A RECURRENT NEURAL NETWORK MODEL for DETECTING FISHING GEAR PATTERNS

Abstract

Estimation of fishing efforts is the key metric for sustainable ocean management. Previous studies have been proposed to detect fishing activities based on analysis of vessel trajectory from Vessel Monitoring System (VMS). However, identification of fishing activity without prior knowledge related to fishing gears may cause detection failure because individual gears of fishing possess specific movement patterns. It is desirable to identify vessel movements made by different fishing gears for accurately detecting fishing events. In this work, we propose a novel method that recognizes a VMS trajectory corresponding to fishing gear types by encoding sequences of GPS points with Recurrent Neural Networks (RNNs). Firstly, we segment a route trajectory using an unsupervised segmentation scheme. After that, each extracted segment is encoded into a semantic space to train a neural network model for identifying a fishing ship of a specific gear. We also demonstrate that RNNs with feature embedding can leverage the discriminative power of classifier. We conduct experiments on real trajectory data of three fishing gear types, including trawl, purse-seine and falling net, collected from a VMS database of the Thailand Command Center for Combating Illegal Fishing (CCCIF). Our experimental results demonstrate embedded bidirectional gate recurrent units achieves over 90% classification accuracy compared with state-of-the-art methods and other RNN models.