Global Tuna Fleet Dynamics Vary Across Continents

Understanding the pattern of fleet dynamics is important for effective fisheries management. Due to differences in fisheries policy, fishing equipment, economic development, social background and cultural characteristics, the fishing fleet dynamics may differ among different countries and regions. Furthermore, countries and regions within the same continent may share similar social, economic and technological conditions, leading to similar fishing fleet dynamics within continent^7,8,9,10. For example, Tidd et al.⁸ found that the dynamics of global artisanal fishing vessels are significantly influenced by climate and economic changes, while countries and regions within the same continent exhibit similar patterns in the fishing efficiency and adaptability of fishing vessels because they share similar climatic conditions and economic structures. Ye et al.⁹ assessed the fishing pressure on global fisheries and discovered that regions with similar social background and fisheries policy exhibited comparable levels of fishing pressure, i.e., developed regions experienced a decline in fishing pressure but economically disadvantaged regions saw a continuous increase in fishing pressure. Kroodsma et al.¹⁰ found that the behavioral differences exhibited by global fishing fleets are influenced far less by marine environments and seasons than by policies and cultures. Therefore, it is interesting to compare the dynamics of fishing fleet among different continents and explore their similarity. For continents with high similarity in fleet dynamics, national and regional fisheries management organizations may better cooperate to jointly develop fisheries policies. Conversely, regions with significantly different fleet dynamics may prefer to develop separate policies tailored to their specific conditions.

The dynamics of fishing vessel includes a variety of features, and the Automatic Identification System (AIS) has emerged as an important and cost-effective tool to extract those features¹¹. The widespread use of AIS provides real-time trajectory data, which can better uncover the dynamics of fleets by analyzing information such as vessel speed, course, and position. However, raw AIS data is large in quantity and updated frequently, so directly extracting behavioral features of fishing vessels from such high-dimensional data puts forward high requirements on data processing algorithms and computational resources. Moreover, the raw AIS data only includes dynamic information of the vessels, i.e., position, speed, and course, but lacks detailed information directly related to the vessel behavior, i.e., gear type, fishing time and distance¹². This information is crucial for analyzing the behavioral characteristics of fishing vessels, but it cannot be directly obtained from AIS data.

Traditional dimensionality reduction methods, such as principal component analysis (PCA), may not well handle the complex nonlinear structure like AIS data¹⁰. These methods tend to lose important information during the dimensionality reduction process^13,14, such as failing to effectively preserve dynamic temporal variations, and the features obtained after dimensionality reduction are usually linear combinations of the original features, which may not directly correspond to the specific behaviors of fishing vessels¹⁵. To explore the characteristics among various types of fishing vessels, many studies have proposed a variety of methods for extracting fishing vessel features from raw trajectory data. For instance, Farrell et al.¹⁶ utilized the temporal and spatial information of vessels to design features such as sailing time, distance, and three related speed metrics, integrating these with course and the depth of fishing locations to accomplish the task of identifying whether a vessel is engaged in fishing activities. Yan et al.¹⁷ constructed geometric features of vessels, i.e., the length and width of vessels, and behavioral features, i.e., the distance, position, and speed, to differentiate among cargo, tanker, fishing, passenger, and tug vessels. These methods extract the static and dynamic characteristics of diverse vessels from multiple perspectives, which provides a foundation for exploring the dynamics of fishing vessel across different regions. However, for different types of fishing vessels, there are differences in their hull and dynamic features, e.g., purse seine vessels often make sharp turns at high speeds during fishing, while longline vessels maintain a straight line and sail at low speeds¹⁸. Additionally, marine environment could change the actual movement trajectories of fishing vessels, but the feature information cannot directly reflect the action of these external forces, resulting in deviations in the interpretation of fishing vessel behavior. For example, a fishing vessel showing low-speed sailing might be due to sailing against a current rather than engaging in fishing operations. Therefore, it seems insufficiently comprehensive to quantify the dynamics of various types of fishing vessels by the multiple feature factors.

In response to the above challenges, this study proposed a novel approach based on machine learning algorithm for the first time to explore the similarity of large-scale tuna fishing fleets dynamics among different continents. The fundamental principle is that if the fleet dynamics between two continents are similar, then a machine learning model trained on dataset from one continent should also be applicable for predicting dataset from the other continent. If prediction fails, it indicates differences in fleet dynamics between the two continents. The process is as follows. First, based on the AIS data of two important types of tuna fishing vessels, longline (LL) and purse seine fleets (PS), static and dynamic features of fishing vessels across six continents were constructed. Fishing vessels from flag states in the same continent are regarded as a fleet. Subsequently, to ensure comprehensive comparison of static and dynamic features, we designed a two-stage experiment which including and removing static features by utilizing the extreme gradient boosting (XGBoost) algorithm to train binary classification model of each continent and quantify the dynamics similarities of fishing vessels. Finally, we ranked the important features which influencing the results of classification models, and analyzed the similarities of tuna fishing vessels dynamics across different continents.

Features distribution of tuna fishing fleets in different regions

The length and tonnage of fishing vessels are typical static features, reflecting the basic physical properties and operational capabilities of vessels. Dynamic features are attributes of fishing vessels that change over time during operations, reflecting the actual operational behaviors and patterns of the fleets. LL and PS are important components of the global tuna fishery, both their static and dynamic features exhibit variation to a certain extent across different fleets of continents due to differences in their target species, fishing capability limitations, economic development and culture. We drew the distribution maps of features of LL and PS across six continents as shown in Fig. 1.

For the two static features, length and tonnage, the peak distribution characteristics of LL across different continents do not show a clear pattern. However, PS on different continents exhibits relatively similar peak distributions but except for Europe. In the dynamic features, the ridge maps of sailing distance and fishing time demonstrate that the consistent distribution trends across different fleets of continents, while the values at which the peaks occur differ. It is noteworthy that the longitudinal distribution trend of tuna fleets shows no commonality across continents, but those of Asia and Oceania are remarkably comparable. There are also distinct differences in the latitudinal range of tuna fleets among continents, with Europe exhibiting the widest distribution and South America the narrowest. Regarding the distribution characteristics of course and speed, the distribution ranges of tuna fleets across all continents are surprisingly almost identical, with the similar number of peaks, but the specific numerical values at which these peaks occur differ.

Feature importance of different regional classification models

According to the differences in the features of two types of fishing vessels, this study constructed a 24-dimensional mixed feature. This feature includes two static features, length and tonnage, and 22 dynamic features, which are the mean (MEAN), upper quartile (UP), median (MID), lower quartile (LOW), and standard deviation (STD) of the vessels’ longitude, latitude, speed, and course, denoted as \({lon\_mean}\), \({lon\_up}\), \({lon\_mid}\), \({lon\_low}\), \({lon\_std}\), \({lat\_mean}\), \({lat\_up}\), \({lat\_mid}\), \({lat\_low}\), \({lat\_std}\), \({speed\_mean}\), \({speed\_up}\), \({speed\_mid}\), \({speed\_low}\), \({speed\_std}\), \({course\_mean}\), \({course\_up}\), \({course\_mid}\), \({course\_low}\), \({course\_std}\), as well as the sailing distance and operational time. To further explore the feature differences between two types of fishing vessels across various continents, we employed the XGBoost algorithm to rank the importance of features in distinguishing between LL and PS, as illustrated in Fig. 2a.

The results of the feature importance ranking indicate that, except for Africa where the second most important feature is \({lat\_mean}\), the top two features for other continents are static features, namely length and tonnage. Especially in South America, these two static features can even clearly distinguish between LL and PS. This finding suggests that the differences in hull characteristics between LL and PS across continents are more pronounced compared to other features. However, while static features are undoubtedly important, the dynamic features of the fleets more effectively reflect the behavioral patterns and regional preferences of different types of fishing vessels during operations, providing valuable insights into fleet dynamics. Therefore, to mitigate the dominance of static features in the classification model, we removed static features and conducted the feature importance ranking again, with the results shown in Fig. 2b.

The ranking results after removing static features reveal both similarities and differences in the important features that distinguish LL and PS across different continents. For African fleets, the upper limit of longitude (\({lon\_up}\)) is the most significant differentiating factor, followed by the fluctuation in course (\({course\_std}\)) and speed (\({speed\_std}\)). The most important distinguishing features of Asian fishing vessels is \({speed\_std}\), with \({lo}{n\_up}\) and \({lat\_up}\) coming next. In Europe, both the upper and lower limits of longitude (\({lo}{n\_up}\), \({lo}{n\_low}\)) and latitude (\(l{at\_up}\), \(l{at\_low}\)) are crucial factors. For North American fleets, \({lat\_up}\), \({lon\_low}\), and \({lon\_mean}\) are significant. In Oceania, \({speed\_std}\) and \({speed\_up}\) are the most important differentiating factors. The \({lo}{n\_low}\) and \({lo}{n\_std}\) are key distinguishing features in South America, followed by the features of \({course\_std}\) and \({speed\_std}\).

In summary, the main factors distinguishing LL and PS in Africa and South America are determined collectively by location, course, and speed information. In Asia and Oceania, differentiating between the two types of fishing vessels relies mainly on speed and location information. For European and North American fleets, the differences are primarily determined by location information.

Similarity analysis of tuna vessel behavior in different regions

To evaluate the degree of feature similarity between LL and PS among different continents, this study used the XGBoost algorithm to construct binary classification models for training datasets from six continents, both including and removing static features. Then, each classification model was used to cross-predict independent testing datasets from other continents. And each independent testing set comprised 1000 LL samples and 1000 PS samples. The classification performance was evaluated using four metrics: \({Accuracy}\), \({Precision}\), \({Recall}\), and \(F1\). If one of the metrics fell below 60%, the model of one continent is considered not transferable to the other continent. The results of cross-validation on the four metrics are shown in Fig. 3 and the schematic diagram of cross-validation results across all continents is shown in Fig. 4.

For the cross-validation results considering all features (Fig. 3a), all classification models from a given continent demonstrated the best classification performance on their own continent’s independent testing dataset, with \({Accuracy}\) even reaching as high as 100%. The cross-prediction results between other continents mostly exceeded 70% across all metrics, which means there are similarities in static features between continents (Fig. 4a). However, European model is applicable to fleets from other continents, but all models from other continents failed to achieve an \({Accuracy}\) exceeding 61% on the European testing set, with \({Recall}\) not surpassing 30%, except for the African classification model, which achieved an \({Accuracy}\) of 75.9%. It seems that the models on all other continents except African are unable to predict the testing set for Europe (Fig. 4b), which reflects the universality and diversity of the static features of the European tuna fishing fleets. Combined with the ranking of feature importance, the tonnage feature of fleets in Africa and Europe may be more similar. In addition, only the European classification model considers sailing distance as an important feature, which is one of the unique aspects of the European model.

When only considering the dynamic features, the cross-validation results underwent a dramatical transformation (Fig. 3b), which no transferability between classification models on most continents. However, the model transferability was high between the Asian and Oceanian fleets, with \({Accuracy}\) even surpassing 90%, indicating significant similarity in the behavioral patterns between Asian and Oceanian fleets (Fig. 4d). From the feature ranking results, it is evident that only Asia and Oceania have the standard deviation of vessel speed (\({speed\_std}\)) as their dominant predictor, suggesting that the variability in speed is most similar between Asian and Oceanian fleets. Moreover, the European model exhibited remarkable distinctiveness (Fig. 4e), whose predictions for the testing sets of all other fleets exceeded 60% on all metrics but not vice versa. The results indicates that the dynamic features of European fleets may be highly representative and general, potentially encompassing the behavioral patterns of fleets on other continents, while the characteristics of fleets on other continents may be more geographically constrained. Notably, the North American model respectively achieved prediction \({Accuracies}\) of 67.3%, 76.9%, and 71.9% for the testing sets of Asian, Oceanian, and South American fleets, implying its dynamic features not only share commonalities with these regions but also retain distinctive elements. In contrast, no transferability was observed between African and other continental fleets apart from European fleets, nor between Asian and South American fleets (Fig. 4f). That is, there are large divergences in dynamic patterns between African and other continental fleets, as well as between Asian and South American fleets.

The findings of this study have significant implications for global tuna fisheries management. The behavioral patterns of global tuna fishing fleets are often influenced by a variety of factors. Countries in the same region usually have similar characteristics of their fleets due to shared human customs¹⁰, fisheries policies¹⁹, and fishing practices of fishers²⁰. For example, Europe has implemented EU-level policies such as the common fisheries policy¹⁹, while Southeast Asia has adopted the ecosystem approach to fisheries management to expand fisheries management scope²¹. Therefore, analyzing behavioral patterns of tuna fishing fleets at the continental scale facilitates the coordination of regional and national fisheries management measures for similar continents and develops specific regional management plans for continents with different characteristics, thereby improving management efficiency.

Understanding the dynamic patterns of fleets across different regions is a critical step in conserving fishery resources. The study of the similarity of the dynamic characteristics of tuna fleets on different continents provides new insights for ecosystem protection, sustainable marine resource development, and effective fisheries management. By analyzing the latitude and longitude characteristics of each fleet activities, it is possible to identify fishing hotspots. If multinational fishing fleets operate with similarly high-intensity patterns in the same waters, this area may exist critical tuna habitats. However, the cumulative fishing pressure from these fleets could threaten local marine ecosystems (e.g., coral reefs, spawning grounds). In such cases, the tRFMOs could implement dynamic Marine Protected Areas (MPAs) activated during peak fishing seasons to achieve ecosystem conservation and sustainable development of marine resources. Tuna is a highly migratory species, with populations spanning the exclusive economic zones and high seas of several continents². Exploration of the similarities in fleet dynamics can facilitate coordinated global or regional resource assessments, assisting tRFMOs in establishing uniform or specific management measures, such as fishing catch quotas, seasonal closures, and marine protected areas, to prevent resource underestimation or overfishing due to fragmented governance. By revealing the commonalities and differences in fleet dynamic across continents, it offers a scientific foundation for multi-scale (national, regional, global) collaboration among stakeholders (governments, industry, non-governmental organizations). For fleets with unique dynamics, it is necessary to develop fisheries policies adapted to local conditions, rather than uniform international regulations. For fleets with high similarity, cross-regional cooperation can be achieved by implementing joint monitoring systems (e.g., sharing AIS tracking and coordinating patrols), thereby enabling effective fisheries management.

Social policies, economic development, and cultural traditions profoundly influence the dynamics of fishing fleets across global continents, shaping unique patterns of fleets in different regions⁷. For example, Kroodsma et al.¹⁰ revealed behavioral variations of global fishing fleets across regions, showing global fishing patterns have surprisingly low sensitivity to short-term economic and environmental changes but strong response to political events and culture. Ye et al.⁹ assessed the fishing pressure on global fisheries and found that regions with similar social background and fisheries policy showed comparable levels of fishing pressure. These findings consistent with our results that there are distinct dynamic patterns of tuna fishing fleets among different continents. Previous studies have shown that economic, cultural, and political factors significantly influence fishing fleet dynamics across nations. For different levels of economy, Sumaila et al.²² pointed out that economically powerful regions, e.g., the European Union (EU), expand their distant-water fleets through subsidies, resulting in high mobility and long voyages of fishing fleet dynamics. Schuhbauer et al.²³ found that differences in subsidy policies lead to significant gaps in fishing efficiency and automated technology between offshore fleets of developing countries (e.g., West Africa) and distant-water fleets of industrialized countries (e.g., Japan, Korea). In the traditional cultures, Kroodsma et al.¹⁰ observed reduced fishing activities during holidays like Chinese New Year and North American Christmas. Cohen et al.²⁴ noted that Pacific Island nations (e.g., Solomon Islands) exhibit small-scale and seasonal fishing dynamics due to community-based tradition, in contrast to industrialized operations of developed countries. In addition, the policy and regulatory intensity of different regions will also make the dynamics of fishing vessels different. Hanich et al.²⁵ compared the policies of Pacific Island countries with those of the EU, and found that island countries (e.g., Palau) restrict foreign fleets through strict licensing systems, while the quota allocation of the EU Common Fisheries Policy leads to competitive fishing by member fleets. Thus, analyzing the fishing fleet dynamics provides insights into historical, economic, and sociocultural drivers of fisheries.

The results of this study show that the classification models trained by European have high transferability across continents when only considering the dynamic features, and the North American model can predict testing sets of Asia, Oceania, and South America, but not vice versa. It suggesting that the European and North American fleets exhibited more diverse and generic dynamics than the other continents. The broad dynamics of European and North American tuna fishing fleets related to their long history of distant-water fisheries²⁶ and transregional fishing activities²⁷. Distant-water fisheries in Europe emerged as early as the late 15th century²⁶, and their colonists (mainly the British, French, and Spanish) initiated distant-water fishing along the eastern coast of North America, such as the Newfoundland fishing grounds²⁸. Since distant-water fisheries developed earlier with advanced fishing technologies in Europe and North America^25,26, the design of their vessels is more automated and the technological standardization²⁵ is more improved compared to other continents. Furthermore, the European and North American tuna fishing fleet operate across the Atlantic and Pacific oceans²⁷, with diverse range of fleets and extensive fishing areas, making it more likely to cover the dynamics characteristics of fleets from other regions. Additionally, the universality of tuna fishing fleets is also linked to fisheries cooperation with some developing countries. For example, the EU has established sustainable fisheries partnership agreements with west African countries^29,30, and the United States has signed bilateral fisheries agreements with some Oceanian countries. These partnerships influence fleet dynamics by shaping access to fishing grounds, resource allocation, and economic dependencies. The global nature of European and North American fleets indicate that cross-regional fisheries management is very important, requiring strengthened communication among regions to prevent overfishing.

On the contrary, the model trained by African fleets performed poorly when applied to vessel classification for the other continents, which exhibiting a completely different fleet dynamics from other continents. It may be because African tuna distant-water fisheries are mostly dependent on foreign fleets (e.g., EU) and domestic fleets of Africa are mostly limited to offshore operations³¹, compounded by developmental constraints³² and weak industrialization capacity. For example, per capita seafood consumption of Africa is much lower than the global average³³. Thus, the African tuna fishing fleets shows significant differences in behavioral patterns of tuna fishing fleets compared to other continents, which should implement regionalized fisheries management measures according to its own development conditions. These findings may be related to the historical development of tuna fisheries across continents, emphasizing the critical role of social and economic factors in determining the dynamics of global tuna fishing fleets. It suggests that RFMOs should not implement the same policies for fisheries management in different regions, but rather develop diversified fisheries management measures to address the distinct developmental needs of different areas.

Notably, the model transferability was high between the Asian and Oceanian fleets, implying strong similarity of their fleet dynamics. Consistent with the findings of Kawamoto et al.³⁴, small-scale tuna longline fisheries in Japan and Australia have much in common, including gear, target species, and relatively small-scale entities. The similarity of the fleet dynamic of two continents is likely attributable to the waters under the jurisdiction of the WCPFC are dominated by countries in Asia (e.g. China, Japan and Korea) and Oceania (e.g. Australia), where most countries operate under the same fisheries policies, cooperate in assessing stocks, and share data³⁵. Moreover, regional cooperation policy of economic and fishery between Asia and Oceania is also one of the reasons. For example, Asia-Pacific Economic Cooperation have jointly promoted trade and sustainable development³⁶, and China has funded port infrastructure improvements in some Oceanian countries³⁷.The dynamics similarity in tuna fishing fleets between Asia and Oceania suggests that tRFMO (e.g., WCPFC) should enhance cooperation to establish universal fisheries management policies for countries with similar dynamic characteristics of the fleets. Analyzing dynamic similarities of global tuna fishing fleets at the continental scale provides new directions for marine management, helping regions and countries formulate effective fisheries management policies to reduce illegal fishing, optimize resource allocation, and jointly promote sustainable fisheries development.

Although this study revealed the similarities and differences of global tuna fishing fleets among different continents, it still has certain limitations. Firstly, the analysis at the continental scale may obscure the heterogeneity of tuna fishing fleets from different countries within the same continent. For example, Asia encompasses major fishing countries such as China and Japan, as well as Southeast Asian countries, where fleet characteristics may vary due to differences in national development levels, fisheries management policies, or operational practices. Therefore, regionalized analysis of tuna fishing fleets by continent alone may ignore information at the national scale. Furthermore, the completeness and accuracy of AIS data may be constrained by regional variations in AIS equipment installation rates, signal coverage, and data reception quality, potentially leading to missing or biases data, which could reduce the model generalizability. While XGBoost algorithm performs well in classification tasks, it depends on training data distribution, which may affect cross-regional validation reliability when applied to larger-scale fishing vessel trajectory datasets or severely imbalanced data categories. In the future, researches could focus on the specific continents or regions, further dividing countries or economic regions (e.g., Northern and Southern Europe) according to historical background, economic development, policies and cultural factors to compare fleet dynamic characteristics. Additionally, integrating data such as VMS, fishing logs, or port records, could complement the limitations of AIS data and improve the comprehensiveness of the datasets. For large-scale data analysis, incorporating deep learning architectures (e.g., long short-term memory or transformer networks) could better explore the dynamic patterns of fishing fleets.

This study is based on a basic principle that if the fleet dynamics between two continents are similar, then a machine learning classification model trained on data from one continent should effectively predict data from the other, and constructed a cross-validation classification models for six continents. First, global AIS data for LL and PS were partitioned by continent. Next, the dataset of each continent was created to combine the static and dynamic features of fishing vessels, and six balanced feature sets were constructed based on the number of samples with fewer LL and PS. From each balanced feature set, 1,000 LL and 1,000 PS samples were randomly selected to form independent testing datasets for model validation, and the remaining data (after removing the testing data) served as training datasets to train the specific continent classification models. Subsequently, the XGBoost algorithm was used to train the binary classification models on the training datasets of Africa, Asia, Europe, North America, Oceania, and South America, then tested on the independent testing datasets on other continents. Finally, by analyzing the results of feature importance rankings and cross-validation results calculated by the XGBoost algorithm, the similarities of fleet dynamics on different continents were discussed. The methodological framework and workflow for assessing similarities are illustrated in Fig. 5.

AIS data

This study is based on the AIS data of the global tuna LL and PS. Firstly, the list of tuna LL and PS fishing vessels was obtained from the International Seafood Sustainability Foundation (ISSF) website (https://www.iss-foundation.org/vessel-and-company-commitments/) and focuses solely on large-scale tuna fishing fleets. Then, according to the corresponding maritime mobile service identity (MMSI) numbers of the fishing vessels, the AIS data was downloaded from the website (https://globalfishingwatch.org/map) of Global Fishing Watch (GFW). Finally, the AIS data was divided by continent, with the year 2023 selected as an example. The number information of tuna fishing vessels and tracks used in this article is presented in Table 1, and the raw data shown in Fig. 6 illustrates the range and trajectory points on each continent.

Extraction features of tuna fishing fleets

Feature extraction is a critical step in identifying vessel types and grasping fishing activity patterns. Trajectory-based methods (e.g., Dynamic Time Warping, DTW) typically require complete sailing sequence data, but AIS coverage gaps or reporting variations often lead to missing data across regions. In this paper, the static and dynamic features of fishing vessels were constructed, which can effectively capture the key behavioral differences between LL and PS, and ensure the robustness of the classification model even under fragmented data conditions.

Static features

Static features are inherent attributes of fishing vessels that do not change with the operating environment or time. The length and tonnage of a fishing vessel are typical static features, reflecting the vessel’s basic physical properties and operational capabilities. The length of a fishing vessel is usually related to its operational range and capacity. Larger vessels may be more suitable for pelagic operations, while smaller vessels may be better suited for offshore operations. Tonnage reflects the load-carrying capacity of a fishing vessel, typically related to the scale of operations, fuel capacity, and fishing capability. Larger tonnage may indicate that the vessel can carry more fishing gear and catch, making it suitable for long-duration, long-distance operations¹⁸. The similarity in static features may suggest consistency in the operational scale and economic investment of fishing vessels across different continents. Therefore, based on the information provided by GFW, the length and tonnage of LL and PS on each continent were selected as the static features in the study.

Dynamic features

Dynamic features are attributes of fishing vessels that change over time during operations, reflecting the actual operational behaviors and patterns of the fleets. In order to explore the similarities and differences in dynamic features between the two types of fishing vessels across different continents, we divided the AIS data of all tuna fishing vessels by day and then analyzed them using statistical values. The most commonly used statistical methods, such as MEAN, MID, UP, and LOW, can reflect the overall trend, central tendency, distribution range, and boundary information of the data, and STD can indicate the dispersion and variability of the data. These statistics can better capture the diversity and complexity of AIS data and describe the distribution of dynamic features of fishing vessels from multiple perspectives. Therefore, we selected the MEAN, UQ, MID, LQ and STD of longitude, latitude, speed and course of each fishing vessel per day in AIS data as part of the dynamic features, resulting in a 20-dimensional dynamic feature vector.

Additionally, sailing distance and fishing time are direct indicators of fishing vessel activity intensity, providing a clear reflection of the activity levels and fishing effort of the vessels, thereby contributing to a better understanding the fleets dynamics across different continents. Therefore, we also added the daily sailing distance and fishing time of fishing vessels as part of the dynamic features. Based on the recorded timestamps in the AIS data, the fishing time of the fishing vessel was calculated by summing the time intervals between consecutive timestamps, the calculation formula is as follows:

where \({t}_{i}\) represents the current time point of the fishing vessel’s operation recorded in AIS, \({t}_{i+1}\) denotes the next time point. The sailing distance is obtained by calculating the Euclidean distance between longitude and latitude at each time interval and summing the results, using the following formula:

where \({{lon}}_{{t}_{i}}\) and \({{lat}}_{{t}_{i}}\) are the longitude and latitude of a fishing vessel at the current time \({t}_{i}\), \({{lon}}_{{t}_{i+1}}\) and \({{lat}}_{{t}_{i+1}}\) are the longitude and latitude at the next time point \({t}_{i+1}\). Ultimately, we obtained a 22-dimensional dynamic feature vector.

XGBoost algorithm

In this study, the static and dynamic features constructed by using AIS data are structured data, and XGBoost serve as the benchmark algorithm for processing such data³⁸. As a machine learning algorithm based on gradient boosting decision trees, XGBoost ensembles multiple weak learners to construct a strong learner, which has superior performance in classification tasks³⁹. The algorithm can automatically select key features, avoid interference from irrelevant features, effectively capture nonlinear relationships in data, achieving high accuracy when fitting complex datasets⁴⁰, and its parallel processing capability further enhances training efficiency⁴¹. Conventional machine learning models (e.g., support vector machines and logistic regression) have weaker nonlinear feature extraction capabilities and heavy dependence on feature engineering⁴², and deep learning models better suited for unstructured data such as images, text, time-series signals⁴³. Due to the limited number of samples obtained from global tuna fishing vessels and the large redundant information in AIS data, XGBoost can automatically handle missing values and has built-in regularization to prevent overfitting with small-scale samples³⁹. But some machine learning models are sensitive to noise data⁴¹, and deep learning models often require massive data to prevent overfitting⁴⁴. Furthermore, XGBoost is easy to adjust parameters and has built-in feature importance scores³⁹, which helps to intuitive determine the contribution rate of static and dynamic features to classification models among continents. While random forest algorithm can also provide feature interpretability, it generally inferior to gradient-optimized XGBoost in prediction accuracy^45,46, and deep learning models are often difficult to directly explain feature effects and adjustment parameters are very complex⁴⁴. Therefore, this study selected the XGBoost algorithm to train binary classification models for tuna fishing fleets among continents in order to make more accurate predictions.

Specifically, the maximum tree depth was set to 3 to control the model complexity, the learning rate was set to 0.1 to ensure stable convergence, the L1 and L2 regularization parameters were set to 0.5 and 1.5 to balance bias and variance. Besides, in order to reduce the evaluation bias caused by a single random data split, five-fold cross-validation was adopted to ensure the assessment reliability. The core principle involves evenly dividing the dataset into five non-overlapping parts, iteratively selecting one as the validation set and the remaining four as the training set. This process is repeated five times, with the average performance of the five results as the final model evaluation. The parameter configurations and functional descriptions of the XGBoost model in this paper are detailed in Table 2.

Classification performance evaluation

To comprehensively evaluate the performance of binary classification models for cross-prediction on different continents from different perspectives, this study employed four key evaluate metrics: \({Accuracy}\), \({Precision}\), \({Recall}\), and \(F1\). \({Accuracy}\) measures the overall proportion of correct classifications, reflecting the general performance of the model. \({Precision}\) represents the proportion of true positive predictions among all positive predictions, focusing on the reliability of the model’s positive outcomes. \({Recall}\) evaluates the ability of model to identify actual positive samples, reflecting its detection completeness. The \(F1\) combines the results of \({Precision}\) and \({Recall}\) to provide a balanced assessment. For the four metrics of performance evaluation, the values range from 0 to 1, and the higher the value is, the more accurate the classification results obtained by the model. The relevant calculations are as follows:

where \({TP}\), \({TN}\), \({FP}\), and \({FN}\) respectively represent the numbers of true positive, true negative, false positive, and false negative samples¹².