An investigation into computational methods for classifying fishing vessels to identify illegal, unreported and unregulated fishing activity

Illegal, unreported and unregulated (IUU) fishing undermines collective efforts to create a global model for sustainable fishing. Countering IUU fishing is an urgent priority given world population growth and increasing dependence on ocean-sourced food. This paper examines deep learning methods for the classification of fishing vessels with the intent to determine illicit fishing operations. This is achieved through supervised learning with highly irregular time series data in the form of signals from the automatic identification system (AIS). To deal with the intermittent frequency of AIS signals, two separate approaches have been followed: feature engineering with zero padding and linear interpolation. Fundamentally, this work suggests there exists a distinct relationship between vessel movement patterns and method of fishing. Two neural network architectures: stacked bidirectional GRUs and 1D CNNs with residual connection blocks, are leveraged on each data pipeline to produce four sets of results. The GRU with feature engineering achieves 95% accuracy despite severe class imbalance in the large datasets. The system can classify a vessel’s fishing method over 24 hours in real-time to monitor behaviour in marine protected areas and detect gear discrepancies, safeguarding fish stocks in the process.