Intelligent Identification of Dolphin Whistle in Acoustic Signals by CNN models

This repository contains the software and example spectrograms associated with the study “Intelligent Identification of Dolphin Whistle in Acoustic Signals via Deep Learning” submitted to the journal “Scientific Reports”. The study presents a convolutional neural network a (CNN)-based approach for the automatic identification of bottlenose dolphin (Tursiops truncatus) whistles from underwater acoustic recordings. The method is based on the generation of spectrograms from audio signals and on the use of vertical Sobel filtering to enhance whistle contours before CNN classification. The material provided in this release includes a basic Python application for testing the trained CNN model and a set of 70 example spectrograms, consisting of 35 positive samples containing dolphin whistles and 35 negative samples without whistles. These files are intended to allow users to run and test the CNN-based approach described briefly below and in details in the associated manuscript.

The CNN is a lightweight sequential architecture designed for binary image classification. It processes grey-scale spectrograms as input and includes three convolutional layers with 32, 64, and 128 filters, respectively. Each convolutional layer uses a Rectified Linear Unit (ReLU) activation function and is followed by a 2 × 2 max-pooling layer to reduce spatial dimensionality and computational complexity. The resulting feature maps are flattened and passed to a dense layer with 128 units and ReLU activation. The final output layer consists of a single neuron with sigmoid activation, providing a binary classification output: class 1 indicates the presence of a dolphin whistle, while class 0 indicates no whistle.

The CNN model was trained on 6222 labeled spectrograms, including balanced positive and negative samples. The dataset was randomly divided into ten folds, and a 10-fold cross-validation procedure was adopted: in each iteration, nine folds were used for training and the remaining fold was used for validation. This ensured that validation spectrograms were never included in the training process. The training procedure was implemented in Python.

• https://www.irbim.cnr.it/progetto-dettagli/life-delfi/
• David Scaradozzi, Rocco De Marco, Laura Screpanti, and Francesco Di Nardo, Intelligent Identification of Dolphin Whistle in Acoustic Signals via Deep Learning, Scientific Reports, Submitted (mail: f.dinardo@univpm.it, d.scaradozzi@univpm.it)