and Machine Learning

Abstract

Accurate identification of ships is vital to ensure safe maritime activities. Research methods for the classification of ship types mostly use traditional radar recognition and optical recognition, but these methods all have their limitations. However, in the case of ship identification based on Automatic Identification System (AIS) data, not only is it less afected by the weather, but also static information and dynamic information can be utilized. So Automatic Identification System (AIS) data applications have been actively researched in ship identification as a more advanced and reliable method than traditional methods. However, incorrect AIS information may be transmitted due to an operator's mistake. In addition, some ships intentionally change AIS data information, such as ship type, to hide abnormal operations or illegal activities. In order to solve this problem, it is necessary to devise a new method to classify ship types correctly. So A ship-type classification scheme based on a ship navigating trajectory with Automatic Identification System (AIS) data is proposed to solve this problem. First, to acquire training data, historical AIS data provided by the Danish Maritime Authority have been converted into ship trajectories based on the Maritime Mobile Service Identities (MMSI), including corresponding ship types. As one of the main challenges in handling raw datasets is cleaning them to ensure the removal of invalid data, pre-processing is applied. Next, we extracted 39 features, including behavioral, geographic properties, and measurement of ship appearance characteristics. We especially proposed new features that could represent the shape of the overall trajectory using ink features designed for sketch recognition. Based on the extracted features, several benchmark classification algorithms (i.e., Decision Tree, Random Forest) are trained to classify four types of ships: Fishing, Passenger, Tanker, and Cargo. Finally, we check which features are valuable for recognizing ship types and which models can implement good performance in ship classification through performance analysis. The results demonstrate that the ink features designed for sketch recognition could express essential characteristics of ship trajectories and could be used for ship classification. Furthermore, Random Forest performs better than other classifiers in the classification of AIS data, and the classification accuracy of the four types of ships could reach 84.05% with a 39-dimensional feature vector.

Keywords

• Learning
• AIS data
• Ship classification
• Machine Learning
• Ship trajectory

1. Introduction

1.1.

Motivation

The international shipping industry is responsible for the carriage of around 90% of world trade[ 1 ]. Ship as the critical transportation tool on the vast ocean always raises much attention. In early 2022, the total fleet of seagoing merchant vessels amounted to 102,899 ships of 100 gross tons and above, equivalent to 2,199,107 thousand deadweight tons of capacity[ 2 ]. Furthermore, the world had an estimated 4.1 million fishing ships in 2020[ 3 ]. Along with the rapid increase in vessels, the maritime trafic environment has become more complex, and the possibility of maritime trafic accidents has increased. Maritime trafic accidents are complex and might result in the loss of human and irreversible economic damage[ 4 ]. Over Australia (T. Hammond) (T. Hammond) in reported shipping losses[ 1 ]. Among them, developing systems capable of monitoring vessel activities is one of the most crucial factors in strengthening navigational safety and security, such as preventing collisions and detecting unreported ships.

Traditional maritime navigation primarily relied on charts, watches, and radars and was judged by the sailor’s long sailing experience. However, it is dificult to quickly and accurately identify numerous ships because it is limited in vision and radar coverage and provides only incomplete information, such as the speed and direction of movement of ships. However, the Automatic Identification System (AIS), which appeared with the development of communication technology, played an essential role in navigational safety, such as collision avoidance and navigation assistance. In the case of vessel identification based on AIS, not only is it less afected by the weather but also static information (vessel’s name, dimensions, vessel’s type) and dynamic information (vessel’s position, speed) can be utilized. In particular, since the type of © 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License ship is closely related to the vessel’s maneuverability, it is crucial information that sailors and Vessel Trafic Service (VTS) controllers must grasp in advance to predict cover a much larger area than AIS data, which is limited the following behavior of ships in a limited area. These to the range of ship-based receivers. However, there are unique characteristics that can afect their behavior can also several disadvantages, such as satellite images havbe represented such as : ing a limited resolution, which can make it challenging to identify smaller vessels, and high-resolution satellite • Speed : Diferent ship types may have diferent images are not always readily available, and acquiring propulsion systems and speeds, afecting their them can be expensive. ability to navigate in a given area, especially in Given these situations, developing a proper ship-type crowded waterways or areas with shallow water. recognizer is vital to solving the missing or tampering • Turn radius : Larger ships with a deep draft and of ship-type information in AIS data. Some papers have slow speeds may have a larger turning radius devised new methods for ship trajectory and type classithan smaller, more maneuverable ones. ifcation based on many AIS dataset. For example, they • Stoppage time : Some ship types, such as con- create a ship’s trajectory image based on AIS data and intainer ships and bulk carriers, may take longer put it into Convolutional Neural Networks (CNN), or they to stop due to their size and weight compared to transform the ship’s trajectory data into graph data and smaller ships, such as tugs and fishing ships. use Graph Neural Networks (GNN) to classify ship types.

However, the following deep learning-based model not By knowing the type of ship and its characteristics from a only makes it dificult to check what process is taken distance, people can quickly recognize the situation, take to identify a ship but also requires a large amount of precautionary measures in a limited area, such as a port training data. In contrast, generating meaningful feaor shipping lane, and improve the safety and eficiency of tures directly for machine learning algorithm operation maritime operations. This information can also be used involves a lot of efort but is more eficient at classifying to optimize trafic management, design navigational aids, objects once they can be found. Some papers obtained and prepare for emergencies. significant classification results using kinetic and geo

In particular, since 2004, the International Maritime graphical information obtained from AIS data as features Organization (IMO) has mandated the installation of AIS for ship identification. However, it is necessary to create on international passenger ships and ships of 300 tons or more meaningful features to improve the classification more to strengthen maritime safety and security. In ad- accuracy of ships. Thus, in this paper, we use various dition, many countries and intergovernmental agencies, ink features used in sketch recognition to create new feasuch as regional fisheries management organizations, are tures practical for vessel identification and compare and creating AIS requirements within their waters. How- analyze the performance of machine learning algorithms ever, with the widespread use of AIS, the reliability in based on these. accuracy of AIS information has been addressed in recent years. In particular, since some static information 1.2. Summary of Solution provided by AIS is directly entered by the ship owner, incorrect and missing information may be provided in- We can take advantage of the Danish Maritime Authortentionally or unintentionally. Furthermore, this may ity AIS data to decide the type of a particular ship. We cause a loss of reliability for the provided information. A assume that ships cruising for diferent purposes would study by Abbas[ 5 ] reveals errors related to the type of have diferent paths crossing the same area during a simship in the AIS data. According to the ”VTS-based AIS ilar time. Given AIS data and analyzing such patterns study” by Abbas, some ships had no available ship type from the data can reveal the type of the ship. Some of the and were defined as ”vessels” rather than a specific ship diferences can be taken from reasoning. For example, type. Meanwhile, researchers and VTS operators were passenger ships, tankers, and cargo ships are likely to unhappy with some of the observed vessel types. There- cruise along a straight path under good weather since fore, research on ship-type identification is needed to they are moving from one destination to another. Howsolve the missing or tampering of ship-type information ever, fishing ships are looking for schools of fish and are in AIS information. One of the other approaches to solve less likely to move along a straight line. There can be this problem is using other types of sensors and technolo- other diferences that could be more obvious. Analyzgies, such as satellite imagery and drones, to supplement ing the AIS data may help us find such diferences. To and improve AIS data for maritime trafic management. achieve the goal, we will preprocess the AIS data so that There are several advantages to using satellite images for the data falsely collected would not afect our classificaship classification compared to AIS data. Satellite images tion. For the data, each CSV file in raw AIS data contains can provide information about ships that do not have AIS the timestamps and locations of a ship. The Maritime or have turned of their AIS, which is a common practice Mobile Service Identities (MMSI) and the type of the ship for some vessels to avoid detection. Moreover, it can are also included. The type of ships other than Cargo, Passenger, Tanker, and Fishing will be deleted. Then we the problem of missing ship type information, it is essenwill build our features and apply diferent classifiers to tial to make various eforts to identify the type of ship the feature set. using other information from AIS data. In particular, as a large amount of trajectory data are generated in realtime based on the AIS system, Ship classification based 2. Related Work on trajectory data can compensate for the deficiency of traditional radar and optical identification.

With the explosion of sensors and GPS-enabled devices, Sanchez et al. [ 16 ] extracted features from spatiotemresearch on trajectory recognition, which is analyzing poral data that represent the trajectories of ships and objects over position and time data, has generated con- used the Support Vector Machine(SVM) and decision tree siderable interest. This research is often performed using to classify ship trajectories into either fishing or nonmachine learning algorithms and is vital in various appli- ifshing ships. Sheng et al. [ 17 ] partitioned each trajeccations, such as object tracking, activity recognition and tory into three basic movement patterns: anchored-of, behavior analysis. As most ships carry an AIS system as turning, and straight-sailing, extracted 17 trajectory feaa matter of law, the data collected from it can be used in tures based on it, and classified fishing and cargo ship by various fields, such as ship navigation route prediction, using a logistic regression model. Li et al. [ 18 ] converted trajectory classification and anomaly detection in ship ship trajectory data into graph data and used Graph Neubehavior[ 6, 7, 8, 9, 10, 11 ]. ral Network(GNN) to classify four types of ships. They could recognize fishing ships, passenger ships, tankers, 2.1. Ship Type Recognition Approaches and containers with an accuracy of 82.7%. Yang et al. [ 19 ] generated ship trajectory images containing operating states such as static, standard navigation, and maneuvering. They used Convolutional Neural Networks(CNNs) to identify eight types of ships from ship trajectory images and achieved an accuracy of 87.5% with the optimal configuration of CNNs. Wang et al. [ 20 ] used static information in AIS data such as width, length, and draught and identified five types of ships with an accuracy of 86.14%.

Yan et al. [ 21 ] extracted some ship appearance and behavior characteristics and classified the five types of ships with an accuracy of 92.7% using the Random Forest model.

Machine learning has been applied successfully in ship classification tasks, providing a promising solution to challenges such as missing or tampering with ship-type information in AIS data. However, the accuracy of machine learning algorithms for ship classification can be afected by factors such as the quality and availability of data, the extraction of valuable features, and the choice of algorithms and models. Researchers have also explored integrating AIS data with other sources of information, such as satellite imagery and radar data, to enhance the accuracy and reliability of ship classification[ 22, 23 ].

Among many applications, accurately identifying the

type of ship is particularly important to ensure the safety and eficiency of maritime operations. This is because knowing the ship’s type and characteristics can help predict its behavior in a limited area and can be used to design navigational aids, and prepare for emergencies.

Along with this need, many eforts have been made to identify ships operating at sea. One of the ship-type classification approaches studied so far is image-based ship-type recognition. In recent years, with the advent of Convolutional Neural Networks(CNNs), which do not require human supervision to identify essential features, many attempts have been made to develop a feasible ship classification system using satellite or aerial imagery[ 12, 13, 14, 15 ]. However, optical sensors have the disadvantage that they cannot be used at night or in adverse weather conditions. And ship classification in Synthetic Aperture Radar (SAR) images still has challenging to identify detailed sub-lists such as cargo and tankers beyond sorting out general categories such as ships and airplanes from diferent types of vehicles[ 15 ].

Also, the number of labeled samples in the SAR domain is limited.

Unlike image-based systems that attempt ship classi- 2.2. Features Used For Recognition ifcation based on images, there are various studies to classify ships using AIS data. Research in this area focuses on developing algorithms and models to classify ships based on their attributes, such as size, speed, using AIS data. Some common approaches in ship classification with AIS data include using machine learning algorithms, such as decision trees, random forests, support vector machines, and neural networks. Although the AIS data already includes information on the type of ship, as mentioned in the introduction, considering human error and Modern sensors and digital devices enable the collection of a large amount of data from moving objects. For example, smartphones, smartwatches, and wildlife with tracking tags generate timestamped position data. Even the handwriting or hand-drawn sketches of users with the digital pen are positional coordinates data created on digital paper over time. Based on the acquired data, we analyze various hidden patterns and use them to solve new problems, such as trajectory and digital ink recognition. Many features have been proposed to solve many diferent recognition problems. As diferent problems Intuitively, we understand that particular features dehave their characteristic, it is essential to add valuable veloped for one domain problem can be valuable for some features to improve the recognition rate for their specific problems in diferent domains. However, until now, reproblem. Jones et al.[ 24 ] grouped trajectory features ac- search papers have yet to extract features from the shape cording to their characters to make it easier to select a of the overall trajectory for ship classification. So our relevant set for a specific problem. These categories for system builds on this kind of work by seeking to identify features can be represented as : ship types with features from a diferent domain, such as sketch recognition. Even if it is a feature for sketch recognition, if it can reflect the characteristics of a general trajectory well, the corresponding feature can be applied to ship type classification. • Kinematic Features: Kinematic Features describe an object’s motion independent of the forces that cause it moves, such as total distance traveled, average speed, and maximum altitude. • Temporal Features: Temporal Features described when some event of interest took place, 3. Methodology such as start and end time, duration, and time when nearest to a fixed point or region. The main objective of this work is to develop and eval• Geospatial Features: Geospatial Features de- uate a ship type classifier using machine learning. In scribe where the trajectory is observed, such as other words, the goal of this project is to address the the place where first seen, the destination point, following two questions: 1) What features are valuable and the nearest distance to a fixed point or region. for recognizing the ship types? 2) How can we achieve • Shape Features: Shape Features describe some reasonable ship type recognition using diferent classifiaspects of the trajectory’s geometry, such as con- cation algorithms?. vex hull area and divergence from a given shape.

3.1. Data Source

AIS data-based ship type recognition has tried applying various trajectory features. Some researchers have The data used in this work were historical AIS data from already proposed various features to classify ship types, the Danish Maritime Authority [ 33 ], with a time distribuincluding kinematic, temporal, geospatial, and geometric tion of 5 days in November 2022. This data includes AIS features. Kraus et al. [ 25 ] extracted geographical charac- information obtained from the coast of Denmark since teristics, such as the distance to the nearest coastline, and 2006, and the size of each file stored per day reaches temporal features, such as if the trajectory starts/ends at approximately 1.5GB. Therefore, it can be used as a sufinight. Yan et al. [ 21 ] extracted various ship appearance cient amount of data to create a trajectory for this project. characteristics, including length, width, and shape com- Considering the subsequent trajectory creation process, plex and Sanchez et al. [ 16 ] extracted kinetic features the information from the AIS data needed includes MMSI, such as average and maximum of course variation from Timestamp, Ship type, Latitude, Longitude, Width, and segments of the trajectory. Length.

Although it is a diferent domain, many features have been proposed for sketch recognition, and they can be divided into two types: gesture-based features and geo- 3.2. Pre-processing metric features[ 26 ]. The first type describes how a sketch The quality of AIS data is crucial for trajectory analywas drawn and used to classify input strokes contain- sis and an essential factor for classification model pering x,y points, and time values into a set of pre-defined formance. We used the Pandas library for data pregestures[ 27 ]. For example, Dean Rubine[ 28 ] proposed processing to remove invalid values in AIS data before thirteen features used for basic shape and gesture recog- converting raw AIS data to trajectories. For example, the nition, and Long et al. [ 29 ] added nine new features to data such as the latitude exceeding 90 degrees or the lonRubine’s existing set. Unlike the first one, the second gitude exceeding 180 degrees was cleared because they type describes the object’s shape and arrangement, focus- were out of range. ing on what the sketch looks like[ 27 ]. Paulson et al.[ 30 ] proposed new features, the Normalized Distance between 3.3. Trajectory Creation Direction Extremes (NDDE) and Direction Change Ratio (DCR), which are suitable for classifying polylines After obtaining the pre-processed AIS data, the next step and curved strokes. Furthermore, Blagojevie et al.[ 31 ] is to generate the trajectory of each ship and label the tracomposed a comprehensive library of 114 ink features jectory according to the type of ship. A trajectory can be from previous work in sketch recognition and developed defined as consecutive coordinates of the ship, and each a taxonomy of feature types such as curvature, density, trajectory coordinate is a tuple comprising a position(latand direction. itude & longitude) and a timestamp. Because the ship trajectory has various lengths, shorter trajectories of less 3.5. Feature Extraction than 3 hours of operation are excluded to ensure the trajectory carries enough information for feature extraction. One of the goals of this work is to find features for recogMoreover, most of the trajectories consist of more than nizing the ship types based on trajectory. Some papers thousands of coordinates, so the computational cost for have already proposed various features to classify ship feature extraction is expensive. A compression algorithm type, including geographical, behavioral, and Geometric is applied to reduces the number of coordinates while features. The author of [ 25 ] extracted geographical charpreserving the character of the shape of the trajectory acteristics, such as the distance to the coast, to classify [ 34 ]. We used MovingPandas, a Python library for han- the type of ship. In [ 21 ], the author extracted various ship dling movement data based on Pandas and GeoPandas, to appearance characteristics, including length, width, and create and generalize the trajectories [ 35 ]. Based on this, shape complex. Moreover, the author of [ 17 ] extracted samples of the trajectory for each ship type is shown in trajectory features based on the fundamental movement ifgure 2. The color shown in the legend expresses relative patterns and divided them into three categories such as time, and the initially acquired coordinates are set to 0, global, straight-sailing, and turning features. However, and the color gradually changes to green as time passes. the most challenging thing is to discover additional pracAccording to the unique MMSI, CSV files, including the tical features that can describe the overall behavior of a information required for trajectory generation, were cre- ship along with the previously used features. However, ated and saved in the folder classified by ship type where until now, research papers have yet to extract features the label matches. Finally, we obtained 1,298 trajectories from the shape of the overall trajectory for ship classifibased on unique MMSI from AIS data. Table 1 shows the cation. For this purpose, we propose new features that number of trajectories for each of the four ship types. could measure an essential characteristic of ship trajectories using ink features designed for sketch recognition.

In [ 31 ], the author composed a comprehensive library Table 1 of 114 ink features from previous work in sketch recogNumber of created trajectories by the four types of ship nition and developed a taxonomy of feature types. We Ship Type Cargo Fishing Passenger Tanker Total have selected several features from these lists that will Quantity 568 197 289 244 1298 be useful for the ship type classification problem. Finally, feature extraction was performed by dividing it into three categories. The first was the trajectory shape feature, which generated 21 features, including Rubine Features, 3.4. Trajectory Data Exploration some Long Features, and DCR(Direction Change Ratio). The second category used latitude and longitude, such as mean, max, and standard deviation, as a geographic features. Lastly, we extracted some features from the measurement of ship appearance characteristics to improve the classification performance[ 36 ]. Finally, we created a 39-dimensional feature vector for each trajectory. The detailed feature list is shown in Table 2.

In order to create meaningful features for ship type classification, it is crucial to identify the corresponding trajectory pattern. Trajectory coordinates are plotted using a diferent color to indicate ship type using the Kelper.Gl, designed for geospatial data analysis. Figure 3 shows some trajectory patterns, such as passenger ships using common routes and fishing ships tending to cluster many points within a limited area intensively.

(a) Trajectory Visualization for Cargo (b) Trajectory Visualization for Fishing (c) Trajectory Visualization for Passenger (d) Trajectory Visualization for Tanker

4. Results and Discussion

Extracted features are tested on classification algorithms

with 10-fold cross-validation provided by the WEKA [ 37 ]. We experimented with five diferent classifiers for our evaluation: J48, Random Forest, Random Tree, Logistic, and Multilayer Perceptron. We used standard classifier evaluation metrics such as Accuracy, Precision, Recall, and F-measures to measure the model’s efectiveness in classifying the four ship types. Accuracy, average precision, average recall, and average F1 score are calculated for each classifier, and the results are shown in Table 3.

According to Table 3, our model has 84% accuracy for the four-ship types altogether and outperforms the GNN model[18]. Although the dataset difers, it suggests our model successfully matches ships to the cargo, tanker,

In weighted average, random forest outperformed other fishing, and passenger category based on ship trajecclassifiers by nearly 4%. tory. Regarding each ship type, the precision, recall, and F-score is 94.6%, 91.7%, and 93.1% respectively for passenTable 4 ger ships; 95.1%, 99%, and 97% respective for fishing ships; Result of the four types of ships with Random Forest 78.6%, 45.1%, and 57.3% respective for tankers and 74.4%, 91.7% and 84% respectively for cargo ships. Though all Ship Types Precision(%) Recall(%) F1-Score(%) indicators for fishing and passenger ships exceed 90%,

Cargo 74.4 91.7 84 it is also alarming that the precision for cargo type and Passenger 94.6 91.7 93.1 all indicators for tanker type are significantly worse. AcFishing 95.1 99 97 cording to Table 4, the major problem is that tankers are Tanker 78.6 45.1 57.3 labeled cargo type. Reviewing the trajectory in figure 2a and figure 2d, the trajectory of cargo and tanker are too similar for our model, which is based on shape features, to recognize their diference. Figure 4a and Figure 4b provide a more intuitive view. Those two trajectories are not a simple straight line or polyline. They make (a) A trajectory of Cargo ship (b) A trajectory of Tanker (c) A trajectory of Fishing 5. Conclusion and Future Work similar turns, and the terminals of their trajectories are similar but diferent. They are even traveling at a similar speed. The color of their trajectories changes similarly, This project proposed a ship classification method with indicating the relative time of their trips. Going over all Danish water’s AIS data considering the overall ship’s trajectories of cargo ships and tankers, it appears that the behavior characteristics to solve the missing and tamperdiference between a cargo ship and another particular ing of ship type information in AIS data. Firstly, trajectanker can have a relatively high chance of being smaller tory shape features and geographic features are extracted than its diference with another cargo ship. The similar- from many AIS data from diferent types of ships. After ity between the trajectories of cargo ships and tankers that, this project used various classification algorithms is reasonable since, in some sense, tanker ships are also such as Random Forest and Decision Tree to compare cargo ships, except that their cargo is oil. To recognize the performance. Results and Discussion showed that them, either the training data is expanded so that extra the Random Forest performs better than other classifiers shape features can be observed to distinguish cargo ships in the classification of AIS data. The classification accuand tankers, or more features need to be applied, and racy of the four types of ships could reach 84.05% with that solution falls out of the scope of this study. 39-dimensional feature vectors. In particular, in the case

By the same reasoning, the exceptional outcome in of fishing and passenger ships, the precision was 0.951 ifshing ships (99% recall and 97% F1-score) suggests that and 0.946, respectively, and very high results were conifshing ships cruise very diferently from the other three ifrmed. This confirmed that the ink features designed for ship types. The most noticeable diference could be that sketch recognition could express essential characteristics the goal of fishing is not traveling from one place to of ship trajectories and be used for ship classification. In another, which is the objective for all three other types. the future, research can be carried out considering the folFigure 4c above is a typical case. lowing directions : (1) focusing on extracting additional features and applying feature selection to improve the performance of AIS data ship classification. In particular, ifnding practical features distinguishing cargo and tanker ships will significantly improve the overall classification model performance. (2) testing the classifier with much larger data volumes to check the classifier’s scalability. (3) Expand the AIS data from Danish waters to diferent regions worldwide to validate its potential in dealing with maritime trafic situations.

References

[1] A. G. C. S. AGCS , Safety and shipping review 2021 , 2021 .

[2] U. Nations , The review of maritime transport 2022 , 2022 .

[3] Food , A. Organization(FAO ), The state of world ifsheries and aquaculture 2022 , 2022 .

[4] V. L. N. R. C.-E. I. T. R. Dominguez-Péry , C. , Reducing maritime accidents in ships by tackling human error: a bibliometric review and research agenda , Journal of Shipping and Trade ( 2021 ). doi:h t t p s : / / d o i . o r g / 1 0 . 1 1 8 6 / s 4 1 0 7 2 - 0 2 1 - 0 0 0 9 8 - y .

[5] A. Harati Mokhtari , A. Wall , P. Brooks , J. Wang , Automatic identification system (ais): Data reliability and human error implications , Journal of Navigation 60 ( 2007 ) 373 - 389 . doi:1 0 . 1 0 1 7 / S 0 3 7 3 4 6 3 3 0 7 0 0 4 2 9 8 .

[6] J.-Y. Oh , H.-J. Kim , S. - K. Park, Detection of ship movement anomaly using ais data: a study , Journal of Navigation and port research 42 ( 2018 ) 277 - 282 .

[7] T. Guo , L. Xie, Research on ship trajectory classification based on a deep convolutional neural network , Journal of Marine Science and Engineering 10 ( 2022 ) 568 .

[8] X. Chen , Y. Liu , K. Achuthan , X. Zhang, A ship movement classification based on automatic identiifcation system (ais) data using convolutional neural network , Ocean Engineering 218 ( 2020 ) 108182 .

[9] Y. Li , H. Ren , Visual analysis of vessel behaviour based on trajectory data: A case study of the yangtze river estuary , ISPRS International Journal of Geo-Information 11 ( 2022 ) 244 .

[10] P. Sampath , Trajectory analysis using automatic identification system (ais) in new zealand waters , 2012 .

[11] J.-G. Lee , J. Han , X. Li , H. Gonzalez , Traclass: trajectory classification using hierarchical region-based and trajectory-based clustering , Proceedings of the VLDB Endowment 1 ( 2008 ) 1081 - 1094 .

[12] Y. Wang , C. Wang , H. Zhang, Ship classification in high-resolution sar images using deep learning of small datasets , Sensors 18 ( 2018 ) 2929 .

[13] Z. Shao , J. Wang , L. Deng , X. Huang , T. Lu , R. Zhang , X. Lv , Q. Ding , Z. Wang , Glsd: The global largescale ship database and baseline evaluations , arXiv preprint arXiv:2106.02773 ( 2021 ).

[14] J. Li , C. Qu , J. Shao , Ship detection in sar images based on an improved faster r-cnn, in: 2017 SAR in Big Data Era: Models, Methods and Applications (BIGSARDATA) , IEEE, 2017 , pp. 1 - 6 .

[15] Y. Xu , H. Lang , Ship classification in sar images with geometric transfer metric learning , IEEE Transactions on Geoscience and Remote Sensing 59 ( 2020 ) 6799 - 6813 .

[16] D. Sánchez Pedroche , D. Amigo , J. García , J. M. Molina , Architecture for trajectory-based fishing ship classification with ais data , Sensors 20 ( 2020 ) 3782 .

[17] K. Sheng , Z. Liu , D. Zhou , A. He , C. Feng, Research on ship classification based on trajectory features , The Journal of Navigation 71 ( 2018 ) 100 - 116 .

[18] T. Li , H. Xu , W. Zeng , Ship classification method for massive ais trajectories based on gnn , in: Journal of Physics: Conference Series , volume 2025 , IOP Publishing , 2021 , p. 012024 .

[19] T. Yang , X. Wang , Z. Liu , Ship type recognition based on ship navigating trajectory and convolutional neural network , Journal of Marine Science and Engineering 10 ( 2022 ) 84 .

[20] Y. Wang , L. Yang , X. Song , X. Li , Ship classification based on random forest using static information from ais data , in: Journal of Physics: Conference Series , volume 2113 , IOP Publishing , 2021 , p. 012072 .

[21] Z. Yan , X. Song , H. Zhong , L. Yang , Y. Wang , Ship classification and anomaly detection based on spaceborne ais data considering behavior characteristics , Sensors 22 ( 2022 ) 7713 .

[22] Z. Yan , X. Song , L. Yang , Y. Wang , Ship classification in synthetic aperture radar images based on multiple classifiers ensemble learning and automatic identification system data transfer learning , Remote Sensing 14 ( 2022 ) 5288 .

[23] H. Lang , S. Wu , Y. Xu , Ship classification in sar images improved by ais knowledge transfer , IEEE Geoscience and Remote Sensing Letters 15 ( 2018 ) 439 - 443 .

[24] J. L. Jones , B. D. Newton , K. L. Wisniewski , A. T. Wilson, M. J. Ginaldi , C. A. Waddell , K. Goss , K. J. Ward , M. D. Rintoul , Large-Scale Trajectory Analysis via Feature Vectors , Technical Report , Sandia National Lab. (SNL-NM), Albuquerque , NM (United States) , 2021 .

[25] P. Kraus , C. Mohrdieck , F. Schwenker , Ship classification based on trajectory data with machinelearning methods , in: 2018 19th International Radar Symposium (IRS) , IEEE, 2018 , pp. 1 - 10 .

[26] P. Corey , T. Hammond, Gladder: Combining gesture and geometric sketch recognition ., in: AAAI , volume 8 , 2008 , pp. 1788 - 1789 .

[27] B. Paulson , P. Rajan , P. Davalos , R. Gutierrez-Osuna , T. Hammond , What!? ! no rubine features?: Using geometric-based features to produce normalized confidence values for sketch recognition , in: HCC workshop: sketch tools for diagramming, 2008 , pp. 57 - 63 .

[28] D. Rubine , Specifying gestures by example , ACM SIGGRAPH computer graphics 25 ( 1991 ) 329 - 337 .

[29] A. C. Long Jr , J. A. Landay , L. A. Rowe , J. Michiels , Visual similarity of pen gestures , in: Proceedings of the SIGCHI conference on Human Factors in Computing Systems , 2000 , pp. 360 - 367 .

[30] B. Paulson , T. Hammond, Paleosketch: accurate primitive sketch recognition and beautification , in: Proceedings of the 13th international conference on Intelligent user interfaces , 2008 , pp. 1 - 10 .

[31] R. Blagojevic , S. H.-H. Chang , B. Plimmer , The power of automatic feature selection: Rubine on steroids ., SBIM 10 ( 2010 ) 79 - 86 .

[32] Google , Google map of danish waters, (n .d.). URL: https://www.google.com/maps/@57.1858914, 10 .0813199,6z, retrieved February 22 , 2023 .

[33] D. M. Authority , Ais data, (n .d.). URL: https: //dma.dk/safety-at-sea/navigational-information/ ais-data, accessed November 3 , 2022 .

[34] N. Meratnia , R. A. de By, Spatiotemporal compression techniques for moving point objects , in: Advances in Database Technology-EDBT 2004: 9th International Conference on Extending Database Technology , Heraklion, Crete, Greece, March 14 -18, 2004 9, Springer, 2004 , pp. 765 - 782 .

[35] A. Graser , Movingpandas: eficient structures for movement data in python , GIForum 1 ( 2019 ) 54 - 68 .

[36] H. Lang , S. Wu , Y. Xu , Ship classification in sar images improved by ais knowledge transfer , IEEE Geoscience and Remote Sensing Letters 15 ( 2018 ) 439 - 443 . doi:1 0 . 1 1 0 9 / L G R S . 2 0 1 8 . 2 7 9 2 6 8 3 .

[37] M. A. H. Eibe Frank , Ian H. Witten , The WEKA Workbench. Online Appendix for ”Data Mining: Practical Machine Learning Tools and Techniques”, fourth ed., Morgan Kaufmann, 2016 .