| It is of great practical significance to comprehensively and accurately grasp the behavior pattern of ships and the spatial distribution of hot spots for ship manipulation at sea to ensure the safe navigation of ships and maintain the order of marine traffic.In recent years,the appearance of AIS(Automatic Identification System)system has pushed ocean management into the digital era,making it possible to monitor ship activities in real time.Digging deep into the rich information contained in massive AIS data to analyze the ship’s trajectory can capture the law of ship’s behavior pattern and the change of maritime traffic situation scientifically and objectively.However,at present,there are still some problems in the analysis of ship’s trajectory,such as inaccurate automatic acquisition of port position,insufficient exploitation of time-series characteristics of AIS data,etc.This paper takes dismantling the composition of port geographical elements,deeply exploring the application potential of AIS data,and facing the global demand for marine resources monitoring,ship navigation safety guarantee and maritime traffic order management,puts forward a set of ship’s trajectory analysis method of "automatic acquisition of port position-extraction of ship’s behavior characteristics-identification of ship’s behavior pattern-detection of maritime maneuvering hot spots".The main research contents and conclusions are as follows:(1)Research on port location acquisition based on ship terminal aggregation.High-precision automatic acquisition of port position is the key prerequisite for analyzing the ship trajectory in a specific port.In order to solve the problem that it is difficult to identify the port target by deep learning technology,this paper puts forward a set of port location acquisition method based on ship terminal aggregation from the perspective of the geographical elements of the port,which is used to automatically acquire the port location in the western Mediterranean.Firstly,the ship and dock targets are identified based on the remote sensing images of the coastline,and then the central points of the identification frame are clustered to obtain the spatial scope and location information of the port.Considering the performance differences of different models,the performance of three mainstream deep learning models,namely Faster R-CNN,YOLO v5 l and Retina Net,is tested by using remote sensing data sets to determine the optimal model.Accuracy and recall are used as evaluation indexes of detection model and port location acquisition method.The results show that YOLO v5 l is more suitable for detecting ships and wharf targets.This method obtains the spatial range and location information of 788 ports in the western Mediterranean,and the accuracy rate and recall rate reach 95.81% and 92.64% respectively,which fully verifies the feasibility of this method.(2)Research on ship behavior feature extraction based on spatio-temporal and kinematic features.According to the abundant information contained in AIS data,this paper proposes a set of ship behavior feature extraction methods based on spatialtemporal and kinematic features.This method mainly starts from the sub-trajectory segment composed of two consecutive trajectory points,deeply digs the attribute information in AIS,extracts nine features of the sub-trajectory segment,such as sailing time,distance,speed and acceleration,and determines the features with high correlation with ship behavior as the input data of the model by calculating Spearman coefficient,and constructs the data set of the ship behavior recognition model according to the determined features.The results show that the absolute values of Spearman correlation coefficients of six characteristics,such as sailing distance,time,speed,acceleration,sailing angle and steering rate,are between 0.21 and 0.65,which have high correlation with ship behavior and are used as input data of the model.In addition,four kinds of ship behaviors(acceleration,normal navigation,deceleration and berthing)are taken as the output of the model.Finally,a data set containing 46,014 behavioral characteristic data of 200 ships and corresponding 46,014 ship behavioral data was formed.(3)Research on ship behavior identification and hot spot detection of offshore maneuvering considering time series characteristics.According to the characteristics of two-dimensional characteristics and serialization of ship behavior characteristic data,this paper combines CNN and LSTM with attention mechanism,constructs a CNN-LSTMAT ship behavior recognition model considering time series characteristics,conducts a performance comparison experiment between the model and the traditional neural network model on the ship behavior data set,and applies the model proposed in this paper to the behavior recognition of two typical merchant ships in the western Mediterranean.Finally,based on the recognition results and the concept of geographical grid,the hot spots of ship manipulation in the western Mediterranean are detected and analyzed.The results show that the F1-Score of CNN-LSTMAT model is 89.45%,which is 4.09%,3.16%,1.66% and 0.99% higher than that of BP,CNN,LSTM and CNN-LSTM models respectively.The Recall of CNN-LSTMAT model is89.65%,which is 3.79%,2.32%,1.44% and 0.64% higher than that of BP,CNN,LSTM and CNN-LSTM models,respectively.The high-hot spots in the map of typical merchant-operated hot spots in the western Mediterranean generated by this method are also highly consistent with the spatial trend of the main routes in the western Mediterranean in the open source route data.The above results fully verify the high accuracy of CNN-LSTMAT model proposed in this paper and the reliability of the detection method of offshore maneuvering hot spots. |
