November 2024 Status update

As COP28 draws near, this is what Aqoustics have been up to.

Aqoustics Status update

Identifying boat and fish sounds in underwater environments presents a unique set of challenges. Underwater sounds can be complex and include a wide range of frequencies, from the low rumble of boat engines to the clicks and chirps of fish.

To tackle this problem we use Convolutional Neural Networks, with the ability to process spatial audio and patterns, it is ideal for classifying boats and fish.

We have connected with two groups who have shared their data with us and allowed us to start classifying boat and fish sounds.

University of Bristol's Boat Sound Classification:

The researchers from the University of Bristol, Clemency White and Cathy Hobbs, have contributed valuable data to our project. They provided us with tagged underwater boat recordings from Miami, which they used for understanding the correlation between boat sounds and the underwater ecosystem's health. To classify these sounds accurately, we've implemented a CNN-based neural network tailored to this specific task.

CNN Architecture for Boat Sound Classification:

  • Input Layer: The input to the network consists of spectrograms representing the audio recordings. Spectrograms break down the audio signal into time-frequency components, making it suitable for CNNs.
  • Convolutional Layers: Multiple convolutional layers are used to learn hierarchical features from the spectrograms. These layers apply filters to detect patterns in different scales, capturing both fine and coarse-grained details in the boat sounds.
  • Flatten Layer: The flattened output from the pooling layers is passed to a fully connected layer.
  • Fully Connected Layers: These layers learn to map the extracted features to the final output classes, which in this case is boat sounds.
  • Output Layer: The output layer uses a sigmoid activation function to produce class probabilities, allowing us to determine the likelihood of each audio segment containing a boat sound.
    With our current configuration we have achieved an accuracy of 95%, we are currently waiting on tagged data from the PhD students to compare human and machine tagging.

Mary Shodipo's Fish Sound Classification:

Mary Shodipo, a field ecologist and acoustician based in the Philippines, has generously provided us with fish audio data tagged by PhD students from the University of Miami. This data is currently being used to manually classify fish species, an essential effort to categorise the diverse range of fish species in underwater environments.
The current approach is using categories of distinct sounds and then later matching them to fish.
Using the Neural Network for the Boat Classifications with a few tweaks, we had a network that could identify if there are fish sounds in a clip, but not which specific sound. With a new batch of data that we are about to receive from Mary, we aim to make a new network that will classify the distinct fish sounds that they have provided us with, with the aim of helping them in their monumental task of classifying fish species.

Both of these neural network architectures leverage the power of CNNs to automatically learn and extract relevant features from spectrograms, enabling the accurate classification of boat and fish sounds. With the collaboration of these two groups and the implementation of these advanced models, we aim to contribute significantly to our understanding of underwater ecosystems and the conservation efforts that rely on precise sound identification.

More articles