Marine mammal call discrimination using artificial neural networks

Area of Science:

• Bioacoustics
• Machine Learning
• Animal Communication

Background:

• Previous methods for detecting bowhead whale song notes include linear spectrogram correlator filters (SCF) and hidden Markov models.
• These methods rely on empirical weighting matrices, which may not capture the full complexity of whale vocalizations.

Purpose of the Study:

• To investigate the effectiveness of a three-layer feed-forward artificial neural network (ANN) for detecting bowhead whale song notes.
• To compare the performance of the ANN against the SCF and other established methods.
• To explore the interpretability of the ANN by analyzing its hidden neurons.

Main Methods:

• A three-layer feed-forward ANN was trained on 1475 bowhead whale sounds (54% training, 46% testing).
• The ANN's performance was evaluated based on its error rate in identifying song notes.
• The function of the ANN's hidden neurons was analyzed in relation to spectrographic features of the whale calls.

Main Results:

• The trained ANN achieved a 1.5% error rate, a twofold improvement over previous methods.
• The ANN demonstrated superior performance compared to the SCF.
• The study showed that ANN hidden neurons could be interrogated to understand the learned operating paradigm, allowing for controlled training and reduced training time.

Conclusions:

• Artificial neural networks offer a more optimal and sophisticated approach to analyzing complex bioacoustic signals like whale songs.
• The interpretability of ANNs through hidden neuron analysis provides insights into the detection process and allows for method optimization.
• ANNs represent a significant advancement in the automated detection and analysis of marine mammal vocalizations.