Ecologically Valid Benchmarking and Adaptive Attention: Scalable Marine Bioacoustic Monitoring
Nicholas R. Rasmussen, Rodrigue Rizk, Longwei Wang, KC Santosh · 2025
Read Ecologically Valid Benchmarking and Adaptive Attention: Scalable Marine Bioacoustic Monitoring by Nicholas R. Rasmussen, Rodrigue Rizk, Longwei Wang, KC Santosh (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-04682.txt.
Introduction
Underwater sound monitoring fails in noisy, shifting conditions more often than standard tests reveal. Nicholas R. Rasmussen, lead author with three co-authors, present a framework meant to expose those failures in a 2025 paper. GetNetUPAM uses nested cross-validation arranged in site and year blocks, so each test fold reflects a distinct environment.
The nested stage measures model stability rather than tuning for inflated scores. With it, they test a network with attention for spectrograms of varying size. Attention acts as a learned noise filter. The goal is more honest estimates of how marine bioacoustic systems will perform when deployed.
Catalog record
| CREATOR | Nicholas R. Rasmussen, Rodrigue Rizk, Longwei Wang, KC Santosh |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 51 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-04682
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Ecologically Valid Benchmarking and Adaptive Attention: Scalable Marine Bioacoustic Monitoring Nicholas R. Rasmussen, Rodrigue Rizk, Longwei Wang, KC Santosh Abstract Underwater Passive Acoustic Monitoring (UPAM) provides rich spatiotemporal data essential for long‑term ecological analysis, but intrinsic noise and complex signal dependencies hinder model stability and generalization. While multilayered windowing has improved target sound localization, the variability induced by shifting ambient noise, diverse propagation effects, and mixed biological and anthropogenic sources demands robust architectures and rigorous evaluation. Conventional methods often oversimplify these challenges, limiting performance in real‑world deployments. Thus, we introduce GetNetUPAM , a hierarchical nested cross‑validation framework in which the nested stage is used not to inflate hold‑out performance, but to quantify model stability under ecologically realistic variability. By partitioning data into distinct site‑year segments, the framework preserves recording heterogeneity and ensures 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. This complementary design yields more reliable assessments for real‑world marine applications. Using GetNetUPAM as the evaluation backbone, we propose the Adaptive Resolution Pooling and Attention Network (ARPA‑N) , a neural architecture tailored for irregular spectrogram dimensions. Adaptive pooling with spatial attention extends the receptive field, capturing global context akin to transformers without excessive parameters. Under GetNetUPAM, ARPA‑N delivers a 14.4% gain in average precision over DenseNet baselines and a log 2 ‑scale order‑of‑magnitude drop in variability across all metrics, ensuring consistent detection across site‑year folds. This robustness advances scalable, accurate bioacoustic monitoring for conservation and ecological research. † † impactstatement: GetNetUPAM enables non‑invasive, real‑time monitoring of vulnerable marine species such as blue whales, supporting conservation while minimizing disturbance and promoting ethical research. Its robust detection in noisy underwater conditions reduces manual annotation and improves reliability. By modeling environmental variability through hierarchical evaluation, GetNetUPAM avoids overfitting to site‑specific noise, ensuring generalization across geographies and acoustic regimes. The modular design, adaptive attention, and scalable architecture extend to domains with sparse annotations and shifting signal profiles—including terrestrial bioacoustics, public‑health sensing, and security surveillance. Optimized for edge deployment, the framework balances stability and efficiency, demonstrating the broader potential of machine learning in safeguarding both natural and human environments. I Introduction Climate change increasingly threatens ecosystems, biodiversity, and the viability of vulnerable species [ 1 ] . Marine environments face compounded risks from rising temperatures, acidification, and shifting food webs. In response, the scientific community has called for integrated frameworks to assess climate vulnerability—capturing species’ sensitivity, exposure, and adaptive capacity [ 2 ] while distinguishing geographical population differences [ 3 ] . For many marine species, acoustic presence remains the only reliable indicator of distribution, making Underwater Passive Acoustic Monitoring (UPAM) critical for long‑term ecological baselines [ 4 , 5 , 6 , 7 ] . Fig. 1: Motivation for GetNetUPAM and ARPA-N. ( Left ) Input spectrograms in UPAM have diverse spectral and temporal resolutions, with colored overlays denoting target whale signals and intrinsic noise sources . Such heterogeneous, odd-dimensional inputs challenge model stability and can conceal performance variance behind a single evaluation score. ( Center ) The ARPA-N convolutional neural network applies spatial attention to suppress noise, standardize aspect ratios, and stabilize representations. ( Right ) The resulting uniform feature maps retain salient signal structure while reducing variability across conditions. This approach supports rigorous generalizability testing under UPAM helping to bolster minimal intrusion conservation methods [ 8 ] ., ensuring that processing steps alter the original signal only as much as necessary to enhance precision and interpretability. In UPAM, hydrophone‑equipped buoys capture underwater sounds and convert them into digital time‑series for machine‑learning analysis [ 9 ] . Although convolutional neural networks (CNNs) have advanced marine bioacoustic detection, they are typically evaluated on curated datasets [ 10 ] and often falter under field variability. For example, [ 11 ] found that a CNN trained on Glacier Bay vessel‑noise misclassified recordings with different noises (e.g., harbor seal roars), showing how shifts in local acoustics can degrade precision and distort ecological metrics—potentially erasing months of monitoring if undetected. State‑of‑the‑art methods also produce irregular spectrograms [ 10 , 12 ] , as shown in Figure 1 , which improve performance and interpretability [ 13 ] but add computational complexity such as non-standard downstream dimensions. These challenges underscore the need for UPAM models that generalize across spatial and temporal scales while remaining efficient and sensitive to global acoustic context. We present Generalization and Efficiency Testing for Neural Networks in UPAM (GetNetUPAM)—the first hierarchical nested cross‑validation framework for UPAM that uses nesting not to inflate hold‑out performance, but to quantify stability and reveal overfitting masked by conventional splits. Partitioning data into site‑year blocks mirrors natural variability and reduces bias from random splits [ 14 ] . Within each hold‑out block, standard cross‑validation enhances advanced detection techniques with rigorous stability testing. This multi‑tiered design quantifies trade‑offs among stability, accuracy, and inference efficiency, accounting for varying signal‑to‑noise ratios as shown in 1 , and supports robust UPAM systems in power‑ and bandwidth‑limited settings. The protocol also provides a reproducible benchmark for comparative studies. A core innovation is the Adaptive Resolution Pooling and Attention Network (ARPA‑N), which integrates complementary modules to boost stability, efficiency, and standardization. Under GetNetUPAM, ARPA‑N surpasses DenseNet baselines in mean precision—a metric often prioritized in conservation decision‑making [ 12 ] —and reduces variability (log 2 scale) by up to three‑fold, maintaining consistent performance across site‑year folds. Unlike CNNs restricted to local patches, ARPA‑N applies spatial attention to dynamically expand its receptive field across the entire signal [ 15 ] , mirroring transformer‑style global context modeling, while adaptive pooling standardizes feature maps and stabilizes predictions. This unified design captures fine‑scale details and non‑local dependencies, remaining resilient even amid ambient noise. Our contributions are summarized as follows: • Ecologically valid benchmarking: GetNetUPAM enforces site‑year partitions preserving environmental heterogeneity while combining nested cross-validation techniques, which are absent from existing benchmarks, enabling reproducible, deployment‑relevant assessments. • Operationally reliable detection: ARPA‑N combines adaptive resolution pooling and spatial attention to natively process irregular spectrograms, achieving higher mean precision and up to a three‑fold reduction in variability (log 2 ) over DenseNet baselines. • Deployment‑ready efficiency: Lightweight architecture reduces parameters and computation, supporting buoy‑mounted and other resource‑constrained platforms without sacrificing accuracy or stability. • Reproducible evaluation: Unified blocked and random cross‑validation quantifies stability–accuracy–efficiency trade‑offs, providing a transparent foundation for future UPAM and ecological‑monitoring studies. II Related Work The Antarctic Blue and Fin Whale Acoustic Trends Project [ 4 ] is one of the most comprehensive circumpolar repositories of whale recordings. Each site contributes at least one full year (ideally two consecutive) of data collected under diverse geographic, temporal, and equipment conditions, yielding 1,880.25 hours of annotated recordings and more than 300,000 hours of supplemental data. This scale and diversity are invaluable but amplify core UPAM challenges: detecting low‑frequency blue whale d‑calls, low‑SNR signals with sparse occurrences and spectral similarity to other frequency‑modulated calls—remains difficult [ 4 , 10 ] . These ecologically critical calls offer insight into population structure and distribution, yet their rarity and acoustic ambiguity make them prone to false detections. Classical UPAM Approaches Traditional large‑scale detection