Fin Whale Detection in Ocean Acoustic Recordings Using a Deep Neural Network

Underwater acoustic recordings provide an efficient way to track marine animals. The analysis of the large volume of recordings available requires automatic call detectors because manual processing is impractical. In this study, we developed a Deep Neural Network (DNN) to detect vocalizations of endangered fin whales in acoustic recordings. The DNN was trained and evaluated on a year's worth of Southern Ocean recordings and expert annotations. Spectrogram images were computed for the classification by the DNN and an enhanced training algorithm was developed to optimize the detection performance. The proposed DNN achieves acceptable performance for automatic fin whale detection for practical applications. The number of false detections is an order of magnitude lower than a traditional energy-based detector. The DNN performance was also compared to a statistical model for data from the same location. The DNN performs similarly to other DNNs that were applied to right whale calls, but it has lower complexity. Future work includes improving the robustness of the DNN using recordings from other locations, detecting whale choruses and other species of whales using the sameDNN.