MARINA: Maritime Acoustic Recognition and Identification with Novel Algorithms

We present MARINA, underwater soundscape awareness in the 10Hz-10kHz range for understanding and monitoring the undersea acoustic environment. Soundscape awareness is a challenging problem due to the environment’s complexity and the limited amount of labeled training data available for automated methods, which rely on abundant amounts of labeled data for success. Instead, we propose to exploit the abundance of publicly available unlabeled acoustic data by creating acoustic foundation models, large deep learning models that learn the structure of unlabeled data. These models demonstrate robust feature repre- sentations useful for downstream tasks like few-shot learning, without requiring abundantly labeled train- ing data. Such models can perform semantic search, allowing an operator to query a new sound against a database of known sounds, obtaining semantically similar matches. Our proposed model design extends current methods by including amplitude and phase representations. Additionally, our model design lever- ages classical signal processing techniques known for enhancing signal quality. This guides our model to learn robust feature representation as well as develop signal enhancement capabilities for countering noise in underwater environments. We also optimize our models for low-SWaP hardware, balancing size and accuracy, and evaluate the effectiveness of our approach using few-shot learning and ablation studies.