Detecting the Presence of Sperm Whales' Echolocation Clicks in Noisy Environments

Guy Gubnitky, Roee DiamantIEEE/ACM Trans. ASLP 2024 · DOI

TL;DR

This paper introduces MPS-CD, a novel unsupervised method for detecting sperm whale echolocation clicks in noisy underwater environments by analyzing the stability of their unique multi-pulse structure (MPS). This approach significantly improves detection accuracy and reduces false alarms compared to existing methods, making it a valuable tool for real-time marine mammal monitoring.

What Problem Does This Paper Address?

Sperm whales (Physeter macrocephalus) use echolocation clicks for navigation and foraging, and these clicks have a distinctive multipulse structure (MPS). Detecting these clicks in the ocean is challenging due to several factors:

  • Noise: The underwater soundscape is filled with transient noises (e.g., from snapping shrimp or ships) that can mimic sperm whale clicks.
  • Variability: The MPS of sperm whale clicks can vary significantly depending on the whale's orientation relative to the receiver, making template-based detection difficult.
  • Multiple Whales: Distinguishing clicks from multiple simultaneously vocalizing whales adds another layer of complexity.

Current methods often rely on supervised learning, requiring large, site-specific labeled datasets, or template matching, which struggles with the inherent variability and noise. This paper aims to address these limitations by developing a robust, unsupervised method that can operate effectively in diverse and noisy marine environments with minimal prior knowledge.

Methods

The core of the proposed method, MPS-CD (MPS-based Clicks Detection), lies in leveraging the stability of the Multi-Pulse Structure (MPS) as a detection metric. Here's a breakdown of the key steps:

  1. Region of Interest (ROI) Identification: The raw acoustic signal is first enhanced (e.g., using a bandpass filter and the Teager-Kaiser Energy Operator - TKEO) to improve the signal-to-noise ratio. Then, potential transient events (ROIs) are identified based on local energy maxima.
  2. MPS Estimation: For each identified ROI, the MPS is calculated. This is defined as the time difference between the two most dominant (intense) pulses within the click. This measure is sensitive to both the internal structure of the click and potential multipath reflections.
  3. Clustering: A series of MPS values from a time buffer are clustered. The goal is to group together MPS measurements that are stable, indicating a consistent underlying signal source (likely a sperm whale). This clustering process is designed to be robust to the number of sources and noise transients.
  4. Verification: Before making a final detection decision, potential clicks within a cluster are further verified against constraints on pulse duration, inter-click interval (ICI), and spectral characteristics. This step helps to eliminate false positives that might have slipped through the initial MPS stability check.
  5. Detection Decision: The final decision of whether a sperm whale is present is based on the stability (low variance) of the MPS values within a validated cluster.

The novelty lies in using the stability of the MPS, rather than just its value, as the primary detection metric, and employing an unsupervised clustering approach to handle the complexities of noise and multiple sources.

Datasets

The authors evaluated their method using three distinct datasets:

  • THEMO observatory: Seven months of recordings from the Mediterranean Sea, characterized by significant ambient noise but no sperm whale clicks. This dataset was used to assess false alarm rates.
  • Dominica Island: Manually labeled data containing approximately 40,000 clicks from multiple sperm whales, along with noise-only recordings. This dataset was crucial for evaluating performance in a multi-whale, noisy scenario.
  • Bahamas (AUTEC): A dataset of 1,203 manually labeled clicks from a single sperm whale, including noise-only recordings. This provided a benchmark for single-whale detection.

The paper also makes their implementation code and a dataset of labeled clicks and noise transients available for reproducibility.

Key Insights and Ideas

  • MPS Stability as a Robust Metric: The core insight is that while the absolute value of the MPS can vary due to whale orientation, the stability of these MPS measurements over a short time buffer is a strong indicator of a sperm whale's presence. This is because noise transients are inherently random, and emissions from a single whale (even with slight orientation changes) will exhibit more consistent MPS patterns.
  • Unsupervised Learning for Real-World Challenges: The paper demonstrates that an unsupervised approach, focusing on intrinsic signal properties like MPS stability, can outperform supervised methods in complex, real-world scenarios where labeled data is scarce or site-specific.
  • Clustering for Source Separation: The clustering algorithm effectively groups similar MPS values, helping to separate potential sperm whale click trains from noise and even from clicks of other whales or different emission types from the same whale.
  • Handling Multipath: The method is designed to be relatively insensitive to multipath reflections, as these are assumed to affect all clicks within a buffer similarly, thus contributing to the overall stability of the MPS estimation.

Key Results

  • Improved Precision-Recall Trade-off: MPS-CD demonstrated a better trade-off between precision and recall compared to two benchmark detectors (Statistical and PamGuard), especially in noisy environments.
  • Significant Reduction in False Detections: The method achieved a substantial reduction in false alarm rates, particularly in the challenging THEMO dataset, which contained only ambient noise.
  • Robustness Across Environments: MPS-CD showed consistent performance across different marine environments and noise conditions, with minimal variation in threshold setting regardless of the training dataset.
  • Real-time Capability: The algorithm's complexity allows for processing in near real-time, making it suitable for on-site detection systems.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that the MPS can sometimes capture multipath reflections, which can affect the measurement. However, they argue that the stability of these measurements mitigates this issue.
  • The verification step might eliminate valid clicks if not carefully tuned, potentially splitting click sequences and impacting the clustering outcome.
  • Future Work:
    • Investigating the MPS measure further, along with statistical relationships between clicks, to distinguish between emissions from different sperm whales.
  • Exploring the possibility of performing the verification step before clustering to potentially improve efficiency and accuracy.
  • Extending the method to handle more complex scenarios, such as very close-range detections or highly dynamic whale movements.

Summary updated

BibTeX
@article{gubnitky2024detecting,
  author = {Guy Gubnitky and Roee Diamant},
  title = {{Detecting the Presence of Sperm Whales' Echolocation Clicks in Noisy Environments}},
  journal = {IEEE/ACM Trans. ASLP},
  year = {2024},
  doi = {10.1109/taslp.2024.3379899},
}
Figure