Fishing boat trajectory classification method based on gradient lifting model

Abstract

The invention relates to a fishing boat trajectory classification method based on a gradient lifting model, and belongs to the field of fishing boat trajectory classification. The method comprises the following steps: acquiring fishing boat trajectory data (including a fishing boat number, a position, a speed, a direction and a timestamp) from an automatic ship identification system, interpolating missing values, and smoothing or eliminating abnormal values; performing feature extraction including spatial features, time features, course and direction features, dynamic change features and statistical features; respectively training three gradient lifting models by using training data, enabling the models to pay attention to minority class samples through loss function weighting, optimizing hyper-parameters in combination with grid search, and evaluating the performance on a verification set by using the accuracy rate, the precision rate, the recall rate and the F1 value; and selecting the model with the optimal performance to classify the fishing boat operation types of the test set. The method can effectively solve the problem of unbalanced fishing boat operation types, has good generalization performance and real-time processing advantages, and can provide powerful technical guarantee for fishery supervision.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of fishing boat trajectory classification, and particularly relates to a fishing boat trajectory classification method based on a gradient boosting model. BACKGROUND

[0002] With the continuous improvement of global fine regulation demand of fishery resources, relying on the fishing boat trajectory data collected by the automatic identification system (AIS) to carry out fishing boat behavior identification has become one of the core tasks of modern marine fishery management. For example, by automatically extracting the fishing boat operation mode through an algorithm model, the typical fishery operation modes such as purse seine, trawl and gillnet can be accurately identified, and the output results can provide scientific and efficient decision support for the on-site patrol scheduling of marine law enforcement departments, intelligent early warning and tracing of illegal fishing behavior, and evaluation of fishery resource reserves and sustainable fishing planning.

[0003] The automatic identification system of the ship can continuously collect the key dynamic data such as the position (latitude and longitude), heading, speed, etc. of the ship through the shore-based base station and satellite receiving network, and has the core advantages of wide global coverage and high spatio-temporal resolution. However, the original AIS trajectory data often has problems such as uneven sampling interval (such as high-frequency sampling during navigation and low-frequency sampling during parking), data packet loss (affected by communication signal shielding or interference), repeated recording (multiple base station signal overlapping reception) and abnormal value (such as positioning jump and speed anomaly), which need to be improved through data cleaning (de-duplication, format standardization), interpolation completion (such as linear interpolation to fill in missing positions), noise removal (such as smoothing filter to correct abnormal values) and other preprocessing methods to improve the data quality. In addition, due to the significant differences in AIS signal coverage intensity in different sea areas (such as nearshore and open sea, busy waterway and remote sea area) and navigation characteristics and data reporting rules of different ship types (such as cargo ship, fishing boat and passenger ship), the generalization ability of the model is limited when applied across sea areas and ship types, so it is necessary to design robust feature engineering (such as spatio-temporal adaptive features) and algorithm framework to adapt to complex application scenarios.

[0004] Existing methods for identifying fishing vessel operation types mostly employ traditional machine learning models (such as random forests and support vector machines) or shallow neural networks. These methods have limitations such as strong reliance on manual features and poor performance in handling imbalanced operation categories. Ultimately, this results in limited model classification accuracy and robustness, making it difficult to implement in actual fisheries supervision systems. In recent years, although some deep learning methods (such as recurrent neural networks, convolutional neural networks, and Transformers) have emerged, which can uniformly model local dynamic features (such as short-term speed fluctuations) and global temporal dependencies (such as the periodicity of multi-day operations) through end-to-end learning, significantly improving the feature extraction capabilities of traditional models, the following bottlenecks still exist: First, they are highly dependent on high-quality labeled data. In fisheries scenarios, the labeling cost of "trajectory + on-site verification" is high, and samples of niche and illegal operations are scarce, further exacerbating the class imbalance. Second, they have high computational overhead and poor real-time performance. Model training and inference require high-performance computing power, making it difficult to adapt to lightweight hardware such as island shore-based stations and shipborne law enforcement terminals. Third, they have weak scenario adaptability and are not robust enough to the "sampling interval fluctuations" of AIS trajectories (such as high frequency during navigation and low frequency during anchoring) and cross-sea area data differences (dense trajectories in nearshore areas vs. sparse trajectories in the open ocean). Fourth, they have poor interpretability. The "black box nature" of deep learning makes it impossible to clearly trace "why the model determines a certain fishing vessel to be a certain type of operation," making it difficult to meet the traceability requirements of law enforcement and supervision for decision-making basis. Summary of the Invention

[0005] The purpose of this invention is to provide a method for classifying fishing vessel trajectories based on a gradient boosting model. Through systematic data preprocessing, multi-dimensional feature engineering, integrated feature selection, and comparative testing and optimal model selection of multiple gradient boosting models, this method achieves high-precision identification of mainstream fishing vessel operation types such as gillnetting, trawling, and purse seine, providing reliable technical support for dynamic fisheries supervision and illegal fishing detection.

[0006] To achieve the above objectives, the technical solution of the present invention is: a method for classifying fishing vessel trajectories based on a gradient boosting model, comprising the following steps:

[0007] Step S1: Obtain fishing vessel trajectory data from the Automatic Identification System (AIS) and perform data preprocessing operations, including: interpolating missing values, smoothing or removing outliers, and standardizing the time format.

[0008] Step S2: Extract 80 dimensions of features, including spatial features, temporal features, heading and direction features, dynamic change features, and statistical features;

[0009] Step S3: Train three gradient boosting models, CatBoost, XGBoost and LightGBM, respectively, using the training data. Weight the loss function to make the model focus on minority class samples, optimize the hyperparameters by combining grid search, and evaluate the performance on the validation set using accuracy, precision, recall and F1 score.

[0010] Step S4: Select the model with the best performance to classify the fishing vessel operation types on the test set and output the operation type labels.

[0011] Furthermore, in step S1, the fishing vessel trajectory data includes the fishing vessel number, location, speed, direction, and timestamp.

[0012] Furthermore, in step S1, the data preprocessing operations are as follows:

[0013] Step S11: The acquired raw fishing vessel trajectory data has missing fields. For trajectory segments with a small missing ratio, linear interpolation is used to fill in the missing points to keep the trajectory time continuous. If the missing data is severe or there is no interpolation basis, the corresponding trajectory data is directly removed. Due to sampling errors or trajectory drift, some points may exhibit anomalies such as sudden changes in speed and changes in direction angle. These are identified by setting thresholds for the rate of change of speed and direction, and then smoothed or removed to eliminate abnormal disturbances in the trajectory.

[0014] Step S12: Standardize the format of the time field in the original fishing boat trajectory data, processing it one by one according to mon dayhh:mm:ss, where mon represents the month; day represents the day; hh represents the hour of the day; mm represents the minute; and ss represents the second.

[0015] Step S13: In other stages of data cleaning, some features may have a small number of missing values ​​due to boundary conditions or data anomalies. To ensure data integrity and model stability, zero values ​​are used to fill in the missing values ​​to prevent abnormal interruption of the training process due to missing data.

[0016] Furthermore, in step S11, the linear interpolation formula is:

[0017]

[0018] in, It represents a value at a specific moment, specifically the position coordinates, velocity, or direction; This indicates the moment when a missing value exists. and Distance The two most recent consecutive moments with sampled values, .

[0019] Further, step S2 is as follows:

[0020] Step S21: Extract spatial features:

[0021] First, calculate the actual heading angle real_angle based on the coordinates:

[0022]

[0023] Where Δy and Δx represent the difference in the projected plane coordinates, with a minimum value of 0.00001 added to avoid division by zero in the calculation;

[0024] Further calculate the angular velocity real_angle_var:

[0025]

[0026] in, Indicates the time difference between two points. This represents the actual heading angle at time t;

[0027] Next, the spherical distance between adjacent time points (haversine) is calculated, and the heading angle (bearing) is estimated using the spatial relationship between coordinate points.

[0028]

[0029]

[0030] Where R is the Earth's radius. and The latitude of the two points is represented by Δlat, which represents the difference in latitude, and Δlon, which represents the difference in longitude. Since trigonometric functions require input in radians, the latitude and longitude information needs to be converted from degrees to radians for calculation.

[0031] Step S22: Extract time features: Extract the hours and minutes from the time field separately, and use sine and cosine functions to encode them for time periodicity to capture the diurnal rhythm of fishing boat operations;

[0032] Step S23: Extract heading and direction features: Perform direction periodic encoding on the heading angle to alleviate the discontinuity of the angle value around 0° / 360°;

[0033] Step S24: Extract dynamic change features: Perform first-order difference on the position latitude and longitude (x,y), velocity, and direction in the trajectory data to calculate the position increments x_var, y_var, distance increments distance_var, velocity increments velocity_var, and direction change angle_var to characterize the trend of the fishing boat's motion state.

[0034] Step S25: Extract statistical features: Perform statistical analysis on each feature according to the trajectory segment, and extract the mean, standard deviation, range, root mean square (rms), and waveform factor boxing to capture the overall distribution pattern of the local trajectory.

[0035] Step S26: To ensure the uniformity of the input sequence, slice each trajectory with a fixed step size SEQ_LEN, and retain segments of appropriate length as model input samples to achieve an effective conversion from the original trajectory to the model input.

[0036] Furthermore, in step S22, the time periodicity encoding is as follows:

[0037]

[0038] in, Indicates hours, Representing minutes, the cyclic time variable is converted into continuous features that are easy for the model to understand through sine and cosine encoding. and , and This represents sine and cosine encoding.

[0039] Furthermore, in step S23, the direction periodic encoding is as follows:

[0040]

[0041] in, Indicates direction; dir_sin and dir_cos represent sine and cosine encodings.

[0042] Furthermore, in step S25, the statistical characteristics are calculated as follows:

[0043] For a certain feature sequence , Let i represent the i-th feature in the feature sequence. The formulas for calculating each statistical feature are as follows:

[0044]

[0045]

[0046]

[0047]

[0048]

[0049] in, and This indicates taking the maximum and minimum values ​​of the feature values ​​within the sequence.

[0050] Further, step S3 is as follows:

[0051] Step S31, Data partitioning: Randomly select 1000 trajectory samples from the training set as the validation set, and use the remaining 6000 samples from the training set to train the classification model; the validation set is used to evaluate the model performance and guide parameter adjustment during the training process.

[0052] Step S32: Optimize the loss function: Assign different weights to different classes to make the model focus more on minority class samples during training; first calculate the class distribution of the training set, then set the weights inversely proportional to the sample size, as shown below:

[0053]

[0054] in, This represents the weight of the c-th category; Indicates the total number of samples; Let K represent the number of samples in the c-th category; K represents the number of categories.

[0055] The final weighted multi-class cross-entropy loss function is set as follows:

[0056]

[0057] in, Indicates sample The true labels belonging to category c are encoded using one-hot encoding, that is, if the sample If it belongs to category c, then ,otherwise 0, The model represents the samples The predicted probability of belonging to category c satisfies ;

[0058] Step S33: Optimize the hyperparameters of the three gradient boosting models, CatBoost, XGBoost, and LightGBM, respectively. Using GridSearchCV combined with cross-validation, evaluate the model performance under different parameter combinations on the validation set using accuracy, precision, recall, and F1 score to find the optimal parameter configuration. First, quickly locate the effective hyperparameter range through a broad-coverage, large-stride search, and then find the optimal hyperparameters that balance training accuracy and generalization ability through a small-stride, high-precision search. The key parameters for optimization of each model are as follows:

[0059] (1) Optimize the key parameters of the CatBoost model, including the learning rate, the maximum depth of the tree, and the number of iterations;

[0060] (2) Optimize the key parameters of the XGBoost model, including the learning rate, the maximum depth of the tree, and the number of generators, n_estimators;

[0061] (3) Optimize the key parameters of the LightGBM model, including learning rate, number of leaves, maximum tree depth, minimum number of samples per leaf node, and number of trees generated.

[0062] Further, step S4 is as follows:

[0063] Step S41, Optimal Model Selection: Based on the performance evaluation results of step S3, select the model with the best performance;

[0064] Step S42, Classification Result Output: Based on the best performing model, classify the fishing vessel operation types on the test set and output the operation type labels.

[0065] Compared with existing technologies, the present invention has the following advantages: the method of the present invention has a high accuracy rate in classifying fishing vessel operation types; it supports multi-class imbalance processing and improves the identification ability of small class samples; it has good generalization ability and real-time processing potential; and it can provide reliable technical support for fisheries supervision and illegal fishing detection. Attached Figure Description

[0066] Figure 1 This is a flowchart of the fishing vessel trajectory data processing of the present invention.

[0067] Figure 2 For feature extraction and data partitioning. Detailed Implementation

[0068] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings.

[0069] This invention provides a method for classifying fishing vessel trajectories based on a gradient boosting model, comprising the following steps:

[0070] Step S1, as follows Figure 1 As shown, fishing vessel trajectory data (including fishing vessel number, position, speed, direction, and timestamp) is obtained from the Automatic Identification System (AIS) and data preprocessing operations are performed, including: interpolating missing values, smoothing or removing outliers, and standardizing the time format.

[0071] The specific data preprocessing operations are as follows:

[0072] Step S11: The acquired raw fishing vessel trajectory data has missing fields. For trajectory segments with a small proportion of missing fields, linear interpolation is used to fill in the missing points, ensuring the continuity of the trajectory time. The linear interpolation formula is:

[0073]

[0074] in, It represents a value at a specific moment, specifically the position coordinates, velocity, or direction; This indicates the moment when a missing value exists. and Distance The two most recent consecutive moments with sampled values, .

[0075] If there is a serious lack of data or no interpolation basis, the corresponding trajectory data will be directly removed. Due to sampling errors or trajectory drift, some points may exhibit anomalies such as sudden changes in velocity or changes in direction angle. These are identified by setting thresholds for velocity and direction change rates, and then smoothed or removed to eliminate abnormal disturbances in the trajectory.

[0076] Step S12: Standardize the format of the time field in the original fishing boat trajectory data, processing it one by one according to mon dayhh:mm:ss, where mon represents the month; day represents the day; hh represents the hour of the day; mm represents the minute; and ss represents the second.

[0077] Step S13: In other stages of data cleaning, some features may have a small number of missing values ​​due to boundary conditions or data anomalies. To ensure data integrity and model stability, zero values ​​are used to fill in the missing values ​​to prevent abnormal interruption of the training process due to missing data.

[0078] Step S2, as follows Figure 2 As shown, 80 dimensions of features are extracted, including spatial features, temporal features, heading and direction features, dynamic change features, and statistical features; the specific implementation is as follows:

[0079] Step S21: Extract spatial features:

[0080] First, calculate the actual heading angle real_angle based on the coordinates:

[0081]

[0082] Where Δy and Δx represent the difference in the projected plane coordinates (unit: meters), and a minimum value of 0.00001 is added to avoid division by zero in the calculation, and the result is real_angle (unit: radians).

[0083] Further calculate the angular velocity real_angle_var:

[0084]

[0085] in, Indicates the time difference between two points (unit: seconds). This represents the actual heading angle at time t;

[0086] Next, the spherical distance between adjacent time points (haversine) is calculated, and the heading angle (bearing) is estimated using the spatial relationship between coordinate points.

[0087]

[0088]

[0089] Where R is the Earth's radius (6371 km). and The latitude of the two points is expressed in degrees, Δlat represents the difference in latitude (in degrees), and Δlon represents the difference in longitude (in degrees). Since trigonometric functions require input in radians, the latitude and longitude information needs to be converted from degrees to radians for calculation.

[0090] Step S22: Extract time features: Extract the hours and minutes from the time field separately, and use sine and cosine functions to encode them for time periodicity to capture the diurnal rhythm of fishing boat operations;

[0091] The time-periodic encoding is as follows:

[0092]

[0093] in, Indicates hours, Representing minutes, the cyclic time variable is converted into continuous features that are easy for the model to understand through sine and cosine encoding. and , and This represents sine and cosine encoding.

[0094] Step S23: Extract heading and direction features: Perform direction periodic encoding on the heading angle to alleviate the discontinuity of the angle value around 0° / 360°;

[0095] The directional periodicity encoding is as follows:

[0096]

[0097] in, Indicates direction (unit: degrees), dir_sin and dir_cos represent sine and cosine encoding.

[0098] Step S24: Extract dynamic change features: Perform first-order difference on the position latitude and longitude (x,y), velocity, and direction in the trajectory data to calculate the position increments x_var, y_var, distance increments distance_var, velocity increments velocity_var, and direction change angle_var to characterize the trend of the fishing boat's motion state.

[0099] Step S25: Extract statistical features: Perform statistical analysis on each feature according to the trajectory segment, and extract the mean, standard deviation, range, root mean square (rms), and waveform factor boxing to capture the overall distribution pattern of the local trajectory.

[0100] For example, a certain feature sequence , Let i represent the i-th feature in the feature sequence. The formulas for calculating each statistical feature are as follows:

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] in, and This indicates taking the maximum and minimum values ​​of the feature values ​​within the sequence.

[0107] Step S26: To ensure the uniformity of the input sequence, slice each trajectory with a fixed step size SEQ_LEN, and retain segments of appropriate length as model input samples to achieve an effective conversion from the original trajectory to the model input.

[0108] Step S3: Train three gradient boosting models—CatBoost, XGBoost, and LightGBM—using the training data. Weight the models using a loss function to focus on minority class samples, optimize hyperparameters using grid search, and evaluate performance on the validation set using accuracy, precision, recall, and F1 score. Details are as follows:

[0109] Step S31, Data partitioning: Randomly select 1000 trajectory samples from the training set as the validation set, and use the remaining 6000 samples from the training set to train the classification model; the validation set is used to evaluate the model performance and guide parameter adjustment during the training process.

[0110] Step S32: Optimize the loss function: Assign different weights to different classes to make the model focus more on minority class samples during training; first calculate the class distribution of the training set, then set the weights inversely proportional to the sample size, as shown below:

[0111]

[0112] in, This represents the weight of the c-th category; Indicates the total number of samples; Let K represent the number of samples in the c-th category; K represents the number of categories.

[0113] The final weighted multi-class cross-entropy loss function is set as follows:

[0114]

[0115] in, Indicates sample The true labels belonging to category c are encoded using one-hot encoding, that is, if the sample If it belongs to category c, then ,otherwise 0, The model represents the samples The predicted probability of belonging to category c satisfies ;

[0116] Step S33: Optimize the hyperparameters of the three gradient boosting models, CatBoost, XGBoost, and LightGBM, respectively. Using GridSearchCV combined with cross-validation, evaluate the model performance under different parameter combinations on the validation set using accuracy, precision, recall, and F1 score to find the optimal parameter configuration. First, quickly locate the effective hyperparameter range through a broad-coverage, large-stride search, and then find the optimal hyperparameters that balance training accuracy and generalization ability through a small-stride, high-precision search. The key parameters for optimization of each model are as follows:

[0117] (1) Optimize the key parameters of the CatBoost model, including the learning rate, the maximum depth of the tree, and the number of iterations;

[0118] (2) Optimize the key parameters of the XGBoost model, including the learning rate, the maximum depth of the tree, and the number of generators, n_estimators;

[0119] (3) Optimize the key parameters of the LightGBM model, including learning rate, number of leaves, maximum tree depth, minimum number of samples per leaf node, and number of trees generated.

[0120] Step S4: Select the best-performing model to classify fishing vessel operation types on the test set and output operation type labels. Details are as follows:

[0121] Step S41, Optimal Model Selection: Based on the performance evaluation results of Step S3, the model with the best performance is selected. In this example, the CatBoost model has an accuracy of 90.50% and an F1 score of 0.9045, indicating good inter-class balance performance. XGBoost achieves an accuracy of 92.20% and an F1 score of 0.9221, ranking second overall. LightGBM achieves an accuracy of 93.40% and an F1 score of 0.9338, performing best among all models. LightGBM possesses an efficient gradient boosting mechanism, a histogram-based splitting method, and a leaf-first decision strategy, enabling it to maintain high accuracy while also considering training speed and memory efficiency, making it suitable for large-scale trajectory classification scenarios.

[0122] Step S42: Classification Result Output: Classify the fishing vessel operation types on the test set based on the best-performing model and output the operation type labels. In this example, the best-performing LightGBM model is selected to classify the fishing vessel operation types on the test set and output the operation type labels.

[0123] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0124] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0125] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0126] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0127] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for classifying fishing vessel trajectories based on a gradient boosting model, characterized in that, Includes the following steps: Step S1: Obtain fishing vessel trajectory data from the Automatic Identification System (AIS) and perform data preprocessing operations, including: interpolating missing values, smoothing or removing outliers, and standardizing the time format. Step S2: Extract 80 dimensions of features, including spatial features, temporal features, heading and direction features, dynamic change features, and statistical features; Step S3: Train three gradient boosting models, CatBoost, XGBoost and LightGBM, respectively, using the training data. Weight the loss function to make the model focus on minority class samples, optimize the hyperparameters by combining grid search, and evaluate the performance on the validation set using accuracy, precision, recall and F1 score. Step S4: Select the model with the best performance to classify the fishing vessel operation types on the test set and output the operation type labels.

2. The method for classifying fishing vessel trajectories based on a gradient boosting model according to claim 1, characterized in that, In step S1, the fishing vessel trajectory data includes the fishing vessel number, location, speed, direction, and timestamp.

3. The method for classifying fishing vessel trajectories based on a gradient boosting model according to claim 1, characterized in that, In step S1, the data preprocessing operations are as follows: Step S11: The acquired raw fishing vessel trajectory data has missing fields. For trajectory segments with a small missing ratio, linear interpolation is used to fill in the missing points to keep the trajectory time continuous. If the missing data is severe or there is no interpolation basis, the corresponding trajectory data is directly removed. Due to sampling errors or trajectory drift, some points may exhibit anomalies such as sudden changes in speed and changes in direction angle. These are identified by setting thresholds for the rate of change of speed and direction, and then smoothed or removed to eliminate abnormal disturbances in the trajectory. Step S12: Standardize the format of the time field in the original fishing boat trajectory data, processing it one by one according to mon day hh:mm:ss, where mon represents the month; day represents the date; hh represents the hour of the day; mm represents the minute; and ss represents the second. Step S13: In other stages of data cleaning, some features may have a small number of missing values ​​due to boundary conditions or data anomalies. To ensure data integrity and model stability, zero values ​​are used to fill in the missing values ​​to prevent abnormal interruption of the training process due to missing data.

4. The method for classifying fishing vessel trajectories based on a gradient boosting model according to claim 3, characterized in that, In step S11, the linear interpolation formula is: in, It represents a value at a specific moment, specifically the position coordinates, velocity, or direction; This indicates the moment when a missing value exists. and Distance The two most recent consecutive moments with sampled values, .

5. The method for classifying fishing vessel trajectories based on a gradient boosting model according to claim 1, characterized in that, Step S2 is as follows: Step S21: Extract spatial features: First, calculate the actual heading angle real_angle based on the coordinates: Where Δy and Δx represent the difference in the projected plane coordinates, with a minimum value of 0.00001 added to avoid division by zero in the calculation; Further calculate the angular velocity real_angle_var: in, Indicates the time difference between two points. This represents the actual heading angle at time t; Next, the spherical distance between adjacent time points (haversine) is calculated, and the heading angle (bearing) is estimated using the spatial relationship between coordinate points. Where R is the Earth's radius. and The latitude of the two points is represented by Δlat, which represents the difference in latitude, and Δlon, which represents the difference in longitude. Since trigonometric functions require input in radians, the latitude and longitude information needs to be converted from degrees to radians for calculation. Step S22: Extract time features: Extract the hours and minutes from the time field separately, and use sine and cosine functions to encode them for time periodicity to capture the diurnal rhythm of fishing boat operations; Step S23: Extract heading and direction features: Perform direction periodic encoding on the heading angle to alleviate the discontinuity of the angle value around 0° / 360°; Step S24: Extract dynamic change features: Perform first-order difference on the position latitude and longitude (x,y), velocity, and direction in the trajectory data to calculate the position increments x_var, y_var, distance increments distance_var, velocity increments velocity_var, and direction change angle_var to characterize the trend of the fishing boat's motion state. Step S25: Extract statistical features: Perform statistical analysis on each feature according to the trajectory segment, and extract the mean, standard deviation, range, root mean square (rms), and waveform factor boxing to capture the overall distribution pattern of the local trajectory. Step S26: To ensure the uniformity of the input sequence, slice each trajectory with a fixed step size SEQ_LEN, and retain segments of appropriate length as model input samples to achieve an effective conversion from the original trajectory to the model input.

6. The method for classifying fishing vessel trajectories based on a gradient boosting model according to claim 5, characterized in that, In step S22, the time periodicity encoding is as follows: in, Indicates hours, Representing minutes, the cyclic time variable is converted into continuous features that are easy for the model to understand through sine and cosine encoding. and , and This represents sine and cosine encoding.

7. The method for classifying fishing vessel trajectories based on a gradient boosting model according to claim 5, characterized in that, In step S23, the direction periodicity is encoded as follows: in, Indicates direction; dir_sin and dir_cos represent sine and cosine encodings.

8. The method for classifying fishing vessel trajectories based on a gradient boosting model according to claim 5, characterized in that, In step S25, the statistical characteristics are calculated as follows: For a certain feature sequence , Let i represent the i-th feature in the feature sequence. The formulas for calculating each statistical feature are as follows: in, and This indicates taking the maximum and minimum values ​​of the feature values ​​within the sequence.

9. The method for classifying fishing vessel trajectories based on a gradient boosting model according to claim 1, characterized in that, Step S3 is as follows: Step S31, Data partitioning: Randomly select 1000 trajectory samples from the training set as the validation set, and use the remaining 6000 samples from the training set to train the classification model; the validation set is used to evaluate the model performance and guide parameter adjustment during the training process. Step S32: Optimize the loss function: Assign different weights to different classes to make the model focus more on minority class samples during training; first calculate the class distribution of the training set, then set the weights inversely proportional to the sample size, as shown below: in, This represents the weight of the c-th category; Indicates the total number of samples; Let K represent the number of samples in the c-th category; K represents the number of categories. The final weighted multi-class cross-entropy loss function is set as follows: in, Indicates sample The true labels belonging to category c are encoded using one-hot encoding, that is, if the sample If it belongs to category c, then ,otherwise 0, The model represents the samples The predicted probability of belonging to category c satisfies ; Step S33: Optimize the hyperparameters of the three gradient boosting models, CatBoost, XGBoost, and LightGBM, respectively. Using GridSearchCV combined with cross-validation, evaluate the model performance under different parameter combinations on the validation set using accuracy, precision, recall, and F1 score to find the optimal parameter configuration. First, quickly locate the effective hyperparameter range through a broad-coverage, large-stride search, and then find the optimal hyperparameters that balance training accuracy and generalization ability through a small-stride, high-precision search. The key parameters for optimization of each model are as follows: (1) Optimize the key parameters of the CatBoost model, including the learning rate, the maximum depth of the tree, and the number of iterations; (2) Optimize the key parameters of the XGBoost model, including the learning rate, the maximum depth of the tree, and the number of generators, n_estimators; (3) Optimize the key parameters of the LightGBM model, including learning rate, number of leaves, maximum tree depth, minimum number of samples per leaf node, and number of trees generated.

10. The method for classifying fishing vessel trajectories based on a gradient boosting model according to claim 1, characterized in that, Step S4 is as follows: Step S41, Optimal Model Selection: Based on the performance evaluation results of step S3, select the model with the best performance; Step S42, Classification Result Output: Based on the best performing model, classify the fishing vessel operation types on the test set and output the operation type labels.

Images