Real-time detection and classification of underwater soundscape signals via a convolutional autoencoder

Abstract

Underwater sounds come with a wide range of frequency components and with a wide range of time intervals. For example, individual killer whale ( Orcinus orca) calls are typically one or two seconds long. Humpback whale (Megaptera novaeangliae) calls range from second long moans to half-hour songs. Speed boats can pass a hydrophone in seconds to minutes. Large ships create signals above background for many minutes. Here fixed 100x100 dimension spectrograms having exponentially increasing time windows of 3, 15, 75, 375, 1875, 9375 seconds are calculated and classified in real time via an autoencoder used as an acoustic pattern detector followed by a supervised convolutional neural network with softmax output. The autoencoder was trained on underwater sound (∼1 Tb) at Orcasound Lab, San Juan Island, WA. The system is designed to classify incoming real-time signals to one or more of twelve different classes (background, killer whale, Southern Resident, Biggs killer whale, humpback whale, breaking waves, kayak paddles, speedboat, displacement boat, large ship, large ship with shaft rub, large ship with repeating sound pattern). The design and efficacy of this detection/classification system will be reported.