Detection and classification of whales calls using band-limited energy detection and transfer learning

Abstract

The Monterey Bay Aquarium Research Institute has been recording since July 2015 almost continuously at the Monterey Accelerated Research System (MARS) cabled observatory in Monterey Bay, California, USA. This long-term recording contains thousands of whale calls to help further our understanding of interannual, seasonal, and diel patterns. Here we report on our highly accurate detection and classification method developed to classify blue whale A, B, and D calls and Fin whale 20 Hz pulses. The foundation of the method is a computationally efficient and tuned decimation filter to convert the broadband hydrophone 128 kHz signal to 2 kHz which preserves the low-frequency signal and avoids any high-frequency aliasing. Detection is done using a band-limited-energy-detection filter to find potential calls in the decimated data. Spectrograms are then generated for potential calls and enhanced with local image normalization followed by smoothing by convolving in either time or frequency. Classification is done using the Google Inception v3 model with a transfer learning method. Overall, false positive rates are very low despite variability in whale call shape and background noise.