Detection of Fishing Activities from Vessel Trajectories

Aida Ashrafi, Bjørnar Tessem, Katja Enberg

Abstract: This work is part of a design science project where the aim is to develop Machine Learning (ML) tools for analyzing tracks of fishing vessels. The ML models can potentially be used to automatically analyse Automatic Identification System (AIS) data for ships to identify fishing activity. Creating such technology is dependent on having labeled data, but the vast amounts of AIS data produced every day do not include any labels about the activities. We propose a labeling method based on verified heuristics, where we use an auxiliary source of data to label training data. In an evaluation, a series of tests have been done on the labeled data using deep learning architectures such as Long Short-Term Memory (LSTM), Recurrent Neural Network (RNN), 1D Convolutional Neural Network (1D CNN), and Fully Connected Neural Network (FCNN). The data consists of AIS data and daily fishing activity reports from Norwegian waters with a focus on bottom trawlers. Accuracy is higher than or equal to 87% for all deep learning models. Example applications of the trained models show how they can be used in a practical setting to identify likely unreported fishing activities.