|
--- |
|
tags: |
|
- fastai |
|
--- |
|
|
|
This model was trained to as part of collaboration between [Mote Marine Laboratory & Aquarium](https://mote.org), [Southeast Coastal Ocean Observing Regional Association](https://secoora.org), and [Axiom Data Science](https://axiomdatascience.com) to develop a model capable of detecting and classifying fish vocalizations from audio files collected from hydrophones. |
|
|
|
More information available at [the project archive repo](https://github.com/axiom-data-science/project-classify-fish-sounds). |
|
|
|
--- |
|
|
|
|
|
# Model card |
|
|
|
## Model description |
|
|
|
This model was trained on spectrograms |
|
|
|
A [reproducible Jupyter notebook](https://github.com/axiom-data-science/project-classify-fish-sounds/blob/main/notebooks/train-resnet101-fastai.ipynb) describing the training of the model is available in the archive repo. |
|
|
|
## Intended uses & limitations |
|
|
|
The model was intended to be a proof on concept to aid researchers identify fish vocalizations through vast amounts of audio data collected from hydrophones. |
|
Although the training data was collected using multiple devices in multiple locations, the model may not be generally applicable to other uses. |
|
|
|
## Training and evaluation data |
|
|
|
A training set of spectrograms of fish calls was created based on annotations of fish sounds in passive acoustic recordings by a hydrophone were provided by Jim Locascio, Max Fullmer, and volunteers from the [Mote Marine Laboratory & Aquarium](https://mote.org). |
|
|
|
Due to severe imbalances in the number of samples per class, the training involved both under-sampling classes with many samples and over-sampling classes with few classes so that the model was trained on 50 samples per class. This number was derived in a completely ad-hoc fashion based on the distribution of class samples. |
|
|
|
### Class label description |
|
|
|
| Call Index | Description | |
|
|------------|-------------| |
|
| 0 | Background noise (no fish vocalizations) | |
|
| 1 | Black grouper 1 | |
|
| 2 | Black grouper 2 | |
|
| 3 | Black grouper grunt | |
|
| 4 | Black grouper spawning rush | |
|
| 5 | Black grouper chorus < 50% of file | |
|
| 6 | Black grouper chrous > 50% of file | |
|
| 8 | Unidentified sound type | |
|
| 9 | Red grouper 1 | |
|
| 10 | Red grouper 2 | |
|
| 17 | Red hind 1 | |
|
| 18 | Red hind 2 | |
|
| 19 | Red hind 3 | |
|
| 25 | Goliath grouper 1 | |
|
| 27 | Multi-phase goliath grouper | |
|
| 28 | Sea trout chorus | |
|
| 29 | Silver perch call | |
|
|
|
### Class indices in trained model |
|
|
|
Some classes did not meet the training criteria, high signal-to-noise ratio and minimum call overlap, and were therefore excluded from the model training. |
|
As such, the number of classes represented in the trained model is few than the amount of labeled classes in the training set. |
|
|
|
| Call Index | Description | |
|
-------------|-------------| |
|
|0 | No call | |
|
|1 | Black grouper call | |
|
|2 | Black grouper call 2 | |
|
|3 | Black grouper grunt | |
|
|4 | Unidentified sound | |
|
|5 | Red grouper 1 | |
|
|6 | Red grouper 2 | |
|
|7 | Red hind 1 | |
|
|8 | Red hind 2 | |
|
|9 | Red hind 3 | |
|
|10 | Goliath grouper | |
|
|11 | Goliath grouper multi-phase | |
|
|
