A marine traffic network mining method and system based on open-source AIS data

Abstract

The application provides a marine traffic network mining method and system based on open source AIS data, and the method comprises the following steps: pre-processing AIS data; identifying and classifying the berthing track by using geographic information and motion information, extracting the ship berthing area containing the berthing function semantic distribution based on the berthing track; dividing the space of the concerned area by screening the track points at the ship turning point, and extracting the ship turning area by using a genetic algorithm; the collection of the ship berthing area and the ship turning area is used as a feature area; a directed graph is constructed according to the feature area, noise data is removed, and a route boundary and a route representation are generated by a kernel density estimation and a triangular segmentation method. The advantages of the application are as follows: high extraction accuracy, low artificial supervision cost and rich regularity semantics; the mining efficiency of the turning area can be greatly accelerated, and the mining accuracy is not affected by the data density distribution.

Description

Technical Field

[0001] This invention belongs to the field of maritime traffic pattern mining, specifically involving a method and system for mining maritime traffic networks based on open-source AIS data. Background Technology

[0002] The ocean is the most extensive area on Earth, covering approximately 71% of the Earth's surface. In recent years, against the backdrop of economic globalization, the demand for international trade transportation has been increasing, making maritime transport the most important mode of international trade, accounting for 90% of global trade. To improve navigation safety and efficiency, maritime agencies in various countries are continuously striving to enhance their Maritime Situational Awareness (MSA) capabilities and have conducted extensive research in maritime intelligent services, such as trajectory prediction, anomaly detection, and route planning. Maritime traffic network mining is a key topic for understanding ship maritime traffic behavior and patterns, which can support various maritime intelligent service applications and is of great significance for maritime intelligence acquisition.

[0003] The International Maritime Organization (IMO) began promoting Automatic Identification System (AIS) equipment at the beginning of this century, stipulating that all merchant ships with a gross tonnage exceeding 300 tons operating on international routes must be equipped with AIS transceivers. These transceivers can transmit real-time location and movement information of ships to shore base stations or satellites, and have collected a large amount of ship trajectory data over the years. Some AIS data service providers collect and aggregate AIS data from AIS transceivers around the world and release it open-source on the Internet, providing data support for research on maritime intelligence.

[0004] Open-source AIS data provides ample resources for mining maritime traffic networks. However, effectively extracting large-area maritime traffic networks from low-quality data is challenging. First, unlike land road networks, ships at sea have considerable freedom of navigation, and their actual trajectories often differ from official routes, thus introducing greater uncertainty into maritime traffic networks. Second, AIS data is near real-time, requiring processing of massive amounts of data when the area of ​​interest is large. Furthermore, open-source AIS data is aggregated from multiple sources, with varying noise levels, densities, and quality; directly performing pattern mining from the raw data may lead to inaccurate results. Therefore, controlling the quality of the raw AIS data is crucial.

[0005] Maritime traffic network mining involves the compression and abstract representation of multi-ship trajectory data, with a focus on deriving traffic route networks for relevant sea areas from complex AIS data in an unsupervised manner. Current research methods can be categorized into three types: grid-based methods, statistical methods, and vector-based methods.

[0006] Grid-based methods subdivide the sea area of ​​interest into a grid space, discretly storing the raw trajectory data in a grid index. Then, the motion characteristics within each grid are summarized and statistically analyzed to represent traffic patterns. This method enables efficient query operations, facilitating map modeling or probabilistic statistical analysis. However, this type of method also has limitations. First, the grid size must be predetermined, making it difficult to represent traffic patterns at different spatial scales. Second, while grid-based methods perform well at small scales, the number of grids increases significantly as the area of ​​interest expands, increasing the computational burden and limiting the analysis of global maritime traffic.

[0007] Statistical methods focus on using quantitative modeling tools to perform statistical analysis on specific water areas. For example, adaptive kernel density estimation (KDE) is used to estimate the distribution of ship positions and speeds, and Ornstein-Uhlenbeck mean regression stochastic processes are used to model ship motion. These methods help determine basic data distributions and thresholds, assist in anomaly detection and ship behavior prediction, and facilitate the estimation of maritime traffic capacity and density. Although these methods are widely used in maritime traffic research, their generalization ability is limited.

[0008] Vector-based methods represent trajectories as interconnected line segments between path nodes, providing a structured and lightweight representation of traffic patterns over large ocean areas. These methods are helpful in constructing global maritime traffic networks. TREAD (Traffic Route Extraction and Anomaly Detection) is a classic framework. This algorithm classifies ship stationary points and entry / exit points and uses density-based incremental noise applied spatial clustering (DBSCAN) to cluster ocean feature regions, thereby extracting navigable trajectories between feature regions and generating route objects. Generally, vector-based methods can reasonably represent maritime traffic networks, but most existing schemes focus on waypoint mining and trajectory classification, simply describing the maritime traffic network as a collection of nodes and connections, lacking a detailed description of the rich semantic information of maritime feature regions and routes.

[0009] Among existing methods for mining maritime traffic networks, some excel at detailed modeling of routes in local sea areas but struggle to extend to large areas. Other methods are adept at extracting vector-described traffic networks from large sea areas but lack detailed descriptions of maritime semantic information. Therefore, there is an urgent need for a feasible method to extract maritime traffic patterns from large areas of low-quality AIS data. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of existing maritime traffic network mining methods, which are difficult to extend to large areas of sea and lack detailed descriptions of maritime semantic information.

[0011] To achieve the above objectives, this invention proposes a method for mining maritime traffic networks based on open-source AIS data, comprising: Step 1: Preprocess the AIS data; Step 2: Identify and classify berthing trajectories using geographic and motion information, and extract ship berthing areas containing semantic distributions of berthing functions based on the berthing trajectories; spatially divide the area of ​​interest by screening trajectory points at ship turning points, and extract ship turning areas using a genetic algorithm; the combination of the ship berthing areas and ship turning areas is used as the feature region; Step 3: Construct a directed graph based on the feature regions, remove noisy data, and generate route boundaries and characterizing routes through kernel density estimation and triangulation methods.

[0012] As an improvement to the above method, step 1 includes: A kinematic interpolation method is used to estimate the missing trajectory points: For two consecutive points in a trajectory and When the time difference between them Greater than the set time interval At the same time, estimate the position and motion information of the lost trajectory points between the two points; express Time; express The time required; the estimation method is as follows: by Establish a local coordinate system with the origin, where The positive direction of the axis is north. The positive direction of the axis is east; calculate and Given the spherical distance and vector direction, we obtain... Coordinates in this coordinate system : in,distance express and The spherical distance between them; course represents and The direction of the vectors between them; Decompose the ship's velocity into two coordinate axes:

[0013] in, Indicates the ship's heading relative to the ground; Indicates the speed of the ship; Indicates decomposition to X The speed of the shaft; Decomposed into Y The speed of the shaft; Estimation using kinematic equations and The missing position and motion information of the trajectory points:

[0014] in, Indicates that the ship is t time X Velocity in the axial direction; Indicates that the ship is t time X The coordinates of the axis; Indicates that the ship is point X Velocity in the axial direction; Indicates that the ship is point X The coordinates of the axis; Indicates that the ship is t time Y Velocity in the axial direction; Indicates that the ship is t time Y The coordinates of the axis; Indicates that the ship is point Y Velocity in the axial direction; Indicates that the ship is point Y The coordinates of the axis; , , and Represents unknown variables; Will and The velocity and position information are substituted into the kinematic equations to solve. , , and The kinematic equations between the two points are obtained; Calculate each interpolation time:

[0015] in, Indicates the first i Interpolation time; By substituting each interpolation time point into the kinematic equations being solved, we obtain estimates of the position and motion information at those times.

[0016] As an improvement to the above method, step 1 further includes: cleaning the ship trajectory data before estimating the missing trajectory points. The deletion interval is less than the first set time threshold. The trajectory; Calculate the time difference between two adjacent points in the trajectory and the speed difference between the direct speed and the actual speed, and compare the results with a first set threshold. Second set threshold Compare the values; when the time difference is greater than... Or the speed difference is greater than The trajectory is divided between these two points; Iterate through all the segmented trajectory segments and set the time difference between adjacent points to be less than 1. And the speed difference between the direct speed and the actual speed is less than The trajectory segments are merged; The interval between deletions is less than the set time threshold. The trajectory.

[0017] As an improvement to the above method, the extraction of the ship berthing area includes: The preprocessed AIS data is used to select berthing points, which are then input into a trained gradient boosting model to output the ship berthing type; the ship berthing type includes berthing and anchoring. Construct a ship berthing area.

[0018] As an improvement to the above method, the process of selecting stopping points includes: Use trajectory points with speeds less than a set speed threshold as stopping points; Exclude docking points where the seawater depth exceeds the set depth threshold; Exclude docking points where the area of ​​the circular region with the nearest land, centered on the docking point, intersects with the nearest land and is less than a set area threshold. If the distance between two stops is less than a set distance threshold or the interval between two stops is less than a second set time threshold, then exclude these two stops. Exclude docking points whose stop time is less than the third set time threshold.

[0019] As an improvement to the above method, the training process of the gradient boosting model includes: The selected docking point data is processed into feature vectors; the feature vectors include: mean distance, standard deviation of distance, 85th percentile of distance, mean speed, standard deviation of speed, 85th percentile of speed, trajectory circularity, proportion of trajectory points within the coastline, average seawater depth, and vessel type. The gradient boosting model is trained using the processed feature vectors and classification labels until it meets the classification accuracy requirements on the test set.

[0020] As an improvement to the above method, the construction of the ship berthing area includes: For each docking point, a convex polygon is generated as the basic docking point position. All intersecting docking points are recursively merged to construct several discrete docking positions. The number of berths, total berthing time, number of vessels that have berthed, and number of vessel types are counted for each berthing berth, and noisy berths with a frequency below a set frequency threshold are removed. A point-order search algorithm is used to cluster the berths to form berthing areas; The Kriging interpolation algorithm is used to fit the geographical distribution of berthing types within the berthing area; the ship berthing area is then constructed.

[0021] As an improvement to the above method, the extraction of the ship turning area includes: The cumulative sum algorithm is used to simplify the ship's departure trajectory, removing trajectory points in the straight-line travel section and retaining only the trajectory points at the turning points; A quadtree is used for spatial partitioning, dividing the region of interest into several subspaces; The main turning regions within each subspace are extracted using a genetic algorithm, and the turning region extraction results at the region boundaries are merged using a point order search clustering method.

[0022] As an improvement to the above method, the construction of a directed graph based on the feature region includes: The region of interest is divided into grids of equal latitude and longitude, and the feature regions are mapped to the grids. For each trajectory point, first obtain the feature regions within its own grid and its 8 neighboring grids, and then perform a second filtering based on the distance between the center of the feature region and the trajectory point to obtain candidate feature regions; The ray method is used to sequentially determine whether the trajectory points are located in the candidate feature regions, thereby obtaining the feature regions to which the trajectory points belong. Construct a directed graph with feature regions as nodes and the paths connecting the feature regions as edges.

[0023] As an improvement to the above method, the generation of route boundaries and the characterization of routes includes: Kernel density is used to estimate the probability density distribution of trajectory points in a directed graph; Each route is meshed using a local equal-area partitioning method, and each trajectory point is assigned to a grid. A Gaussian kernel is used for calculation to determine the route boundary. Using the Delaunay triangulation method, a route triangulation network is constructed; Triangles outside the flight path are filtered out based on the kernel density estimate of the centroid coordinates of the triangles; Based on the number of sides they share, triangles are divided into three categories: single-neighbor triangles, double-neighbor triangles, and triple-neighbor triangles. Single-neighbor triangles are located at the entrances and exits of waterways, triple-neighbor triangles are located at the intersections of waterways, and double-neighbor triangles are located on the waterways themselves. Calculate the number of neighbors of each triangle and the location of the maximum estimated kernel density within it. Find the single-neighbor triangle closest to the starting point as the starting point, and recursively find the path connecting all triangles using a depth-first search. Cut off branches shorter than a set length threshold. The Savitzky-Golonin filtering algorithm is used to smooth the preserved path to obtain the characterizing route.

[0024] This application also provides a maritime traffic network mining system based on open-source AIS data, implemented using the above method. The system includes: The data preprocessing module is used to preprocess AIS data; The feature region acquisition module is used to identify and classify berthing trajectories using geographic and motion information, and extract ship berthing areas containing semantic distributions of berthing functions based on the berthing trajectories; by filtering trajectory points at ship turning points, the region of interest is spatially divided, and a genetic algorithm is used to extract ship turning areas; the set of the ship berthing areas and ship turning areas is used as the feature region; and The module for generating route boundaries and characterizing routes is used to construct directed graphs based on feature regions, remove noisy data, and generate route boundaries and characterizing routes through kernel density estimation and triangulation methods.

[0025] Compared with the prior art, the advantages of the present invention are: 1. A complete unsupervised mining framework is proposed that can accurately extract the transportation network of large sea areas from low-quality AIS data. The framework consists of three parts: data preprocessing, feature region extraction, and maritime transportation network generation. It is easy to apply to multiple fields of maritime intelligent services and has the advantages of high extraction accuracy, low cost of manual supervision, and rich semantic patterns. 2. In view of the difference between anchoring and berthing, this method innovatively uses geographical and motion information to extract the anchoring and berthing trajectories of ships, and constructs a functionally distributed maritime berthing area, realizing in-depth exploration of the maritime berthing pattern; 3. In view of the uneven distribution of data density at sea, a method for mining turning areas at sea based on quadtree dissection and genetic algorithm is proposed. This method can greatly accelerate the mining efficiency of turning areas while ensuring that the mining accuracy is not affected by the data density distribution. Attached Figure Description

[0026] This embodiment utilizes AIS data for the waters surrounding the United States in 2019 (https: / / coast.noaa.gov / htdata / CMSP / AISDataHandler / 2019 / ) published by the National Oceanic and Atmospheric Administration (NOAA). This data covers multiple sea areas including the Atlantic Ocean, Pacific Ocean, Gulf of Mexico, and Caribbean Sea, with longitude ranging from -180° to -61° and latitude ranging from 11° to 73°. The dataset stores navigation records for approximately 31,000 vessels (37 types), constituting a highly complex maritime traffic scenario. The Earth imagery data used in this application originates from Google Earth, a satellite remote sensing data service platform provided by Google. The selected Earth imagery is all high spatial resolution satellite imagery within the coverage area of ​​the AIS data, clearly presenting the geographic details of the area of ​​interest in a visual form. All English labels in the figures are place names.

[0027] Figure 1 The diagram shows the technical framework of a maritime traffic network mining method based on open-source AIS data. Figure 2 The diagram shown is a flowchart of the trajectory cleaning process. Figure 3 The diagram shown is a schematic of trajectory standardization. Figure 4 The image shows a visual example of a manually labeled docking trajectory sample. The sample data is based on open-source AIS data, where: the label indicates whether it is a docking trajectory or an anchored trajectory; MMSI is the Maritime Mobile Service Identifier, which is equivalent to a unique identifier for the vessel; Figure 5(a) shows a schematic diagram of the process for constructing a maritime berthing area; Figure 5(b) shows a visualization of the corresponding offshore berthing area construction process. Figure 6 The image shown is a visualization of the quadtree spatial partitioning results based on the turning points in the open-source AIS data. Figure 7 The diagram shows the evolution process of the genetic algorithm. Figure 8 The diagram shows a schematic of a trajectory point feature region allocation algorithm based on grid partitioning. Figure 9(a) shows the probability density distribution of waypoints in the route estimated using KDE; Figure 9(b) shows the route boundary map determined with a probability threshold of 80%. Figure 9(c) shows a schematic diagram of the route triangle network; Figure 9(d) shows a schematic diagram of the filtering results based on the KDE estimate of the centroid coordinates of the triangle; Figure 9(e) shows a schematic diagram of the path connecting all triangles recursively found using a depth-first search method; Figure 9(f) shows a schematic diagram of the smoothing result of the retained main path using the Savitzky-Golay filtering algorithm; Figure 10 The image shows the results of berthing berth excavation for the Ports of Los Angeles and Long Beach; the polygons in the image represent berthing berths extracted by the algorithm; the same numbers in the image represent the same geographical area. Figure 11 The image shows the results of the berthing area excavation in the ports of Los Angeles and Long Beach; the thin-lined polygons in the image represent the extracted berthing positions, and the thick-lined polygons represent the berthing areas. Figure 12 The image shows a comparison between the berthing areas extracted from the Commonwealth of Puerto Rico and the ports in the World Ports Index; the rectangles and rhombuses in the image represent the extracted port berthing areas and sea berthing areas, respectively; the triangles represent the ports in the World Ports Index; the dots represent the coverage area of ​​the AIS data; the polygons represent the coverage area of ​​the extracted berthing areas; the same numbers in the image represent the same geographical area. Figure 13 The graph shown illustrates the relationship between data size and computation time. Figure 14 The image shows a heatmap of the Florida Straits trajectory and the results of turning area extraction; the dark areas in the image represent high-frequency areas of ship navigation, and the light areas represent low-frequency areas; the polygons in the image represent the extracted turning areas. Figure 15(a) shows a comparison between the typical route mining results from San Diego to the Port of Los Angeles and the planned route of the modern navigation software Searchoutes. The lines with dots in the figure are routes generated by the modern navigation software Searchoutes. The lines without dots are typical routes generated by this method, and the dashed lines are the boundaries of typical routes. Figure 15(b) shows a comparison between the mining results of a typical route from the Port of Winnie-Miles to San Francisco and the planned route of the modern navigation software Searchoutes. The lines with dots in the figure are routes generated by the modern navigation software Searchoutes. The lines without dots are typical routes generated by this method, and the dashed lines are the boundaries of typical routes. Detailed Implementation

[0028] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] This application proposes a method and system for mining maritime traffic networks based on low-quality open-source AIS data covering large areas of sea. It aims to provide maritime agencies with a solution for automatically completing maritime situational awareness and understanding with minimal human supervision. The method comprises three components: data preprocessing, feature region extraction, and maritime traffic network generation. It represents the maritime traffic network as a directed graph composed of feature regions containing rich semantic information and typical routes. This method can accurately extract ship maritime traffic behavior and represent actual maritime traffic patterns, providing valuable knowledge support for various marine intelligent applications.

[0030] The main objectives of this application include: A comprehensive open-source AIS data processing framework is proposed to address the issues of numerical anomalies, spatially anomalous trajectories, and temporal discontinuities in raw data. First, numerical anomalies are removed from the raw AIS data, and a set of raw tracks based on ship identifiers is constructed. Then, a track cleaning algorithm is designed to effectively remove spatially anomalous trajectories, thereby improving the quality of the track data. Finally, a kinematic interpolation method is employed to fill in portions of the tracks with excessively large time intervals, ensuring the spatiotemporal continuity of the tracks.

[0031] A method for extracting berthing areas by combining motion and geographic information is proposed. First, berthing points are identified and berthing tracks are generated. Then, a gradient boosting model is used to classify the berthing tracks into berthing tracks and anchoring tracks. Finally, maritime berthing areas are constructed based on the spatial distribution of berthing tracks, and the functional distribution within the berthing areas is estimated using the Kriging interpolation method.

[0032] A method for extracting turning regions is proposed. First, the CUSUM method (cumulative sum algorithm) is used to identify turning points in the ship's navigation trajectory. Then, a quadtree structure is used to divide the sea area into several sub-regions. Finally, a genetic algorithm is used to capture the regions in each sub-region where turning behavior frequently occurs as turning regions, and a point order search clustering method is used to fuse the turning region extraction results at the region boundaries.

[0033] A method for generating maritime traffic networks is proposed. First, based on the semantic model of ship behavior of "stop-turn-turn-stop", a grid method is used to efficiently construct a set of ship navigation trajectories. Then, noisy routes are removed by weighting the routes. Finally, kernel density estimation and the Delaunay Triangulation algorithm are used to generate route boundaries and characterizing routes.

[0034] In summary, the method provided in this application can effectively remove outliers from the original data while ensuring the spatiotemporal continuity of the trajectory. By extracting maritime stop-and-turn feature regions, it can effectively identify important maritime destinations and high-frequency turning areas, and generate typical routes that accurately reflect the traffic patterns between feature regions. Tested with actual data, this method can accurately mine maritime traffic patterns, exhibiting robustness and scalability. It provides a feasible solution and a complete research framework for maritime traffic network mining, and can be applied to trajectory datasets of other data types or small-scale sea areas.

[0035] The goal of the maritime traffic network mining method proposed in this application is to automatically extract traffic networks from large-scale, noisy raw AIS data. For example... Figure 1 As shown, this method consists of three main parts: a data preprocessing module, a feature region extraction module, and a maritime traffic network generation module. The preprocessing module converts raw AIS data into high-quality ship trajectory data. Then, the feature region extraction module uses a learning method to identify berthing and turning areas. Finally, the maritime traffic network generation module extracts representative routes between different areas.

[0036] 1. AIS Data Processing Data preprocessing workflow as follows Figure 1 As shown at the bottom, it includes three steps: data access, trajectory cleaning, and trajectory standardization.

[0037] During data access, due to the low reliability or potential absence of some attribute items, only nine attributes are retained as shown in Table 1: MMSI (Maritime Mobile Service Identity), BaseDateTime (signal transmission time), LAT (latitude), LON (longitude), SOG (speed over land), COG (heading over land), VesselType (vessel type), Length (vessel length), and Width (vessel width). Next, data items containing obvious anomalies are removed according to rule-based verification, as detailed in the last column of Table 1. Before conducting vessel trajectory research, vessel aggregation is required for the trajectory data. MMSI is used to separate AIS data samples from different vessels, and the trajectory point data for each vessel is sorted in ascending order of time. A maximum time interval threshold is also introduced. The ship's trajectory is divided into different navigation trajectory segments.

[0038] Table 1. AIS Data Attributes and Abnormal Data Judgment Rules

[0039] Besides numerical errors, the raw AIS data also contains spatiotemporal anomalies in its trajectories, which require comprehensive analysis of the entire trajectory to detect. One example is trajectory points appearing in illogical locations. Furthermore, during ship trajectory extraction, the MMSI is considered a unique identifier for the ship; however, due to unforeseen events such as malicious tampering with AIS information, multiple ships may use the same MMSI. Additionally, the raw trajectory dataset contains many trajectories with excessively short sailing times or low temporal density. These factors negatively impact the mining results; therefore, this application proposes an efficient trajectory cleaning method (such as...). Figure 2 As shown in Table 2, this can effectively solve the above problems.

[0040] This trajectory cleaning method mainly consists of four steps. The first step is to utilize... (Time threshold) Removes excessively short trajectories to reduce unnecessary calculations. The second step is trajectory segmentation, which calculates the time difference between adjacent points and the difference between the direct speed and the actual speed, and compares the results with... (Time difference threshold between two adjacent points) and The result is compared to the threshold of the difference between the direct speed and the actual speed. or The trajectory is divided from these two points. The third step is trajectory merging, which iterates through all the resulting sub-trajectory segments and merges those that meet the merging condition (the time difference between two adjacent points is less than 1). And the difference between the direct speed and the actual speed is less than The sub-trajectory segments are merged. The final step utilizes again... Check all sub-trajectory segments to ensure that the output trajectory segments meet the requirements.

[0041] Table 2 Trajectory Cleaning Algorithm

[0042] In addition, the actual time intervals of ship trajectory data are uneven, with some trajectories having long time gaps, which affects the extraction quality of maritime traffic networks. Since ships typically maintain a stable motion during navigation, this application employs a kinematic interpolation method to estimate missing trajectory points. In this application, the maximum time interval of the trajectory is set to... When the time difference between adjacent trajectory points is greater than this value, it is necessary to estimate the motion information of the missing time.

[0043] like Figure 3 As shown, given two consecutive points on a trajectory... and The time difference between them Greater than the preset time interval To obtain the position and motion information of the lost trajectory points between these two points, the following steps are required.

[0044] First of all, with Establish a local coordinate system with the origin at point , where The positive direction of the axis is north. The positive direction of the axis is east. Then, calculate... and The Haversine distance (spherical distance) and vector direction are used to obtain... The coordinates in this coordinate system.

[0045]

[0046] in, distance express and Haversine distance between them; course represents and The direction of the vectors between them.

[0047] Next, the ship's speed is also decomposed into two coordinate axes using the heading:

[0048] Use kinematic interpolation methods to estimate and The position and motion information of unobserved trajectory points are mainly based on the object's acceleration formula, and the velocity and position formulas are obtained through integration. Taking the axial direction as an example, the kinematic equations used are as follows:

[0049] in, Indicates that the ship is t time X Velocity in the axial direction; Indicates that the ship is t time X The coordinates of the axis; Indicates that the ship is point X Velocity in the axial direction; Indicates that the ship is point X The coordinates of the axis.

[0050] Where the acceleration formula Since it is unknown, we set it to linear form:

[0051] The above formula contains two unknown variables. and By and exist Substituting the velocity and position along the axis into formulas (3) and (4), the distance between the two points can be calculated. The equations of motion along the axial direction. Then, following the same method, we can obtain... The equations of motion along the axis. To determine the interpolation time, it is also necessary to estimate the number of interpolation points. :

[0052] Each interpolation time point can be obtained in the following way:

[0053]

[0054] Finally, by substituting the interpolated times into the previously solved equations of motion, we obtain estimates of the position and motion information at these times.

[0055] 2. Extract feature regions The process of mining maritime traffic networks can be viewed as constructing a directed graph, where nodes represent intersections or destinations for ships at sea, and edges represent possible routes between these nodes. Feature region extraction focuses on analyzing the spatiotemporal information in AIS data and studying the geospatial distribution of these behavioral locations to extract feature regions as nodes of the maritime traffic network. Since stopping and turning behaviors have different characteristics and descriptions, different mining methods are required for each behavior. Figure 1 The middle section shows the core steps of feature region extraction.

[0056] 2.1 Extracting the docking area Identifying berthing trajectory points is fundamental to berthing behavior recognition. In the marine environment, ships are not completely stationary when berthed due to the effects of sea winds and waves; therefore, a speed threshold is introduced. This method filters track points to determine if they are in a docked state. Additionally, it considers further filtering based on geographic factors. Specifically, a circular area with a fixed radius is used to describe the docking location. Then, the area where this location intersects with land and the seawater depth are combined to determine if it meets pre-set threshold requirements. Since ports and anchorages are typically close to land, this method effectively eliminates noisy data or unexpected low-speed navigation in the open ocean. Detailed judgment methods are given in Table 3.

[0057] Table 3. Stop Point Recognition Algorithm

[0058] Next, by calculating the distance and time difference between the points, a sample of ship berthing trajectories is constructed. To reduce noise interference, excessively short trajectories need to be removed to optimize the sample quality. The specific algorithm process is shown in Table 4.

[0059] Table 4. Dock Trajectory Generation Algorithm

[0060] The berthing and anchoring behaviors of ships exhibit significant spatial differences. Berthing trajectories are more convergent, while anchoring trajectories are more divergent, typically exhibiting a semi-circular shape. Since ship berthing trajectories are time-series sequences of indeterminate length, to complete the classification task, we first manually labeled some samples using trajectory visualization (e.g., ...). Figure 4 (as shown), and then the original trajectory is processed into a 10-dimensional feature vector, including the mean, standard deviation and 85th percentile of the distance, the mean, standard deviation and 85th percentile of the velocity, the circularity of the trajectory, the proportion of trajectory points within the coastline, the average seawater depth and the ship type.

[0061] This application uses Gradient Boosting as the classification model. The Gradient Boosting model is trained using 10-dimensional feature vectors and classification labels until it meets the classification accuracy requirements on the test set. Then, the trained classification model is used to label the parking category for each parking trajectory sample.

[0062] Next, this application proposes a method for constructing berthing areas based on berthing trajectories to achieve accurate extraction and semantic matching of berthing areas at sea. The algorithm flowchart is shown in Figure 5(a), and its corresponding graphical representation is shown in Figure 5(b). First, for each berthing trajectory, an outer convex polygon is generated as the basic berthing point, and all intersecting berthing points are recursively merged to construct several discrete berthing positions. Subsequently, the number of berthing trips, total berthing duration, number of berthed vessels, and number of vessel types for each berthing position are counted, and noise berths with too low frequency are removed. Since a berthing area is generally composed of multiple berthing positions of varying sizes, the number and density of berths it contains are also different. Therefore, this application uses the OPTICS algorithm (point order search algorithm) to cluster the berthing positions. Finally, the Kriging Interpolation algorithm is used to fit the geographical distribution of berthing types within the berthing area. A standard Kriging interpolation model was adopted, and Spherical was selected as the semi-variogram function to measure the spatial autocorrelation between data. Triples (Lon, Lat, Label) composed of the latitude and longitude of the location of the center of gravity of the docked ship and the classification results of the trajectory were used to characterize the samples of a single docking event, and the parameters of the interpolation model were estimated using these sample points.

[0063] 2.2 Extracting the turning area Important maritime traffic intersections reflect key patterns in local maritime traffic. Extracting turning regions aims to identify areas where ships frequently turn from AIS data and incorporate these areas as nodes into the maritime traffic network. To this end, this application proposes a three-step overall scheme. First, the CUSUM method is used to simplify the ship's departure trajectory, removing trajectory points from straight-line travel and retaining only those at turning points. Then, a quadtree is used for spatial partitioning, dividing the region of interest into several subspaces. Finally, a genetic algorithm is used to extract the main turning regions within each subspace, and a point-order search clustering method is used to merge the extracted turning region results at the region boundaries.

[0064] In a ship's navigation trajectory, there are numerous straight sections. The CUSUM algorithm is used to detect significant changes in the ship's heading during navigation, preserving turning points and thus compressing the trajectory. Let... For a single flight path, here is a sequence of changes in heading:

[0065] The decision function of the bilateral CUSUM algorithm is defined as follows:

[0066]

[0067] In the above formula Indicates the first Positive CUSUM statistics for each track point; Indicates the first Reverse CUSUM statistics for each track point; This represents the average value of the change in trajectory heading. The noise term can be estimated by calculating the standard deviation of the trajectory heading change and multiplying it by a fixed coefficient. To determine the significance of the heading change, a decision threshold is set. .when or Greater than the set threshold At that time, the algorithm considers the first... The headings of the points changed significantly, so they were retained.

[0068] A significant challenge in mining maritime traffic data is its uneven spatial distribution, which is why most clustering algorithms struggle to produce good clustering results. Similarly, genetic algorithms for mining turning regions are also affected by this uneven spatial distribution. High-density areas represent high-quality data, leading to clustering results in high-density regions while low-density areas become less competitive. Furthermore, the time complexity of genetic algorithms is primarily influenced by chromosome length and the number of candidate points. As the computational area increases, the computation time of genetic algorithms increases exponentially, making them difficult to apply directly in large-area marine scenarios. In a quadtree structure, each node can grow four child nodes, recursively subdividing a high-density region into four smaller regions. The final partitioning results are stored in the leaf nodes of the quadtree. To ensure the algorithm's adaptability, a maximum depth threshold also needs to be set. When the node depth of a partition reaches this threshold, further partitioning stops to prevent the final partition area from becoming too small. A schematic diagram of the spatial partitioning result is shown below. Figure 6 As shown.

[0069] Genetic algorithms require encoding the problem into a "population-chromosome-gene" format. In this method, each gene represents a candidate turning region, containing three parameters: latitude, longitude, and radius, formally expressed using a circle equation. Each chromosome consists of... Each population consists of several chromosomes. This method randomly initializes a population in each leaf partition of a quadtree. After multiple generations of evolution, the chromosome with the highest fitness in that population is the extraction result for the turning region of that partition. Figure 7 The evolution of genetic algorithms is demonstrated.

[0070] Fitness functions are used to evaluate the merits of each chromosome in a population and serve as the basis for genetic operations. In this application, the calculation method for chromosome fitness scores is defined by formula (11). Each gene represents a circular region in geographic space, and the number of trajectory points contained within this region is counted and used as the base score. Simultaneously, an adjustment coefficient is introduced. This is to ensure that the radius of the turning region does not increase indefinitely. The final score of each gene is obtained by multiplying the base score by an adjustment factor. The final fitness of the chromosome is the sum of the scores of all genes, divided by the total number of trajectory points within the partition. .

[0071]

[0072] in, This indicates the minimum radius of the turning area.

[0073] To prevent multiple genes within a chromosome from stacking in high-density regions, a penalty function is introduced: if there is spatial overlap between genes on a chromosome, the fitness score of that chromosome is 0.

[0074] Through genetic evolution, a population can produce chromosomes with superior fitness. This application employs a roulette wheel method to select parents, where the probability of each chromosome being selected is proportional to its fitness score. New chromosomes are generated using a single-point crossover method. A gene locus is randomly selected, and the parents exchange genes following that location. Each crossover operation generates two new chromosomes. In the new population, each chromosome has a certain probability of mutation, i.e., a randomly selected gene locus is used to replace the original gene with a new, random gene. During the survival-of-the-fittest phase, chromosomes with lower fitness scores are removed to restore the population size. Once the set number of generations has been reached, the chromosome with the highest fitness score in the population is the optimal solution for extracting the turning region. For the sub-partitions obtained from spatial partitioning, genetic algorithms can be used in parallel to extract the turning region.

[0075] 3. Generate a maritime transportation network The generation of the maritime traffic network involves three steps. First, the historical ship tracks are reconstructed into a (from_areaX, to_areaY) form using the generated feature regions (the berthing and turning areas obtained in step 2). Since numerous noisy connections and some disconnected discrete regions are generated between nodes, denoising processing is necessary. Finally, kernel density estimation and the Delaunay Triangulation algorithm are used to generate a set of navigation tracks between nodes to represent the routes.

[0076] Based on a standard track database, this method determines whether the location of each track point lies within a feature region and tracks the feature region transformation sequence of the entire track, ultimately obtaining a set of tracks between each pair of feature regions. While this work itself is not particularly challenging, the process of assigning feature regions to each track point is very time-consuming, considering the sheer volume of AIS data. This application employs a gridded method to accelerate the feature region selection process. The algorithm flow is as follows: Figure 8 As shown. First, the region of interest is divided into grids of equal latitude and longitude, and the feature regions are stored in the corresponding grids. This allows for the calculation of features based on given latitude and longitude coordinates. The algorithm calculates the grid corresponding to the target location and its neighboring grids in a time-complexity manner. For each trajectory point, it first obtains the feature regions within its own grid and its 8 neighboring grids, then performs a secondary filtering based on the distance between the center of the feature region and the trajectory point to obtain candidate feature regions. Finally, it uses the ray casting method to sequentially determine whether the trajectory point is located within the candidate feature regions. In actual testing, the optimized algorithm can save approximately 80% of the computation time.

[0077] After the above steps, a basic maritime traffic network structure can be obtained (the resulting network structure is a directed graph with feature regions as nodes and the tracks connecting feature regions as edges). Due to factors such as data interruptions in open-source AIS data, a large number of noisy connections will be generated. Therefore, the above results need to be post-processed. First, the number of tracks for each edge is counted as a weight value. If a node is connected by multiple edges, and one of the edges has a very low weight value, then that edge is considered noisy data. Because a large amount of AIS data is missing in offshore areas, mining large areas of sea will result in multiple disconnected graphs. For some smaller graphs, they can be regarded as noisy regions and removed from the network.

[0078] Using the filtered routing data, typical flight paths between characteristic regions can be generated. Figures 9(a)-9(f) The key steps in this process are illustrated. First, the probability density distribution of waypoints in the route is estimated using KDE (as shown in Figure 9(a)). KDE is a non-parametric estimation method suitable for two-dimensional traffic density distribution modeling. Each route is then meshed using a local equal-area partitioning method, and each waypoint is assigned to a grid. Gaussian kernels are used for calculation, and a probability threshold of 80% is set to determine the route boundaries (as shown in Figure 9(b)).

[0079] Subsequently, the Delaunay Triangulation method was used to construct a route triangulation network (as shown in Figure 9(c)). Triangles are distributed both inside and outside the route. Triangles outside the route are meaningless for extracting typical routes, so they are filtered based on the KDE estimates of the centroid coordinates of the triangles, as shown in Figure 9(d). Triangles are categorized into three types based on the number of shared edges: single-neighbor triangles, two-neighbor triangles, and three-neighbor triangles, used to determine key points on typical routes. Single-neighbor triangles are located at the entrances and exits of the route, three-neighbor triangles are located at the intersections of the route, and two-neighbor triangles are located on the route itself. The number of neighbors of each triangle and the location of its maximum KDE value are calculated. The single-neighbor triangle closest to the starting point is found as the starting point, and a depth-first search is used to recursively find the path connecting all triangles (as shown in Figure 9(e)), pruning excessively short branches. Finally, the Savitzky-Golay filtering algorithm is used to smooth the retained main path to obtain the characterization route, as shown in Figure 9(f). It can be seen that the characterization route can well express the main direction of the route.

[0080] The following uses real-world data to validate this method. This example uses AIS data for the waters surrounding the United States published by the National Oceanic and Atmospheric Administration (NOAA) in 2019. This data covers multiple sea areas including the Atlantic, Pacific, Gulf of Mexico, and Caribbean Sea, with longitude ranging from -180° to -61° and latitude ranging from 11° to 73°. The dataset stores navigation records for approximately 31,000 vessels (37 types), forming a highly complex maritime traffic scenario.

[0081] In data preprocessing, numerical and spatial outliers were removed, excessively short trajectories were filtered out, and the trajectories were standardized to a time-uniform form. Ultimately, 212,316 standardized trajectories from 22,302 ships, totaling 1,312,161,124 trajectory points, were retained and stored in a PostgreSQL database to provide data query support for subsequent data mining. Approximately 30.3% of the trajectory points were removed during data access, with an average processing speed of approximately [missing information - likely a unit of time]. Points. During the trajectory cleaning process, approximately 35.7% of the points were deleted, with an average processing speed of approximately [number] points per minute. Points. Approximately 12.1% of trajectory points were added during the trajectory standardization process, with an average processing speed of approximately [number] points per minute. point.

[0082] To verify the effectiveness of the docking trajectory classification algorithm, 2000 docking trajectories were randomly selected from all ship trajectory data and manually labeled as classification samples. Figure 4 Examples of labels are shown. The experimental data covers six types of vessels from the AIS dataset, including 1744 berthed samples and 256 anchored samples. The experimental data is divided into training and test datasets in a 7:3 ratio; these two datasets are mutually exclusive. The classification results are compared with those of three commonly used pattern classification methods: Support Vector Machine (SVM), Random Forest, and Multilayer Perceptron (MLP). Classification accuracy and Kappa coefficient are selected as evaluation metrics.

[0083] Table 5 shows the classification results of the four classification methods on the test set. It can be seen that the classification accuracy and kappa coefficient of the gradient boosting-based classification model are the highest among the four methods. The results indicate that the classification results of this application best match the actual berthing behavior of ships, providing an accurate data foundation for the functional semantic analysis of ship berthing areas.

[0084] Table 5. Dock Trajectory Classification Results

[0085] To verify the accuracy of this method, the results of the constructed berthing positions and berthing areas are visually demonstrated using the Port of Los Angeles (LA) and the Port of Long Beach (LB) in California, USA. Figure 10 The document showcases 12 main wharves and their constructed berths at the two ports. An overlay demonstration of the berth locations with images of the wharves' berths demonstrates a high degree of consistency between the berth locations generated using the method described in this application and the berth images. Figure 11 The berthing areas shown correctly cover the land terminals and nearshore areas of the Ports of Los Angeles and Long Beach. The generated berths are located within the ports, while the anchorages are located in shallow water areas near the ports. Furthermore, the functional semantic estimation results are consistent with the port images, further validating the high accuracy of the proposed berthing area generation method.

[0086] When the proportion of berth samples in a berthing area exceeds 25%, the berthing area is classified as a port berthing area. Subsequently, the port berthing areas are compared with Google Earth images; if no port image is found within the berthing area, the result is considered invalid. After manual verification, 782 of the 912 berthing areas were identified as port berthing areas, of which 764 were considered correct (accuracy rate 97.7%). Furthermore, to verify the completeness of the method, the port berthing areas are compared with ports in the World Port Index (WPI). Figure 12As shown, the analysis uses data from the Free State of Puerto Rico as an example. The WPI includes 12 ports in the region, and the method in this application extracted 15 port berthing areas and 5 maritime berthing areas. Nine ports in the WPI are consistent with the results of this application's method, while Aguadilla, Arroyo, and Playa De Humacao were not extracted. Furthermore, it should be noted that there are discrepancies between the port locations in the World Association of Ports and Harbors (WAPH) and the extraction results of this application's method. In the case of Las Marias, this discrepancy arises because the port locations in the WPI are marked on the mainland. In contrast, the areas extracted by the method in this application are automatically mined and distributed in shallow water areas. In addition, ports not marked in the WPI, such as berthing areas 0, 2, 3, 5, 8, 11, and 16, were also extracted; therefore, the mining results provided by the method in this application are more comprehensive.

[0087] The quadtree-based genetic algorithm enables efficient data processing. To evaluate its time performance, experiments were conducted using four datasets of different sizes: 371, 1961, 9353, and 14370 points. For standard GA, the chromosome length was set to 10, the population size to 100, the mutation rate to 10%, and the computational upper bound to 100 epochs. In our method, spatial dissection of the region yields 10 turning regions. The parameters of the genetic algorithm within each subregion, including population size, mutation rate, and computational upper bound, are consistent with those of the standard genetic algorithm. Figure 13 The runtime variations of these two methods on four datasets are illustrated. As the dataset grows, the computation time of standard GA increases rapidly and linearly, making it unsuitable for processing large datasets. In contrast, the method proposed in this application effectively manages the computation time of each subtask by controlling the number of candidate points and chromosome length within each subregion. This strategy ensures that the overall computation time gradually increases with data size, thus achieving efficient processing of large-scale datasets.

[0088] Another important evaluation involves the accuracy of the maritime turning areas extracted by the method in reflecting maritime traffic patterns. Given the lack of standardized ground facts or evaluation criteria for these turning areas, a visual evaluation method was chosen. The selected test area is the Florida Straits between the Florida peninsula in the United States and Cuba. This region is a vital gateway from the northern Gulf of Mexico to the Atlantic Ocean. The key evaluation metric is the degree of match between the turning areas generated by the method and actual traffic flow. A higher degree of match indicates higher extraction quality. Figure 14 As shown, the heatmap of the trajectory in this area is overlaid with the results of the extracted turning area. The turning area shows a significant consistency with the main route pattern in this area, verifying the accuracy of the method proposed in this application.

[0089] The Gulf of California was chosen as the test area. This narrow Pacific coastal bay stretches from north of San Diego to south of San Francisco, and is characterized by heavy maritime traffic. Two routes were generated in this area using the modern navigation software Searchoutes: one from San Diego to Los Angeles and the other from Hunem Port to San Francisco. These routes were compared with turning areas and typical routes extracted using the method described in this application. The extraction results in Figures 15(a) and 15(b) are consistent with the routes in Searchoutes, indicating that the maritime traffic network obtained here accurately reflects actual maritime routes.

[0090] This application also provides a maritime traffic network mining system based on open-source AIS data, implemented using any of the above methods, the system comprising: The data preprocessing module is used to preprocess AIS data; The feature region acquisition module is used to extract the ship berthing area using the area where the ship's berthing location intersects with the land and seawater depth information; by filtering the trajectory points at the ship's turning points, the area of ​​interest is spatially divided, and a genetic algorithm is used to extract the ship's turning area; the combination of the ship berthing area and the ship turning area is used as the feature region; and The module for generating route boundaries and characterizing routes is used to construct directed graphs based on feature regions, remove noisy data, and generate route boundaries and characterizing routes through kernel density estimation and triangulation methods.

[0091] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for mining maritime traffic networks based on open-source AIS data, comprising: Step 1: Preprocess the AIS data; Step 2: Identify and classify berthing trajectories using geographic and motion information, and extract ship berthing areas containing semantic distribution of berthing functions based on the berthing trajectories; By selecting trajectory points at the ship's turning points, the area of ​​interest is spatially divided, and a genetic algorithm is used to extract the ship's turning area; the combination of the ship's berthing area and the ship's turning area is used as the feature area. Step 3: Construct a directed graph based on the feature regions, remove noisy data, and generate route boundaries and characterizing routes using kernel density estimation and triangulation methods; The extraction of ship berthing areas includes: The preprocessed AIS data is used to select berthing points, which are then input into a trained gradient boosting model to output the ship berthing type; the ship berthing type includes berthing and anchoring. Construct ship berthing areas; The constructed ship berthing area includes: For each docking point, a convex polygon is generated as the basic docking point position. All intersecting docking points are recursively merged to construct several discrete docking positions. The number of berths, total berthing time, number of vessels that have berthed, and number of vessel types are counted for each berthing berth, and noisy berths with a frequency below a set frequency threshold are removed. A point-order search algorithm is used to cluster the berths to form berthing areas; The Kriging interpolation algorithm is used to fit the geographical distribution of berthing types within the berthing area; the ship berthing area is then constructed.

2. The method for mining maritime traffic networks based on open-source AIS data according to claim 1, characterized in that, Step 1 includes: A kinematic interpolation method is used to estimate the missing trajectory points: For two consecutive points in a trajectory and When the time difference between them Greater than the set time interval At the same time, estimate the position and motion information of the lost trajectory points between the two points; express Time; express The time required; the estimation method is as follows: by Establish a local coordinate system with the origin, where The positive direction of the axis is north. The positive direction of the axis is east; calculate and Given the spherical distance and vector direction, we obtain... Coordinates in this coordinate system : ； in, distance express and The spherical distance between them; course represents and The direction of the vectors between them; Decompose the ship's velocity into two coordinate axes: ； in, Indicates the ship's heading relative to the ground; Indicates the speed of the ship; Indicates decomposition to X The speed of the shaft; Indicates decomposition to Y The speed of the shaft; Estimation using kinematic equations and The missing position and motion information of the trajectory points: ； in, Indicates that the ship is t time X Velocity in the axial direction; Indicates that the ship is t time X The coordinates of the axis; Indicates that the ship is point X Velocity in the axial direction; Indicates that the ship is point X The coordinates of the axis; Indicates that the ship is t time Y Velocity in the axial direction; Indicates that the ship is t time Y The coordinates of the axis; Indicates that the ship is point Y Velocity in the axial direction; Indicates that the ship is point Y The coordinates of the axis; , , and Represents unknown variables; Ship and Substitute the velocity and position information of the point into the kinematic equations and solve. , , and The kinematic equations between the two points are obtained; Calculate each interpolation time: ； in, Indicates the first i Interpolation time; By substituting each interpolation time point into the kinematic equations being solved, we obtain estimates of the position and motion information at those times.

3. The method for mining maritime traffic networks based on open-source AIS data according to claim 2, characterized in that, Step 1 further includes cleaning the ship trajectory data before estimating the missing trajectory points: The deletion interval is less than the first set time threshold. The trajectory; Calculate the time difference between two adjacent points in the trajectory and the speed difference between the direct speed and the actual speed, and compare the results with a first set threshold. Second set threshold Compare the values; when the time difference is greater than... Or the speed difference is greater than The trajectory is divided between these two points; Iterate through all the segmented trajectory segments and set the time difference between adjacent points to be less than 1. And the speed difference between the direct speed and the actual speed is less than The trajectory segments are merged; The interval between deletions is less than the set time threshold. The trajectory.

4. The method for mining maritime traffic networks based on open-source AIS data according to claim 1, characterized in that, The process of selecting stops includes: Use trajectory points with speeds less than a set speed threshold as stopping points; Exclude docking points where the seawater depth exceeds the set depth threshold; Exclude docking points where the area of ​​the circular region with the nearest land, centered on the docking point, intersects with the nearest land and is less than a set area threshold. If the distance between two stops is less than a set distance threshold or the interval between two stops is less than a second set time threshold, then exclude these two stops. Exclude docking points whose stop time is less than the third set time threshold.

5. The method for mining maritime traffic networks based on open-source AIS data according to claim 1, characterized in that, The training process of the gradient boosting model includes: The selected docking point data is processed into feature vectors; the feature vectors include: mean distance, standard deviation of distance, 85th percentile of distance, mean speed, standard deviation of speed, 85th percentile of speed, trajectory circularity, proportion of trajectory points within the coastline, average seawater depth, and vessel type. The gradient boosting model is trained using the processed feature vectors and classification labels until it meets the classification accuracy requirements on the test set.

6. The method for mining maritime traffic networks based on open-source AIS data according to claim 1, characterized in that, The extraction of the ship's turning area includes: The cumulative sum algorithm is used to simplify the ship's departure trajectory, removing trajectory points in the straight-line travel section and retaining only the trajectory points at the turning points; A quadtree is used for spatial partitioning, dividing the region of interest into several subspaces; The main turning regions within each subspace are extracted using a genetic algorithm, and the turning region extraction results at the region boundaries are merged using a point order search clustering method.

7. The method for mining maritime traffic networks based on open-source AIS data according to claim 1, characterized in that, The construction of the directed graph based on the feature regions includes: The region of interest is divided into grids of equal latitude and longitude, and the feature regions are mapped to the grids. For each trajectory point, first obtain the feature regions within its own grid and its 8 neighboring grids, and then perform a second filtering based on the distance between the center of the feature region and the trajectory point to obtain candidate feature regions; The ray method is used to sequentially determine whether the trajectory points are located in the candidate feature regions, thereby obtaining the feature regions to which the trajectory points belong. Construct a directed graph with feature regions as nodes and the paths connecting the feature regions as edges.

8. The method for mining maritime traffic networks based on open-source AIS data according to claim 1, characterized in that, The generation of route boundaries and the characterization of routes include: Kernel density is used to estimate the probability density distribution of trajectory points in a directed graph; Each route is meshed using a local equal-area partitioning method, and each trajectory point is assigned to a grid. A Gaussian kernel is used for calculation to determine the route boundary. Using the Delaunay triangulation method, a route triangulation network is constructed; Triangles outside the flight path are filtered out based on the kernel density estimate of the centroid coordinates of the triangles; Based on the number of sides they share, triangles are divided into three categories: single-neighbor triangles, double-neighbor triangles, and triple-neighbor triangles. Single-neighbor triangles are located at the entrances and exits of waterways, triple-neighbor triangles are located at the intersections of waterways, and double-neighbor triangles are located on the waterways themselves. Calculate the number of neighbors of each triangle and the location of the maximum estimated kernel density within it. Find the single-neighbor triangle closest to the starting point as the starting point, and recursively find the path connecting all triangles using a depth-first search. Cut off branches shorter than a set length threshold. The Savitzky-Golonin filtering algorithm is used to smooth the preserved path to obtain the characterizing route.

9. A maritime traffic network mining system based on open-source AIS data, implemented according to any one of the methods described in claims 1-8, characterized in that, The system includes: The data preprocessing module is used to preprocess AIS data; The feature region acquisition module is used to identify and classify berthing trajectories using geographic and motion information, and extract ship berthing areas containing semantic distributions of berthing functions based on the berthing trajectories; by filtering trajectory points at ship turning points, the region of interest is spatially divided, and a genetic algorithm is used to extract ship turning areas; the set of the ship berthing areas and ship turning areas is used as the feature region; and The module for generating route boundaries and characterizing routes is used to construct directed graphs based on feature regions, remove noisy data, and generate route boundaries and characterizing routes through kernel density estimation and triangulation methods.

Images