TL;DR
This paper introduces a low-complexity Convolutional Neural Network (CNN) for robustly detecting and classifying fin whale vocalizations (20 Hz pulses) from diverse acoustic recordings. The study reveals significant long-term trends in fin whale song, including a linear increase in inter-pulse intervals (IPIs) and a seasonal decrease in pulse center frequency, offering valuable insights into the evolution of marine mammal communication.
What Problem Does This Paper Address?
Fin whales produce complex, low-frequency vocalizations known as "songs," which are crucial for their social interactions, particularly for mating. Understanding how these songs evolve over time is vital for conservation efforts and for deciphering the cultural and environmental factors influencing marine mammal communication. However, analyzing these long-term trends is challenging due to the vast amount of acoustic data, the variability in recording equipment, and the presence of environmental noise. Previous studies often relied on smaller datasets or less robust detection methods, limiting the scope and reliability of their findings. This paper aims to address this by developing a more efficient and accurate automated system for analyzing fin whale songs over an extended period.
Methods
The core of this research lies in a novel, low-complexity Convolutional Neural Network (CNN) designed for automated detection of fin whale 20 Hz pulses. This CNN was trained using a supervised, iterative process (active learning) on a diverse dataset of acoustic recordings. Key methodological aspects include:
- CNN Architecture: A relatively simple CNN with 3 depth-wise convolutional layers was employed, making it computationally efficient.
- Data Pre-processing: Input waveforms were standardized, and synthetic brown noise was added to improve robustness against varying noise levels. Mel spectrograms were computed as input features.
- Pulse Detection: The CNN was trained to identify the presence of a 20 Hz pulse within 5-second segments. Post-processing involved extracting pulse timings with high temporal resolution.
- Pulse Classification: Pulses were classified into two types (A and B) based on their center frequency, using a Gaussian Mixture Model (GMM).
- Spectro-temporal Analysis: Detailed pulse characteristics, including exact time position, center frequency, bandwidth, and Signal-to-Noise Ratio (SNR), were extracted.
- Inter-Pulse Interval (IPI) Measurement: IPIs were calculated as the time difference between consecutive pulses, with sequences longer than 45 seconds considered part of the same song.
- Trend Analysis: Inter-annual and intra-annual trends in IPI and pulse center frequency were analyzed using linear regression on quantized data.
The paper highlights the CNN's advantage over traditional template matching methods, demonstrating superior performance and robustness in challenging acoustic conditions.
Datasets
The study utilized passive acoustic recordings from two hydrophone stations in the Pelagos Sanctuary (Western Mediterranean Basin) collected between 2008 and 2018.
- Training Data: The Magnaghi dataset (1999) was used solely for training the CNN.
- Evaluation Data: Data from the Boussole (2008-2009) and Bombyx (2015-2018) stations were used for testing and analysis.
- External Benchmark: The model's performance was also compared against a dataset published by Madhusudhana et al. (2021), which also used a CNN for fin whale pulse detection.
The paper notes that the Magnaghi data was not included in the subsequent analysis due to overlapping songs in available segments, which the proposed method does not currently handle.
Key Insights and Ideas
- Robust Pulse Detection: A low-complexity CNN can effectively detect fin whale 20 Hz pulses even with significant variability in recording devices and high levels of ambient noise, outperforming traditional template matching.
- Stereotypical IPIs: The study confirms that inter-pulse intervals (IPIs) are strongly determined by the type of pulse sequence (e.g., 'AA', 'AB', 'BA', 'BB'), with different sequences exhibiting distinct typical intervals.
- Linear IPI Increase: A significant finding is the consistent, linear increase in stereotypical IPIs across all analyzed pulse sequence types, averaging approximately 0.1 seconds per year over the past two decades. This suggests a long-term evolutionary or cultural shift in fin whale song structure.
- Intra-annual Frequency Trend: Fin whale pulse center frequencies exhibit a clear intra-annual trend, decreasing from August to February, with an estimated slope of -0.10 Hz/month. This contrasts with previous findings in other populations and suggests a unique seasonal pattern in the Mediterranean.
- No Inter-annual Frequency Trend: Unlike some other populations, no significant inter-annual trend in pulse center frequency was observed for Mediterranean fin whales, potentially indicating a unique characteristic of this isolated population.
- Weak Correlation between IPI and Frequency: While previous hypotheses suggested a link between decreasing frequency and increasing IPI, this study found only a weak correlation between IPI and center frequency in the Mediterranean population, challenging that specific explanation for the observed IPI trends.
Key Results
- The developed CNN achieved high performance, with AUC values of 0.992, 0.943, and 0.997 on the Bombyx, Magnaghi, and Boussole test sets, respectively.
- The template matching method achieved an AUC of 0.898, significantly lower than the CNN.
- The study confirmed a linear increase in stereotypical IPIs for 'AA', 'AB', and 'BA' bi-grams, with slopes of 0.091, 0.096, and 0.097 s/year, respectively, all with p-values < 0.001.
- An intra-annual decrease in pulse center frequency was observed with a slope of -0.10 Hz/month (p-value = 0.03), and a coefficient of determination of 0.73.
- Pearson correlation analyses showed weak negative correlations between IPI and center frequency for 'BA' (-0.37), 'AB' (-0.22), and 'AA' (-0.35) bi-grams, all with p-values < 0.001.
Limitations and Future Work
- Overlapping Songs: The current CNN model does not effectively handle overlapping fin whale songs, which limited the inclusion of the Magnaghi dataset in the final analysis. Future work could focus on developing methods to disentangle overlapping vocalizations.
- Seasonal Sampling Bias: The study acknowledges that unbalanced seasonal sampling through the years might introduce observational bias in the calendar of singing activity. More comprehensive data collection across seasons and years is needed.
- Mediterranean Specificity: The lack of an inter-annual frequency trend in the Mediterranean population, compared to other regions, suggests unique factors at play. Further research is needed to understand why this population might be an exception.
- Hypothesis Testing: While the study tested hypotheses regarding the drivers of IPI and frequency shifts (e.g., cultural, environmental, population density), definitive conclusions require more data and potentially integration with other environmental or biological datasets.
- Reproductive Seasonality: The paper notes the lack of a specific reproductive season for Mediterranean fin whales, which complicates the interpretation of intra-annual IPI trends that are often linked to mating cycles in other populations. Further investigation into the reproductive biology of this population is warranted.
Summary updated