Bowhead Whale Deep Learning System: Guarding Arctic Marine Ecology with Autoencoders

This project is a PyTorch-based deep learning autoencoder system for detecting, classifying, and localizing bowhead whale calls in the Alaskan Beaufort Sea. Combining MATLAB preprocessing and deep learning techniques, it provides an automated acoustic monitoring solution to support marine biodiversity conservation and ecological research. Developed by Aaron Thode (Scripps Institution of Oceanography, University of California, San Diego), the source code is available on GitHub (link: https://github.com/AaronThode/BowheadDeepLearningMATLAB) and was released on June 15, 2026.

Bowhead whales are among the Arctic's endemic and longest-lived mammals, relying on complex calls for communication. However, climate change and human activities threaten their survival. Traditional monitoring relies on manual analysis of hydrophone data, which is time-consuming and labor-intensive, making it difficult to handle large-scale continuous data from over 30 hydrophones in the Alaskan Beaufort Sea. There is an urgent need for automated tools.

Data Preprocessing (MATLAB)：Convert raw audio to spectrograms, extract directional information, assemble labeled datasets, and normalize them. Key scripts include master_create_datasets.m (generates spectrograms) and master_cluster_analysis.m (clustering), etc. Deep Learning Model (PyTorch)：The core is a convolutional autoencoder (unsupervised learning), available in LD16 (16-dimensional latent space, fast iteration) and LD32 (32-dimensional, complex features) versions. Use t-SNE/UMAP dimensionality reduction to visualize latent representations and assist in cluster analysis. Application Flow：Training → Feature Extraction → Clustering → Visual Validation → Performance Evaluation.

1. Unsupervised Feature Learning：No need for large labeled datasets, reducing costs；2. Multi-scale Feature Capture：Convolutional architecture is suitable for complex frequency modulation patterns of bowhead whale calls；3. Interpretability：Understand model features through latent space visualization, facilitating scientific applications.

Ecological Conservation：Track bowhead whale migration and habitat changes, assess impacts of climate change and human activities；Methodology Promotion：Can be applied to monitoring other marine mammals like humpback whales and dolphins；Technical Demonstration：Combine traditional signal processing (MATLAB) with modern deep learning (PyTorch), providing an interdisciplinary example.

The project includes an interactive tool called "Whale Call Theremin", which plays 7 types of bowhead whale calls (about 70,000 samples) in the browser and visualizes frequency contours in real time. Using the Web Audio API to synthesize audio, it helps the public understand the diversity of whale calls and provides an interesting entry point for science popularization education.

Current Limitations：Dependent on training data quality, susceptible to artificial noise interference, requires high-performance computing resources；Future Directions：Develop real-time edge computing versions, expand to multiple species, use transfer learning to accelerate adaptation to new environments, and improve models with active learning.

Bowhead Deep Learning is a model example of applying deep learning to marine ecological conservation, providing an "acoustic window" for wildlife monitoring. It represents an interdisciplinary collaboration model (oceanographers, acoustic experts, ML engineers) and is of great significance for protecting vulnerable Arctic life in the context of climate change.