2025_programme: Advancing Marine Monitoring Through Edge-Optimized Deep Learning: Hydrotwin’s Approach to Underwater Acoustic Surveillance
- Day: June 16, Monday
Location / Time: D. CHLOE at 14:50-15:10
- Last minutes changes: -
- Session: 08b. Bioacoustics and Soundscape
Organiser(s): Jennifer Miksis-Olds, Giacomo Giorli
Chairperson(s): Jennifer Miksis-Olds, Giacomo Giorli
- Lecture: Advancing Marine Monitoring Through Edge-Optimized Deep Learning: Hydrotwin’s Approach to Underwater Acoustic Surveillance
Paper ID: 2219
Author(s): Astrid van Toor, Guilherme Beleza Vaz
Presenter: Astrid van Toor
Abstract: Effective marine bioacoustics monitoring is essential for biodiversity conservation and ecosystem health assessment. This paper introduces Hydrotwin, a scalable multi-model acoustic monitoring platform that integrates underwater sound measurement with edge-optimised deep learning for real-time marine species detection and anthropogenic noise analysis in computationally constrained deployments. The system employs two specialised convolutional neural networks: Model A for vessel classification across low (≤112Hz) to medium (112-2500Hz), and high-frequency (2500-5000Hz) bands plus dolphin whistle detection (5000-20000Hz), and Model B for blue and fin whale call identification in the 5-125Hz spectrum. Model development leveraged extensive multi-year data collection efforts comprising 206 hours of expert-labelled recordings across Portuguese waters and the Azores for vessel/dolphin detection, and 1880 hours of baleen whale data. Through spectrogram analysis, multi-band frequency decomposition, and optimised wavelet feature extraction, our ultra-lightweight models (8K-120K parameters) are optimised for high-precision performance: 76-78\% precision for vessel detection, 94\% precision for dolphin whistles, 70\% precision for blue whale ABZ calls, and 77\% precision for fin whale pulse calls. We emphasise methodological rigour through temporally unbiased validation splits that prevent data leakage and provide realistic field performance assessment, contrasting with traditional random splits that artificially inflate metrics to unrealistic 90\%+ ranges. Real-time acoustic processing occurs in sliding windows, with detection events logged alongside environmental data from integrated sensors measuring wave height, wind speed, water temperature, and dissolved oxygen for subsequent correlation analysis. Field deployments demonstrate Hydrotwin's capacity for autonomous three-month monitoring cycles while supporting conservation policies through precise identification of anthropogenic disruptions and marine mammal presence patterns in diverse ocean environments.
Download the full paper
This paper is a candidate for the "Prof. John Papadakis award for the best paper presented by a young acoustician(under 40)"
- Corresponding author: Ms Astrid van Toor
Affiliation: BlueOasis
Country: Portugal
