Machine learning for efficient segregation and labeling of potential biological sounds in long-term underwater recordings

This script collection is intended for re-use of the method presented in:

Parcerisas, C.; Schall, E.; Te Velde, K.; Botteldooren, D.; Devos, P.; Debusschere, E.. Machine learning for efficient segregation and labeling of potential biological sounds in long-term underwater recordings. Submitted to Frontiers in Remote Sensing (2024). https://doi.org/10.3389/frsen.2024.1390687

All the data and the configuration files necessary to run the outputs can be found on IMIS https://www.vliz.be/nl/imis?module=dataset&dasid=8543

The scripts are ready to be used in other locations. All the scripts are organized around "datasets", which basically means a folder with these subfolders:

dataset folder
├── images                      # segmented spectrograms for yolo prediction (optional)
├── labels                      # txt files yolo format (optional) obtained from manual annotations
├── predictions                 # txt files yolo format (optional) obtained from model predictions
├── wavs                        # raw wav files (optional)
└── annotations.txt             # Raven selection table (optional)

As you see, all the components are optional, and this is because depends on what you want to run some parts will be necessary or not.

For each of these folders (dataset), a config.json needs to be described. An example is provided in the file config_example.json. The parameters are:

{
    "duration" : 20, 
    "overlap" : 0.5,
    "desired_fs" : 24000,
    "channel" : 0,
    "color": true, 
    "log": false,
    
    "nfft" : 2048,
    "win_len" : 2048,
    "hop_ratio" : 8,
    
    "normalization_style": "noisy",
    
    "wavs_folder" : "path_to_the_wavs",
    "dataset_folder" : "path_to_dataset_folder",
    
    "annotations_file": "path_to_annotations.txt"
}

• duration: duration of each segment in seconds
• overlap: 0 to 1, overlap between segments
• desired_fs: sampling frequency to resample to, in Hz
• channel: channel number
• color: bool, true converts spectrograms to RBG values, otherwise grayscale
• log: bool, true makes spectrograms with a logarithmic frequency scale
• nfft: number of FFT bins to apply to the segment to generate spectrogram
• win_len: length of the window to apply the FFT to
• hop_ratio: 0 to 100, percentage of hop between windows (1-overlap)
• normalization_style: "noisy" or "low_freq"
• wavs_folder: path to a folder where all the wavs to be analyzed are (they can be in subfolders!)
• dataset_folder: path to the main folder of the dataset, with the images and the labels folder (if applicable)
• annotations_file: path to the annotations file, can be from predictions or manual annotations. It does not necessarily need to be in the dataset folder

Analyze a location

To just analyze a new location using the existing model:

1. Download the model weights (choose which one)
2. Run analyze_one_location.py and specify the path to the model and the path to the config file. The wavs folder does not necessarily need to be under the dataset folder. The annotations file should be dataset_folder/roi_annotations_clean.txt. A folder names predictions/labels under the dataset folder will be created, but the images will not be saved to save space.

Train models

To train models, you will have to:

1. Manually create an annotations file using Raven (and set the correct path in the config.json file)
2. Create a dataset folder, and create the subfolders: "train", "valid", "images", "labels"
3. Create the spectrograms using create_long_spectrograms.py. This will also
4. Run split_train_test_valid.py. This will split your images and your labels from images and labels folder to the train and valid folders.
5. If for some reason you want to re-join all the images and labels, run reshuffle_files.py
6. Fill in the correct model_example.yaml with the paths to the valid and train paths (folders just created)
7. Run the retraining.py script and pass the model_example.yaml file path when prompt with the question

Active Learning

1. Create an unlabeled pool with an associated config.json file. This means create a dataset_folder with the necessary subfolders, and one subfolder called active_learning
2. Run the create_long_spectrograms.py script for the unlabeled pool config (this will speed up things)
3. Run the active_learning.py script, it will guide you through the steps with an interactive console

Other (maybe) useful scripts

• cluster.py -> Cluster manual (or predictions) from one location
• predict.py -> Predict the events for one dataset

Pre-trained models

Name	Pretraining data	Hours	Link to pretrained weights
Model Final (MF)	All annotated BPNS data	23.6h	Download
Model Base (MB)	Initial training set	1.5h	Download
Active Learning (AL)	Intial training set + 10 actively selected wavs	2.3h	Download
Random Selection (RS)	Intial training set + 10 randomly selected wavs	2.3h	Download

If you're interested in weights of other of the trained models (2 more models per row available), please contact [email protected]

References

This work includes some scripts from to extract the CAE features. https://gitlab.lis-lab.fr/paul.best/repertoire_embedder

Best P, Paris S, Glotin H, Marxer R (2023) Deep audio embeddings for vocalisation clustering. PLOS ONE 18(7): e0283396. https://doi.org/10.1371/journal.pone.0283396

This repository includes the filterbank.py script. Copyright (c) 2017 Jan Schlüter https://github.com/f0k/ismir2015

"Exploring Data Augmentation for Improved Singing Voice Detection with Neural Networks" by Jan Schlüter and Thomas Grill at the 16th International Society for Music Information Retrieval Conference (ISMIR 2015)