Ulisses x AI: Underwater Targets Detection using DIFAR Sonobuoys

Advanced Underwater Domain Awareness (UDA) is critical in the maritime domain and relies on analyzing acoustic signals from platforms like sonobuoys. Traditional manual analysis is increasingly challenged by data volume and modern naval stealth, especially in low signal-to-noise ratio (SNR) environments. This thesis introduces a novel approach, shifting from traditional vessel-type identification to a more granular, multi-label classification of predefined frequency ranges to achieve a detailed characterization of acoustic signatures. A primary contribution of this research is a novel synthetic dataset designed to simulate challenging passive sonar operations, characterized by an average SNR of -1.14 dB. The proposed solution adapts a state-of-the-art Audio Spectrogram Transformer (AST) for this unique multi-label task by modifying its classification head and interpolating positional embeddings to fit custom spectrogram dimensions. Empirical results demonstrate the model's effectiveness, achieving a micro-averaged F1 score of approximately 0.71: a greater than four-fold improvement over a random baseline. Explainable AI (XAI) analysis confirmed the model's interpretability, showing that its decisions were grounded in genuine acoustic features. Furthermore, the model's practical utility was validated by its successful generalization to real-world data. This research establishes the viability of the proposed approach as an integration tool to augment the situational awareness of human sonar operators in complex underwater environments.