and removes artifacts, treating them as a distinct "noise" class, which is crucial for robust analysis.

Key Results

  • 17 Acoustic Categories: The methodology successfully categorized the BPNS soundscapes into 17 distinct acoustic classes.
  • High Predictive Accuracy: The Random Forest model achieved a mean CV accuracy of 91.77%, indicating that the environmental parameters were strong predictors of the identified soundscape categories.
  • Dominant Environmental Drivers: Key environmental parameters influencing soundscape categorization included instrument depth, time of day, distance to coast, salinity, and moon phase.
  • Ecological Interpretability: The SHAP analysis provided detailed ecological descriptions for each cluster, linking them to specific environmental conditions (e.g., "high instrument depth, day, close to the coast, low salinity, spring" for Cluster 0).
  • Artifact Identification: The artifact removal process successfully identified and excluded clusters dominated by non-ecological sounds, which constituted 15.30% of the initial dataset.

Limitations and Future Work

  • Redundant Variables: The study acknowledges that redundant environmental variables can complicate interpretation. While efforts were made to address this, further refinement might be needed.
  • Temporal Resolution: The study prioritized spatial coverage over long temporal sampling at each site. Future work could benefit from longer deployments to capture seasonal and annual cycles more comprehensively.
  • Frequency Band: The chosen frequency band captured significant information, but expanding it could potentially reveal new clusters or finer distinctions.
  • Generalizability: While tested in the BPNS, the authors suggest applying the method to other datasets and acoustic contexts to assess its robustness and generalizability.
  • Manual Analysis of Misclassified Data: The authors suggest manually analyzing misclassified data to identify missing explanatory parameters or events not captured by the model.
  • Long-Term Monitoring: The method is well-suited for long-term monitoring, and future work could focus on applying it to assess acoustic trends and ecosystem changes over time.

Summary updated

Abstract

Natural marine soundscapes are being threatened by increasing anthropic noise, particularly in shallow coastal waters. To preserve and monitor these soundscapes, understanding them is essential. Here, we propose a new method for semi-supervised categorization of shallow marine soundscapes, with further interpretation of these categories according to concurrent environmental conditions. The proposed methodology uses a nonlinear mapping of short-term spectrograms to a two-dimensional space, followed by a density-based clustering algorithm to identify similar sound environments. A random forest classifier, based on additional environmental data, is used to predict their occurrence. Finally, explainable machine learning tools provide insight into the ecological explanation of the clusters. This methodology was tested in the Belgian part of the North Sea, and resulted in clearly identifiable categories of soundscapes that could be explained by spatial and temporal environmental parameters, such as distance to the shore, bathymetry, tide or season. Classifying soundscapes facilitates their identification, which can be useful for policy making or conservation programs. Soundscape categorization, as proposed in this work, could be used to monitor acoustic trends and patterns in space and time that might provide useful indicators of biodiversity and ecosystem functionality change.
BibTeX
@article{parcerisas2023categorizing,
  author = {Clea Parcerisas and Irene T. Roca and Dick Botteldooren and Paul Devos and Elisabeth Debusschere},
  title = {{Categorizing Shallow Marine Soundscapes Using Explained Clusters}},
  journal = {Journal of Marine Science and Engineering},
  year = {2023},
  doi = {10.3390/jmse11030550},
}
Figure