Analysis Of Fishing Behavior Of Fishing Boats Based On AIS Data

Fishing from ancient times to the present has accompanied human beings,with the change of the times,many things have undergone earth-shaking changes,but the proposition that human beings need materials to survive is eternal,and fisheries provide a large amount of material for human survival.In recent years,the serious food security problem has highlighted the importance of fisheries,and although the total output of China’s fisheries has increased year by year,there is illegal fishing and excessive utilization of fishery resources,resulting in a decline in inshore fish species and the depletion of resources.In the information age,emerging technologies such as artificial intelligence can be used to better monitor and manage fishery resources,among which accurately identifying fishing status and accurately identifying fishing vessel operation types are the keys to information monitoring and management.Based on the automatic identification System data of the vessel,the paper proposes the two propositions of identifying the fishing status of fishing vessels and the type of fishing vessel operation,and the main work and main innovations of the paper are as follows:(1)Due to the non-disclosure of logbook or VMS data,the difficulty of data collection,the inaccuracy of manual records,and the low temporal and spatial resolution,there are usually not enough such open data to study the fishing status and types of fishing vessels.Thanks to the openness,comprehensiveness and high recording frequency of AIS data,it is very suitable for research.Therefore,the original data used in this study is AIS data with fishing tags,and the original data is disordered.This article starts with the processing of the original data,the purpose of which is to filter and construct a single fishing behavior of each boat from the massive disordered data.The track sequence is selected,and the track sequence of research significance is selected to form a dataset for studying the fishing behavior of fishing boats.The data processing process includes exception processing,data grouping,derivative feature calculation,voyage division,single voyage trajectory interpolation,etc.In the process of constructing the data set,on the basis of comprehensive analysis of the original data,the aim is not to destroy the original data of a large area and not to generate a large area of new data,and try to make the data set more standard and authoritative.(2)The practical problem of identifying the fishing status of fishing vessels.For the track sequence of a fishing boat,it is necessary to determine the fishing point and the navigation point,which is essentially a sequence labeling problem.The recurrent neural network variant Long-Short Term Memory(LSTM)is often used to process sequence data.Conditions in the field of machine learning Conditional Random Field(CRF)can also be used for sequence labeling.The former uses the powerful nonlinear fitting ability of neural networks to calculate labels,while the latter uses probability factors to constrain the relationship between labels.Therefore,this paper proposes an LSTM-CRF model combining LSTM and CRF to solve the problem of fishing state identification.Finally,a labeling accuracy of more than 92% was achieved on the selfbuilt dataset.(3)The identification of the fishing status of a fishing vessel depends on the type of fishing vessel.The fourth chapter of this paper puts forward the practical problem of distinguishing the types of fishing boats,and analyzes the spatial information of fishing boats with different types of operations.From the principle of fishing boats,it is demonstrated that trawl nets and gillnets account for more than 72% of fishing boats.The difference of the job trajectories is put forward to image the trajectory information,and then identify the two job types by image processing.In the experiment,image datasets of these two types of fishing boats were generated,and the differences between the two types of trajectories visible to the naked eye were also found,mainly in the smoothness of the trajectory and the turning frequency.Then three classic networks Alex Net,VGG-16,Res Net-34 were used for comparative experiments,in which the Res Net-34 network was used to achieve 92% classification accuracy.Finally,a method of circular convolution(Cycle-CNN,C-CNN)is proposed to further improve the classification accuracy of Res Net-34.The core of circular convolution is to enable the feature map to achieve multi-position information fusion in one layer.It is verified by experiments that the Res Net-34 network experiment using circular convolution can improve the accuracy rate by 2%,so the accuracy rate finally reaches 94%.