Transfer learning for distance classification of marine vessels using underwater sound

Marine environments are increasingly affected by human activities, which generate underwater noise as a by-product. Acoustic data from these environments can offer valuable insights for tracking human activity and improving the monitoring of sensitive areas, such as Marine Protected Areas (MPAs) and offshore wind farms. This study presents a convolutional neural network (CNN) trained to classify vessel distances from passive acoustic recordings. We constructed an open-source, diverse dataset by integrating 116 days of acoustic data from two stations in the Belgian North Sea with Automatic Identification System (AIS) data. The CNN was trained to classify acoustic clips into discrete distance bins, representing the proximity of the nearest vessel. Our results demonstrate that the model can effectively distinguish between distance categories using underwater sound alone, confirming the feasibility of passive acoustic monitoring for vessel activity. This technology provides an innovative approach to enhance MPA oversight and represents a first step in a promising pathway for conservation efforts.