A Comparative Study of Fishing Activity Detection Approaches in Maritime Surveillance

Maritime authorities (MA) must track fishing vessels to ensure that fishing activities are limited to permitted areas. If illegal fishing is suspected, resources must be allocated to intercept and inspect the vessels. Thus, a false flag by the MA is costly, so it is important to use accurate detection methods. We compare the accuracy and computational time of the main approaches for detecting fishing activities described in the literature, using the Global Fishing Watch (GFW) dataset. We find that Long Short-Term Memory (LSTM) neural networks achieves an optimal accuracy of 1.00, while the random forest approach comes second with an accuracy of 0.87. Given the high cost of mistakes for MA, we conclude that the LSTM's high computational cost is worthwhile.