Discovery and promotion of unknown sounds into operational detection targets for underwater passive acoustic monitoring under false alarm constraints
Abstract
Earth’s oceans cover 70% of the planet yet remain among the least observed regions. Underwater acoustics is a primary sensing modality, but recordings span biological, anthropogenic and natural sounds while labels remain sparse. We present an unknownness-aware monitoring framework for underwater acoustic sensing platforms: it adapts a pre-trained audio representation model to underwater soundscapes, defines an embedding-space unknownness score, and couples it with detector outputs in a Detect–Group–Promote–Union (DGPU) loop tuned under false positives per hour (FP/h) constraints: Detect surfaces candidate events from continuous audio; Group clusters unknown candidates in embedding space; Promote trains lightweight classifiers from sparse expert labels on selected clusters; and Union combines baseline and promoted detectors. Across a 56-class underwater Sound Event Detection (SED) benchmark and two cetacean field datasets, the framework attains a dataset-level bandwidth-matched out-of-distribution AUROC of 1.000 on Hawaiian survey data (HICEAS), raises right-whale upcall detection recall on continuous recordings (FRDR) from 0.027 to 0.390 at FP/h ≈ 10 with a supervised Promoter, with 300 labelled events (~ 32% of training-fold labels) reaching 0.37 (95% of the full-label recall), and classifies seven HICEAS cetacean species with the Promoter classifier alone (AUC 0.924–0.992 within-day and 0.51–0.97 cross-day with a fixed classifier), supporting stethoscope-like ocean monitoring.
Related
- Nanoscale-enabled cross-scale chemical sensing for underwater monitoring and beyond
- Towards operational epidemiology of Pyricularia oryzae in sports turf: contribution of a multi-matrix molecular monitoring strategy at stadium scale
- Structure-Guided Discovery of High-Affinity Methotrexate Aptamers for Electrochemical Drug Monitoring
- Parameter-Efficient Adaptation of Protein Language Models for Dual-Target Antimicrobial Peptide Discovery Under Strict Homology Constraints
- A first description of the ichthyofauna of the Namibian Islands Marine Protected Area using baited remote underwater video systems
BibTeX
@article{noda2026discovery,
title = {{Discovery and promotion of unknown sounds into operational detection targets for underwater passive acoustic monitoring under false alarm constraints}},
author = {Takuji Noda and Takuya Koizumi},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-73751-w},
url = {https://www.nature.com/articles/s41598-026-73751-w},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Discovery and promotion of unknown sounds into operational detection targets for underwater passive acoustic monitoring under false alarm constraints
AU - Takuji Noda
AU - Takuya Koizumi
PY - 2026
DA - 2026/10/02
JO - Scientific Reports
DO - 10.1038/s41598-026-73751-w
UR - https://www.nature.com/articles/s41598-026-73751-w
AB - Earth’s oceans cover 70% of the planet yet remain among the least observed regions. Underwater acoustics is a primary sensing modality, but recordings span biological, anthropogenic and natural sounds while labels remain sparse. We present an unknownness-aware monitoring framework for underwater acoustic sensing platforms: it adapts a pre-trained audio representation model to underwater soundscapes, defines an embedding-space unknownness score, and couples it with detector outputs in a Detect–Group–Promote–Union (DGPU) loop tuned under false positives per hour (FP/h) constraints: Detect surfaces candidate events from continuous audio; Group clusters unknown candidates in embedding space; Promote trains lightweight classifiers from sparse expert labels on selected clusters; and Union combines baseline and promoted detectors. Across a 56-class underwater Sound Event Detection (SED) benchmark and two cetacean field datasets, the framework attains a dataset-level bandwidth-matched out-of-distribution AUROC of 1.000 on Hawaiian survey data (HICEAS), raises right-whale upcall detection recall on continuous recordings (FRDR) from 0.027 to 0.390 at FP/h ≈ 10 with a supervised Promoter, with 300 labelled events (~ 32% of training-fold labels) reaching 0.37 (95% of the full-label recall), and classifies seven HICEAS cetacean species with the Promoter classifier alone (AUC 0.924–0.992 within-day and 0.51–0.97 cross-day with a fixed classifier), supporting stethoscope-like ocean monitoring.
SN - 2045-2322
ER -