Untangling the Ocean Soundscape: A Deep Learning Approach to Acoustic Source Separation and Classification

The ocean soundscape is an active combination of biological, geophysical, and man-made acoustic sources. Precise separation and identification of these sources are key to useful ecological monitoring and marine biodiversity conservation. The suggested method utilizes a Conv1D-based framework motivated by the MixIT model for unsupervised time-domain source separation. Unlike usual models, the Dynamic MixIT model does not use ground-truth labels of the sources and can be capable of separating multiple overlapping sources adaptively in an unsupervised setting, which is extremely efficient in practical underwater acoustic scenarios. After source separation, a Graph Attention Fusion (GAF) classifier is proposed. This classifier constructs a graph representation of various acoustic features like MFCCs, chroma, and spectrograms and uses attention mechanisms to both observe temporal dynamics and inter-feature relationships. This results in better discrimination between bioacoustic signals and non-biological noise. Experimental evaluation against real-world ocean recordings confirms that the suggested framework outperforms standard BSS and CNN/LSTM-based classifiers (SOTA) both in source separation quality and classification accuracy. The integration of graph-based multi-feature fusion with dynamic source separation offers a scalable and robust passive acoustic monitoring solution for monoaural sensors. The method presents a new direction in underwater sound analysis that has wide implications in marine science and environmental management.