Key findings
- 1
To the authors' knowledge, this is the world's first biologging system that uses deep learning to classify sound sources and transmits the results when the tagged animal surfaces.
- 2
Sound is classified in a 3-level, 52-class scheme starting from biophony / geophony / anthrophony. Top-1 accuracy was 91.3% (Level 1), 86.7% (Level 2) and 62.3% (Level 3); top-5 was 97.1% / 96.3% / 86.3%.
- 3
It was demonstrated on seven green turtles (Chelonia mydas) that came to nest at Ibaruma Beach, Ishigaki Island, Okinawa. Deployments lasted on average a little under a day per individual (3.73–24.33 hours).
- 4
Over LTE-M, 778 position and 153 sound-classification/sensor points were recovered from five individuals; among individuals that transmitted, 91.6% of sound-classification/sensor data and 95.3% of location data were retrieved (recovered devices yielded 1,092 position and 240 sound-classification points across all seven).
- 5
Most classifications were background noise (76.4 ± 15.8%); among biological sounds, the giant moray (Gymnothorax javanicus, 39.0 ± 40.2%) and short-finned pilot whale (Globicephala macrorhynchus, 33.5 ± 22.7%) dominated, and 73.7% of sound observations were made shallower than 2 m.
Study overview
The system records underwater sound at 44.1 kHz / 16-bit with an omnidirectional AS-1 hydrophone (Aquarian Hydrophones; −208 dB re V/μPa sensitivity, 1 Hz–100 kHz) and runs deep-learning sound-source classification on an onboard microcontroller (STM32 H7 series). It also logs 3-axis acceleration and magnetism plus depth and temperature, and on surfacing it fixes position with FastLoc GPS and transmits over LTE-M. An Argos satellite variant was also developed for the open ocean.
The field trial ran in summer 2022 at Ibaruma Beach, Ishigaki Island, Okinawa (24°20′N, 123°50′E), on seven nesting green turtles (Chelonia mydas). It operated on a duty cycle of 5 minutes of recording every 30 minutes, and the loggers were retrieved after confirming the turtles' normal nesting behavior.
The LoggLaw device and its role
Streaming raw audio is not realistic over cellular or satellite bandwidth. Rather than the sound itself, this logger sends the 'classification results computed on the device' (up to 560 bytes per packet), enabling near-real-time underwater sound monitoring. The MobileNet v1 model is 8-bit quantized with TensorFlow Lite to about 370 kB so it runs within the microcontroller's RAM.
The unit is about 160 × 100 × 85 mm and weighs roughly 2 kg in air / 0.6 kg in water. With a 30,000 mAh lithium-polymer battery, it was designed for up to 40 days of continuous operation (without transmission). Integrating recording, classification, sensing, positioning and transmission in a single device, it is a core outcome of our work in underwater acoustics.
Why it matters
An underwater soundscape reflects the state of organisms, the seabed and human activity, but it has traditionally been observed with moored recorders fixed at a single point and recovered later — making wide-area, multi-point, near-real-time observation difficult. Using the animal itself as a platform lets us capture the sound environment while roaming the coastal zone without fixed infrastructure, and send results back when the animal surfaces.
Accumulating multi-point, long-term biologging observations points toward detailed 'soundscape maps' that reflect the distribution of sound sources. Our work in underwater acoustics took shape here as a world-first new method for monitoring marine ecosystems and environments.
Authors & collaborators
- Takuji Noda (Biologging Solutions Inc.) — corresponding author
- Takuya Koizumi (Biologging Solutions Inc.)
- Naoto Yukitake, Daisuke Yamamoto, Tetsuro Nakaizumi (Biologging Solutions Inc.)
- Kotaro Tanaka (Japan Fisheries Science and Technology Association / Ocean Policy Research Institute of the Sasakawa Peace Foundation)
- Junichi Okuyama (Fisheries Technology Institute, Japan Fisheries Research and Education Agency)
- Kotaro Ichikawa (Field Science Education and Research Center, Kyoto University)
- Takeshi Hara (Japan Fisheries Science and Technology Association)
Source
Takuji Noda, Takuya Koizumi, Naoto Yukitake, Daisuke Yamamoto, Tetsuro Nakaizumi, Kotaro Tanaka, Junichi Okuyama, Kotaro Ichikawa, Takeshi Hara (2024) Animal-borne soundscape logger as a system for edge classification of sound sources and data transmission for monitoring near-real-time underwater soundscape. Scientific Reports.
https://doi.org/10.1038/s41598-024-56439-xSupplementary materials
Related products & use cases
LoggLaw CAM — underwater video logger with hydrophone
A production underwater video logger that synchronously records video and sound (hydrophone) for marine observation including acoustics.
LoggLaw C Series (data loggers for marine animals)
Ultra-compact data loggers (depth, acceleration, magnetism, etc.) for turtles, fish and seabirds.
What is biologging — principles, history, applications
A systematic explainer of how each device — including hydrophones and acoustic loggers — works and the data it captures.
Related research
Digest by
A graduate of Kyoto University's Graduate School of Informatics and UC Santa Cruz's School of Environmental Studies. As co-founder of Biologging Solutions Inc., a Japan-based biologging equipment manufacturer, he oversees deployments of the company's products with municipalities, universities, and international consortia.
A biologging researcher with field experience including video-logger studies of penguin behavior in Antarctica. As co-founder of Biologging Solutions Inc., he leads the development of compact data loggers, GPS collars, and video loggers built directly around real-world research needs.