Sinatra-Blue/embedding-viewer

Interactive 3D UMAP of self-supervised acoustic embeddings from Gabon marine PAM data # Marine Acoustic Embeddings — Interactive 3D UMAP An interactive 3D visualisation of self-supervised acoustic embeddings learned from marine passive acoustic monitoring (PAM) data. Part of a PhD project at the University of Exeter developing foundation-model and self-supervised learning approaches for scalable marine ecological monitoring. ## Live viewer **View the interactive 3D embedding** Rotate, zoom, and hover over individual points. Toggle between two colour modes using the buttons above the plot: - **Colour by cluster** — HDBSCAN cluster assignments in the embedding space, showing the fine-grained acoustic structure the model has discovered without any labels. - **Colour by recorder** — segments coloured by which of the two VAALCO offshore hydrophones (VAALCO01 or VAALCO02) they came from. The two colours are mixed throughout the space, showing that recorder-hardware differences have been effectively removed. ## About the data The viewer shows a 30,000-segment subsample of a larger corpus of approximately 1.2 million two-second segments extracted from raw 96 kHz passive acoustic recordings collected in Gabonese offshore waters. Segments were embedded into a 256-dimensional space using a self-supervised hierarchical encoder (five-layer 1D convolutional front-end followed by a four-layer transformer aggregator), trained with a temporal contrastive objective on unlabelled data. The embeddings were then projected to three dimensions via UMAP for visualisation. ## Methodological context Marine passive acoustic monitoring datasets are typically collected by hydrophones whose slowly-varying hardware characteristics can dominate the learned representation in a standard contrastive-learning pipeline. Early results in this project showed this failure mode directly: the two Gabonese hydrophones formed two spatially separated regions of the embedding space, and 136 clusters at their boundaries could be identified as recorder-specific artefacts at harmonic frac …