An HMM-DNN-Based System for the Detection and Classification of Low-Frequency Acoustic Signals from Baleen Whales, Earthquakes, and Air Guns off Chile

Marine passive acoustic monitoring can be used to study biological, geophysical, and anthropogenic phenomena in the ocean. The wide range of characteristics from sounds sources makes simultaneous automatic detection classification these a significant challenge. Here, we propose single Hidden Markov Model-based system with Deep Neural Network (HMM-DNN) for low-frequency biological (baleen whales), geophysical (earthquakes), (air guns) sounds. Acoustic data were obtained Preparatory Commission Comprehensive Nuclear-Test-Ban Treaty Organization station off Juan Fernandez, Chile (station HA03) annotated by an analyst (498 h audio containing 30,873 events 19 different classes), then divided into training (60%), testing (20%), tuning (20%) subsets. Each frame was represented as observation vector through filterbank-based spectral feature extraction procedure. HMM-DNN procedure carried out discriminatively setting HMM states targets. A model Gaussian Mixtures Models (HMM-GMM) trained obtain initial set target states. Feature transformation based on Linear Discriminant Analysis Maximum Likelihood Transform also incorporated. displayed good capacity correctly detecting classifying events, high event-level accuracy (84.46%), weighted average sensitivity precision (89.54%). Event-level increased higher event signal-to-noise ratios. metrics per class showed that our generalized well most classes but performances best either had number exemplars (e.g., generally above 50) and/or signals low variability features, duration, energy levels. Fin whale Antarctic blue song air guns performed particularly well.