AI in marine science: Classifying vessels and co-occurance with mammals using CNNs based on underwater acoustics

Marine Protected Areas (MPAs) have been established to safeguard coastal regions against
increasing human activities. Within these areas, human interactions with the environment
require regulation. However, effectively monitoring these regions presents significant
challenges, often leading to illegal activities. An effective way to approach the presence of
human activity, is through ocean acoustics. By using machine learning techniques, object
detection and classification can be performed to lay the groundwork for monitoring
underwater acoustics.
A machine learning model is trained based on AIS data and passive acoustic hydrophone
recordings. Although AIS data’s main purpose is to avoid collisions (in real time), the
historical data contains information about the voyage information (e.g., type of vessel) and
the position (e.g., longitude). The acoustic data is based on over 100 days of acoustic data
obtained from two North Sea stations. By combining the recordings with their relative
distance to by-passing boats, and their type, a database was established to detect and
classify marine vessels. From this dataset, a convolutional neural network (CNN) was
created. It aims to predict the distance, activity, and type of vessel based on their sound
signature.
The DCLDE dataset serves two primary purposes. Firstly, it validates the performance of the
created model by predicting vessel distance based on recordings. Secondly, it predicts the
co-occurrence of mammals and ships. In future steps, the model could be retrained using
the DCLDE dataset to further enhance its performance.