Application and comparison of neural nets for marine mammal call classification

Recent work has successfully applied a linear matched filter to calls made by Bowhead whales recorded off the coast of Alaska in frequency‐time (spectrogram space) to detect and classify marine mammal calls. This method relies on an empirical matrix weighting for the matched‐filter coefficients. A neural net, trained on spectrogram estimates as the feature vector space, offers two advantages over this approach; (a) the equivalent weighting matrix is determined by training and may coverage to a more optimal solution and (b) the response of a neural net is nonlinear and can embody more sophisticated responses. A simple three‐layer feedforward neural net is ideally suited to this application and has been implemented on 204 calls, of which 163 were used for training and 31 kept as ‘‘unseen’’ test data. The neutral net was configured to identify both whale calls and other mammal interference. The success rate including failures in both estimates on training data was 88%. The combined false‐positive and false‐negative whale detection errors on unseen data was only 7%, which compares very favorably with any other known method. The neural net approach is compared with the matched filter and the role of the hidden neurons and equivalent weighting matrix are discussed. [Work supported by the Office of Naval Research.]