A team from University College London employs a convolutional neural network pretrained on YouTube audio to extract embeddings from minute-long coral reef recordings. They combine unsupervised clustering and supervised random forests to classify habitat types and individual sites, showcasing a scalable passive acoustic monitoring workflow.

Key points

• Pretrained VGGish CNN processes 0.96-sec log-mel spectrograms into 128-D embeddings per one-minute recording.
• Compound index combines eight acoustic metrics across three frequency bands into a 44-D feature vector.
• Trained CNN (T-CNN) fine-tunes VGGish architecture on reef audio for direct classification.
• UMAP reduces embeddings to 2D or 10D for visualization and affinity propagation clustering.
• Random forest classifiers use P-CNN and index embeddings to predict habitat types and site identity with up to 100% accuracy.
• Datasets span three biogeographic locations: Indonesia, Australia, French Polynesia.

Why it matters: By integrating pretrained AI models with passive acoustic data, this work paves the way for low-cost, scalable monitoring of marine ecosystems. It demonstrates that transfer learning can unlock ecological insights without extensive manual annotation or specialized hardware.