Abstract
Underwater Passive Acoustic Monitoring (UPAM) provides rich spatiotemporal data for long-term ecological analysis, but intrinsic noise and complex signal dependencies hinder model stability and generalization. Multilayered windowing has improved target sound localization, yet variability from shifting ambient noise, diverse propagation effects, and mixed biological and anthropogenic sources demands robust architectures and rigorous evaluation. We introduce GetNetUPAM, a hierarchical nested cross-validation framework designed to quantify model stability under ecologically realistic variability. Data are partitioned into distinct site-year segments, preserving recording heterogeneity and ensuring each validation fold reflects a unique environmental subset, reducing overfitting to localized noise and sensor artifacts. Site-year blocking enforces evaluation against genuine environmental diversity, while standard cross-validation on random subsets measures generalization across UPAM's full signal distribution, a dimension absent from current benchmarks. Using GetNetUPAM as the evaluation backbone, we propose the Adaptive Resolution Pooling and Attention Network (ARPA-N), a neural architecture for irregular spectrogram dimensions. Adaptive pooling with spatial attention extends the receptive field, capturing global context without excessive parameters. Under GetNetUPAM, ARPA-N achieves a 14.4% gain in average precision over DenseNet baselines and a log2-scale order-of-magnitude drop in variability across all metrics, enabling consistent detection across site-year folds and advancing scalable, accurate bioacoustic monitoring.
Introduction
Underwater Passive Acoustic Monitoring (UPAM) is a vital tool for tracking vulnerable species like blue whales. However, the ocean is a noisy, unpredictable environment. Shifting ambient noise, ship traffic, and complex sound propagation mean that a model trained on data from one location or year often fails when deployed in another. This "hidden overfitting" problem occurs when standard machine learning evaluations (using random data splits) mask the fact that models are memorizing the specific "acoustic fingerprint" of a sensor or a particular season. Furthermore, existing deep learning architectures like standard CNNs often struggle with the irregular dimensions of audio spectrograms or require too much computational power for edge deployment on remote, solar-powered hydrophone buoys.
The GetNetUPAM Framework
To address these challenges, we introduce GetNetUPAM, a hierarchical nested cross-validation scheme. Instead of random splits, it partitions data by site-year segments.
The "outer loop" holds out an entire site-year as a test set, forcing the model to prove it can handle a completely unseen environment. The "inner loop" quantifies model stability, ensuring the results aren't just a "lucky" fluke of data partitioning. This "site-year" block is the gold standard for "ecologically valid" testing, ensuring models generalize to real-world ocean variability rather than local noise artifacts.
The ARPA-N Architecture
The Adaptive Resolution Pooling and Attention Network (ARPA-N) is designed specifically for the unique demands of bioacoustics. It addresses the need for lightweight yet powerful feature extraction in noisy underwater environments.
The architecture incorporates two key features:
- Adaptive Pooling: This standardizes feature maps from irregular spectrogram dimensions, allowing the model to handle various input shapes natively.
- Spatial Attention (CBAM): By integrating the Convolutional Block Attention Module, the model learns to "look past" background noise—such as the constant hum of the ocean or the drone of a passing ship—to focus on salient spectro-temporal regions (the actual whale calls). This provides the "global context" benefits of a Transformer without the massive parameter count.
Key Results
The study utilized the Antarctic Blue and Fin Whale Acoustic Trends Project annotated library, characterized by extreme class imbalance (up to 170:1).
- Superior Precision: ARPA-N delivered a 14.4% gain in Average Precision (AP) over DenseNet baselines. In the most challenging "no support" scenario (Balleny Islands), ARPA-N maintained high ranking quality (0.367 AP vs. 0.298 for DenseNet).
- Reliability and Stability: ARPA-N showed a log2-scale drop in variability across different sites compared to baselines, proving it is significantly more consistent across diverse environments.
- Efficiency: Despite its performance, ARPA-N is lightweight, using only 4.97M parameters (compared to 24.1M for ResNet-50), making it ideal for low-power edge devices.
- Interpretability: Saliency maps showed that ARPA-N focused precisely on the whale call trajectories, whereas older models often focused on random background noise.
Conclusion
ARPA-N and the GetNetUPAM framework provide a robust path forward for scalable marine bioacoustic monitoring. By prioritizing ecological validity in evaluation and efficiency in architecture, these tools enable more accurate tracking of marine life in harsh, remote conditions. Future work involves "in-the-wild" testing on physical buoy hardware to measure real-time power consumption and expanding the architecture to handle multi-species monitoring simultaneously.
Summary updated