Smart Acoustic Monitoring for Automated Whale Song Classification and Migration Insight

Thesis Abstract

Whales rely on complex acoustic signaling for communication, navigation, and social structure, yet manual analysis of vast acoustic datasets is time-consuming and prone to subjectivity; this study addresses the need for scalable, objective whale song classification integrated with migratory inference to inform conservation and management. The aim is to develop an automated, ICT-driven framework that classifies whale vocalizations and extracts migration patterns from passive acoustic monitoring (PAM) data. Specific objectives are to (1) curate a labeled, multi-species whale vocalization dataset from long-term PAM deployments; (2) design and validate a hybrid deep learning pipeline combining convolutional neural networks (CNNs) and recurrent neural networks (RNNs) for robust species- and song-type recognition; (3) implement a feature-extraction module leveraging spectro-temporal receptive fields, mel-frequency cepstral coefficients (MFCCs), and continuous wavelet transforms (CWTs) to improve discrimination of overlapping calls; (4) develop an automated migration inference model integrating acoustic detections with environmental covariates (sea surface temperature, chlorophyll concentration, and bathymetry) using generalized additive models (GAMs) and hidden Markov models (HMMs) to infer migratory corridors and timing; (5) evaluate system performance against expert-human annotations and assess practical deployment in real-time alerting for maritime stakeholders. The methodology adopts a multiple-case, longitudinal design implemented on datasets from five offshore listening stations across temperate to subpolar regions, spanning a 24-month period and yielding approximately 3,200 labeled whale vocalization events. The population comprises baleen and toothed whale species with distinct song repertoires, including humpback, blue, fin, and minke whales, along with echolocation clicks from sperm whales. A stratified random sampling approach ensures representative coverage across seasons and acoustic environments. Data collection instruments include autonomous hydrophone arrays with 96 kHz sampling rate, synchronized time-stamping, and onboard pre-processing; supplementary metadata are sourced from satellite-derived environmental variables (MODIS-Aqua, Jason-3 Altimeter) and local meteorological observations. The data analysis workflow employs a supervised learning framework where the CNN-RNN hybrid classifier operates on log-magnitude spectrogram inputs, enhanced by attention mechanisms to focus on salient acoustic features. Feature extraction leverages MFCCs, Mel-scaled log spectrograms, and CWT-based scalograms to capture both tonal and broadband components. Model validation uses k-fold cross-validation (k=5) and temporal holdout tests to simulate real-world deployment, with performance metrics including precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). The migration component utilizes GAMs to relate detection rates to environmental covariates while HMMs infer state transitions corresponding to migratory phases. Ethical considerations address non-invasive monitoring, data privacy, and potential ecological disturbances, with all procedures aligning to institutional animal care, environmental impact assessment standards, and international maritime guidelines. Expected findings indicate high classification accuracy across target species (F1-scores >0.88 for humpback and blue whale song types; >0.80 for other species) and robust discrimination of song types in noisy gradient environments. The migration inference model is anticipated to reveal consistent migratory corridors and peak passage windows aligned with SST gradients and chlorophyll blooms, with predictive confidence intervals capturing seasonal variability. The integration of acoustic classification with environmental covariates is expected to yield actionable insights for maritime risk management, fisheries management, and protected-area delineation. The study contributes to knowledge by advancing a generalizable, end-to-end acoustic monitoring framework that couples state-of-the-art machine learning with ecological modelling to translate passive acoustic data into species-specific behavior and movement understanding, thereby filling gaps in long-term population connectivity assessments and real-time conservation decision support. In terms of theoretical contributions, the research tests the applicability of hybrid CNN-RNN architectures for bioacoustic signal processing in marine environments and extends theoretical models of animal movement by integrating acoustic evidence with GAM and HMM-based migratory state inference. Practical implications include an open-source toolkit for automated whale song classification and migration analysis, standardized evaluation benchmarks, and a decision-support dashboard for coastal managers and shipping authorities. Recommendations emphasize expanding multi-site deployments to capture broader geographic variability, incorporating transfer learning to adapt models to novel acoustic contexts, and establishing real-time alerting mechanisms to mitigate anthropogenic disturbances during critical migratory periods.