Path Prediction for Fishing Boats Using Attention-Based Bidirectional Gated Recurrent Unit

Obtaining trajectory predictions for fishing boats in complex and unpredictable seas is essential for preventing ship collisions. In this study, we propose a deep-learning model that predicts the paths of fishing boats. We choose offshore trap fishery as the fishery type for which the movement paths of fishing boats are predicted. It is because offshore trap fishing boats sail in sudden changes of 90 degrees or more. Piecewise cubic Hermite interpolation polynomial (PCHIP) is used to interpolate regular-interval data. We focus on extracting feature variables that consider the impact of daytime and nighttime conditions on fishing operations. Trajectory windows are constructed using a sliding-window approach to create input datasets for deep learning. The framework employed is based on the sequence-to-sequence (Seq2Seq) architecture with an attention mechanism. The experimental results demonstrate the superiority of Seq2Seq with attention over Seq2Seq without attention. The performance of our proposed method has increased by at least 7.0%, 12.0% on average, compared with the GRU and LSTM. The technology developed in this study is expected to prevent collision accidents between autonomous ships and fishing boats in the future. In addition, because it is possible to predict the future path of the fishing boat, this technology can be used in the decision-making system of vessel traffic service operators.