OrcaCNN: Detecting and classifying killer whales from passive acoustic data

Passive acoustic data has the potential to greatly increase our knowledge of the presence, habitat, and routines of elusive killer whales by providing time-series of vocalizations from individuals and pods on hourly to seasonal time-scales. However, extracting useful data and information about killer whales from these time-series is difficult due to the sheer size of the data. Specifically, it is difficult to quickly and accurately identify killer whale vocalizations in acoustic data sets because manually detecting calls by a trained human listener is intractable and legacy software detection methods are inaccurate and still require substantial human time for verification. Here we present an automated detection and classification pipeline capable of quickly detecting the presence of killer whales in acoustic data and subsequently identifying the source pod of the detected vocalization trained on data gathered in the Northeast Pacific off the coast of Southeast Alaska, USA. The pipeline consists of two convolutional neural networks developed with Keras using TensorFlow as the backend and trained on GoogleCloud. The first model detects the presence of killer whale calls in the audio samples with 95% accuracy. The second model identifies the pod source of the vocalization with 60% accuracy. We also present the architecture of the models, provide comparisons to legacy results, and demonstrate the deployment of the model pipeline as a web-based application.