Automatic detection of unidentified fish sounds: a comparison of traditional machine learning with deep learning
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Many species of fishes around the world are soniferous. The types of sounds fishes produce vary among species and regions but consist typically of low-frequency ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="m1"><mml:mo><</mml:mo></mml:math> 1.5 kHz) pulses and grunts. These sounds can potentially be used to monitor fishes non-intrusively and could complement traditional monitoring techniques. However, the significant time required for human analysts to manually label fish sounds in acoustic recordings does not yet allow passive acoustics to be used as a viable tool for monitoring fishes. In this paper, we compare two different approaches to automatically detect fish sounds. One is a more traditional machine learning technique based on the detection of acoustic transients in the spectrogram and the classification using Random Forest (RF). The other is using a deep learning approach and is based on the classification of overlapping segments (0.2 s) of spectrogram using a ResNet18 Convolutional Neural Network (CNN). Both algorithms were trained using 21,950 manually annotated fish and non-fish sounds collected from 2014 to 2019 at five different locations in the Strait of Georgia, British Columbia, Canada. The performance of the detectors was tested on part of the data from the Strait of Georgia that was withheld from the training phase, data from Barkley Sound, British Columbia, and data collected in the Port of Miami, Florida, United States. The CNN performed up to 1.9 times better than the RF ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="m2"><mml:msub><mml:mrow><mml:mi>F</mml:mi></mml:mrow><mml:mrow><mml:mn>1</mml:mn></mml:mrow></mml:msub></mml:math> score: 0.82 vs. 0.43). In some cases, the CNN was able to find more faint fish sounds than the analyst and performed well in environments different from the one it was trained in (Miami <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="m3"><mml:msub><mml:mrow><mml:mi>F</mml:mi></mml:mrow><mml:mrow><mml:mn>1</mml:mn></mml:mrow></mml:msub></mml:math> score: 0.88). Noise analysis in the 20–1,000 Hz frequency band shows that the CNN is still reliable in noise levels greater than 130 dB re 1 <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="m4"><mml:mi>μ</mml:mi></mml:math> Pa in the Port of Miami but becomes less reliable in Barkley Sound past 100 dB re 1 <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="m5"><mml:mi>μ</mml:mi></mml:math> Pa due to mooring noise. The proposed approach can efficiently monitor (unidentified) fish sounds in a variety of environments and can also facilitate the development of species-specific detectors. We provide the software FishSound Finder, an easy-to-use open-source implementation of the CNN detector with detailed documentation.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it