Classification of fishing vessel types using machine learning methods on vessel monitoring system data

Abstract: The oceans around the world have been heavily impacted by overfishing due to very intensive commercial fishing in recent times. A large number of fish stocks have already been fully exploited. Vessel Monitoring System has been put in place to regulate fishing vessels and enforce sustainable fisheries management. Data coming from such systems can be used for the detection of illegal, unregulated, and unreported fishing. In this thesis, we present various machine learning models for the classification of fishing trip trajectories. To train these models, we develop a trajectory segmentation algorithm to create trip trajectories out of raw data and design a graphical user interface for labeling the trip trajectories into fishing and non-fishing. We also examine the impact of the temporal resolution of the data. In conclusion, the CNN-Transformer network performed the best on the binary classification of two different fishing vessel types. During the project, we realized that segmentation of real trajectory data into trips poses many problems and presents the biggest obstacle. The experiment on the varying temporal resolution of the data showed that having a higher temporal resolution gives better modeling results but only to a certain point.