Research Of Fishery Prediction Using AIS Data

Fishery prediction is to determine the main activity area of marine fish according to all kinds of information,and the core of fishery prediction research is to identify the fishing behavior of ships.At present,the related research has given a number of solutions,which can be divided into two categories: 1.Methods based on traditional statistical learning.This kind of unsupervised method solves the problem that most AIS data have no navigation behavior label,but the prediction effect is worse than the supervised deep learning method.2.Method based on deep learning.This kind of method has good prediction effect,but its data acquisition depends on the active learning.Therefore,the cost of these deep learning solutions is the main disadvantage.In order to solve this problem,this thesis uses the idea of semi-supervised learning to find the balance between cost and effect.In this thesis,fuzzy entropy and Pu-learning methods are used to clean AIS data set and get enough reliable labels by utilize the differences between fishing and non-fishing tracks.Based on this data set,several methods commonly used in fishery prediction are used to verify.The results show that most of the deep learning models have good performance on this data set,and the data processing improves the accuracy of the model by 20%.It shows that the data processing method can partly replace the active learning.In the real ship track,because the frequency of ship AIS data is affected by satellite signal and delay,the time interval before and after AIS data is not equal.The time interval is a key factor to predict the continuous behavior of the navigation state.To solve this problem,this thesis proposes a variable time long short-term memory network(VTLSTM).VTLSTM designs two time gates for the time interval in the gating mechanism of LSTM,and extracts the time interval information between the track points through the time gates.The experimental results show that the VTLSTM model proposed in this thesis improves the accuracy by 1.5% compared with the traditional model.Finally,from the perspective of ship behavior,this thesis improves the prediction results of fishing vessel and fishing behavior to the prediction of fishery through statistical method.The result can provide new data and perspective for the research of marine ecological protection.