BlueTIDE 2026

BlueTIDE 2026 | August 27, 2026 | Newport, Rhode Island

Kitware is excited to be part of this year’s BlueTIDE 2026 event. Hosted by Polaris Tech Bridge in collaboration with the U.S. Navy and the Naval Undersea Warfare Center Division Newport, BlueTIDE is an annual technology demonstration event. It brings together government, industry, academia, and investors to accelerate the development of maritime and dual-use technologies.

At BlueTIDE, we’ll highlight our research in underwater acoustic analysis, including technologies for passive acoustic monitoring and underwater acoustic target recognition. Our work addresses challenges in underwater acoustic machine learning, from developing foundation models with limited labeled data to improving model evaluation through benchmark datasets and open source software for acoustic data analysis.

Learning from Unlabeled Acoustic Data

As part of the Maritime Acoustic Recognition and Identification with Novel Algorithms (MARINA) project, funded by DARPA, Kitware is developing AI foundation models for underwater acoustic environments.

The underwater domain presents several challenges for machine learning, including complex propagation effects, expensive data collection, and limited labeled datasets. To address these challenges, Kitware applies hierarchical pretraining. The approach begins with out-of-domain data and progresses through increasingly relevant acoustic domains.

By leveraging publicly available unlabeled acoustic data, hierarchical pretraining supports few-shot learning for underwater acoustic classification. The resulting feature representations support downstream tasks including semantic search and underwater sound classification.

Kitware develops and optimizes foundation models for different computing platforms. Model distillation and pruning create smaller models that meet size, weight, and power requirements for deployment on resource-constrained underwater vehicles.

The work also expands Kitware’s open source software capabilities for acoustic data analysis. These capabilities include sound classification using established taxonomies, semantic search, automated labeling, and visualization for underwater acoustic data.

In addition to developing foundation models for underwater acoustic analysis, Kitware is advancing benchmark datasets that reduce data leakage and support more reliable evaluation of underwater acoustic machine learning models.

Reducing Data Leakage in Model Evaluation

UniqueShip is a benchmark dataset for underwater acoustic target recognition that is planned for open release and will be presented at the upcoming OCEANS 2026 conference in Monterey, CA. It includes more than 2,474 hours of audio from 4,222 unique vessel sources across 11 vessel classes and a background class.

Existing ship-classification datasets can contain data leakage when recordings from the same vessel appear in both the training and test sets. This can produce falsely optimistic test results and significant performance declines during real-world deployment.

UniqueShip minimizes this leakage by separating individual vessels, identified by Maritime Mobile Service Identity, across data partitions. It also groups ambient or noise-dominated recordings by calendar date and provides explicit partitions that follow these constraints.

The reported results demonstrate the effect of leakage on model evaluation. The highest class-averaged accuracy was 66.8% without leakage and 80.5% when leakage was introduced. Across the evaluated architectures and input representations, data leakage increased class-averaged accuracy by 9% to 14%. These findings show that leakage can distort estimates of model generalizability.

Connect with Our Team

If you would like to schedule a demo of our technology or speak with our technical developers, please contact our team.