das-finwhale-vocalization

Description

Distributed Acoustic Sensing (DAS) is a technology that enables continuous, real-time measurements along the entire length of a fiber optic cable. When there are acoustic disturbances in the ocean, strain is produced along the cable which creates patterns in time and distance dimensions that differ based on the source of the disturbance. Our goal is to identify all instances of fin whale vocalizations, cluster them together, and visualize these results to identify patterns in the behavioral.

Configuration

This project requires an installation of miniconda which can be found here. After the installation, follow the instructions below in a terminal:

1. Create the conda environment using the environment.yml file provided and activate it:

   > conda env create -f environment.yml
   > conda activate ooi_dev_2
   
   

2. Configure module discovery using the setup.py file (this will help us find modules written in sub-diretories):

   pip install -e .
   

Data

For this project, all data extracts have been backed up to Google Drive here. This folder also contains the final poster that was used for the poster presentation.

We describe some of the extracts that can be found:

• south_dist_chunked_10000to47495_dx5_fs100_bpf14to35Hz.zip contains the 256x256 chunked images along the cable distance with each images covering 15s of time . We used this data in the (b) series of notebooks. The data has been downsampled to 100Hz, with 5 channels skipped, bandpass filtered from 14 to 35 Hz and channel number 10,000 to 47,495 used
• south_ch_10000to47499_fs100_bpf14to35Hz.zip contains the same as above with the difference that no channels were skipped and the images are 128x128
• processed_clip1k_fs100_bpf1435 contains minute long samples of the entire cable length with the first 5k channels skipped. Used mostly for exploration purposes
• processed_256x256_Hunter contains 256x256 minute long extracts of a shorter range of channels. Used mostly for exploration purposes

Project Usage Guide

Based on the questions that you are looking to answer, there are different ways of accessing and utilizing the code. We provide some insight into the use cases for the various Notebooks below:

• Note:
  • Notebooks labelled (a) use full length of the cable in each input image
  • Notebooks labelled (b) and (c) use chunked cable images (10 chunks over the entire cable) with (b) using latent embedding size as 16 and (c) using embedding size as 32 in the autoencoder notebooks.
  • We have completed the project right till GMM clustering for (b). The (a) and (c) notebooks are retained in the repository in case any further work is desired to be done on different data and embedding sizes.

If you are interested in the patterns of Fin Whale vocalizations in the chunked November 2021 data set:

• 05_b_Heatmap Visualization.ipynb contains code to display a heat map over distance and time domain of the data for embedding length 16 of chunked data over the distance dimension

If you are interested in visualizing the clusters of the data:

• 04_b_GMM Clustering Full Dataset 16 Embedded.ipynb creates and visualizes the clusters for the entire data with the cable length broken into 10 chunks and a latent embedding size of 16.
• Note the comments in the code and follow their instructions based on whether you desire to load in current clustering results or create your own.

If you are interested in training the Autoencoder:

• 03_b_PyTorch Autoencoder Training- all data distance chunked - embedding 16.ipynb has code to train the Autoencoder model whose embedded layer is the basis for our clustering algorithm

If you are interested in clustering and training an Autoencoder:

• 02_b_Data Acquisition - cable length chunked.ipynb has code to pull and prepare a dataset from Azure
• 02_Data Acquisition and Processing - from piweb.ipynb has code to pull and process data from the piweb website which was given to us before azure acces

Finally,

• 01_Data Processing Exploration.ipynb contains some data exploration code

Acknowledgements

We want to thank Dr Shima Abadi, Dr William Wilcock and John Ragland from the Ocean Data Lab for their guidance and support throughout the project, and Dr Megan Hazen and the MSDS staff for facilitating this venture.