Underwater radiated noise (URN) from ships is an increasingly critical issue in sustainable maritime transport, with significant implications for vessel efficiency and decarbonization efforts. Traditional noise prediction methods, including empirical models like those of Ross and RANDI, are limited by their reliance on outdated datasets and oversimplified assumptions, while physics-based methods, although theoretically robust, demand substantial computational resources and often fail to generalize across diverse ship types. Recent validation studies reveal that even functional regression models, when applied to vessel noise datasets, account for only 25-50% of noise variability. This underscores the complexity of accurately capturing real-world URN patterns and highlights the need for more advanced, data-driven solutions to address this challenge effectively.

To address these limitations, this study leverages machine learning to predict underwater noise emissions, using a dataset from the Marine Acoustic Research Station project. The dataset includes over 2,000 acoustic signatures from vessels in the St. Lawrence shipping corridor. Preprocessing focused on frequencies above 500 Hz. Machine learning Algorithms like Random Forest, XGBoost, AdaBoost, CatBoost, and neural networks were applied, with CatBoost achieving the best performance (RMSE = 7).

Feature importance analysis identified ship speed and draft as the most influential factors, followed by key vessel characteristics such as length, width, and age. These insights enable the development of scalable models for real-time noise monitoring using open-source databases like AIS. This research provides a data-driven framework for mitigating vessel noise, supporting sustainable maritime transport and ecosystem preservation.