Repository used for the paper A public benchmarking hydroacoustic dataset for geophonic signals detection task

This repository gives the code used to develop, train and evaluate detection models on OHASISBIO-2018, OHASISBIO-2020 and HYDROMOMAR-2013 datasets. The project is made of several modules used to import and process the data and to create learning models. To use these modules, several notebooks are proposed. In addition to this, a visual interface is proposed to enable an easy visualization of the data.

Requirements

This project was developed using Python 3.10.12. Libraries requirements are given in data/requirements.txt. To run the code, please install Python 3.10.12, open a terminal at the root of the project and run the following:

pip3 install -r data/requirements.txt

The project dependencies and the GUI were also tested with Python 3.10.14. In case this version is required, consider running the following:

pip3 install -r data/requirements-3.10.14.txt

How to run

Data

To run the code of this project, acoustic data are necessary. Consider downloading the OHASISBIO-2018 dataset, formatted with .wav files named with a timestamp. The OHASISBIO-2020 and HYDROMOMAR datasets are also available for download. To download them, using the GNU tool wget, one can run the following in a terminal:

wget -r --no-parent https://data-dataref.ifremer.fr/hydro-ac-passive-public/

Visualizing the data

A graphical software is made available with this project. This enables to quickly take a look at the data, by first choosing a directory (e.g. one of the three datasets obtained with the download) to inspect and then by exploring it with spectrograms. The following gif shows an example.

To run this software, simply open a terminal in the src directory and run the following:

python3 -m GUI.main

Some shortcuts are available :

• + or - enable to respectively zoom in or zoom out (dividing/multiplying the window length by two, keeping the same center).
• left mouse click enable to center a spectrogram on the given time instant.
• left arrow or right arrow enable to respectively move backward or forward in time, by half the window length.
• * or / enable to respectively decrease or increase the maximum displayed frequency.
• up arrow or down arrow enable to respectively decrease or increase the minimum and maximum displayed frequency together.
• shift+enter enables to format the other spectrograms as the one we focus on.
• enter enables to listen to the sound contained in the current window, with a x20 speedup for a better experience.

Creating datasets

To create a dataset for training/evaluation, consider running a notebook. The directory src/notebooks/dataset_building is purposely made for this task. Open the relevant notebook, modify the parameters according to the needs and run the cells.

Training and evaluating models

Directories src/notebooks/detection/training and src/notebooks/detection/eval were made to train and run the models used in the paper. Open the relevant notebook, modify the parameters according to the needs and run the cells. The directory src/notebooks/detection/eval/figures_generation enables to use the data produced by the evaluation notebook to create figures.