North Atlantic Right Whale call detection with very deep convolutional neural networks

Abstract

Ship collision is one of the main threats to the North Atlantic right whale, which is in danger of extinction. One popular way to reduce the collision is monitoring for the occurrences of whales by detecting their sounds on data recordings. We explore the application of very deep convolutional neural network in this detection problem. For feature extraction, we compute Mel-frequency cepstral coefficients (MFCCs) along with their first and second temporal derivatives, and Fourier-transform-based filter-banks for all sound clips. MFCCs were calculated with Hamming window, and the filter-banks were calculated in range of 50—650 Hz, and include 72 coefficients, distributed on mel-scale, for each of the 97 time steps. For classifier modeling method, we apply the very deep convolutional Neural Network (CNN) in our task. The CNN architecture s 22 layers, which consists of alternating convolutional layer and pooling layer, while the last layers are full-connected neural network. Dropout is used in our fully connected layers with a rate of 0.4. By using the data provided by the Cornell University Whale Detection data, our model provides area under the ROC curve (AUC) performance of 0.985, which achieves the state-of-the-art performance presently.