This study processes the 19 dynamic environmental factors and the calculated temperature gradient, thermocline temperature and 22 EKE variables in total to establish a regression model for the fishing effort of the vessel and calculate the score of the variance inflation factor. The results show that the highest score is therm, and the VIF score reached 55. There was a serious multicollinearity problem among the 22 independent variables obtained. The correlation coefficients between 22 independent variables were calculated, and the results are shown in Supplementary Material.

The results show that therm and sst, chl0 and pp0, and tem200 and sss are all highly correlated, and the correlation coefficients are all above 0.9. To avoid the influence of multicollinearity on the model, according to a study of tuna habitat conditions and marine environmental factors by relevant scholars, we deleted chl0 and tem200 for reference and kept pp0 and salt200. Between therm and sst, related studies have shown that sea surface temperature has an important influence on the distribution of purse seine tuna fishing grounds. This article chooses sst.

Figure 2 shows the monthly distribution of the spatial trajectory points of the tuna purse seine from 2015 to 2019. According to the trajectory of the vessel in the image, the space coverage of the vessel is 140°-210°E, 15°S-10°N. There was no significant change in the area covered by vessels from January to June. Starting in July, the concentrated area of vessels decreased and was mainly distributed near the equator, with the lowest number of vessel locations in September. Starting in October, the area of vessels expanded to the south.

Figure 3 is a superimposed spatial map of the fishing effort predicted by the model in 2020 and the actual fishing effort of fishing vessels in 2020. The image shows that each month has a good spatial distribution contrast.

Based on the results of two model, the BRT model was used to assess the effects of key environmental parameters on potential tuna purse seine fisheries. According to the space map from January to June, the areas with the largest amount of fishing effort predicted by vessels are concentrated in 150°E to 160°W and 10°S to 5°N. The images in January and February show that there are three areas with significant fishing effort-intensive forecasts [(1) 150°E to 170°E, south of the equator to 10°S; (2) 165°E to 180°, 5°S to 5°N; and (3) 170°W to 155°W, south of the equator to 10°S]. The actual fishing effort of vessels covers these three dense areas and spreads in these three areas. The images in March and April showed that the third fishing effort intensive area disappeared, but the first two areas expanded and merged and concentrated at 140°E to 180°, 10°S to 5°N, and the actual fishing effort also has the corresponding change trend. Compared with March and April, the images in May and June have no obvious change trend.

From July to December, the concentrated area of the spatial image decreased significantly. The images in July, August, and September show that starting in July, the predicted fishing effort area has decreased and become more discrete, extending eastward as a whole, and the actual fishing effort of vessels has also decreased. Until October, the forecasted fishing effort area of vessels in October and November became larger and concentrated in 140°E to 170°W, 10°S to 5°N, and the actual fishing effort of vessels increased. By December, the image predicts that the fishing effort of vessels will be less spatially distributed mainly in some small areas from the south of the equator to 10°S. The actual fishing effort of the vessels, that is, the staying time of the vessels in the unit grid, is more consistent with the area where the fishing effort of the vessels is concentrated every month.

The average contribution rate of the variable value to the fishing effort of the vessel see Supplementary Material for details. It can be seen from the bar chart as a whole that tem100 and lon are the most important factors affecting the fishing effort of tuna purse seiners, and their contribution rates to the fishing effort of vessels are 12.38% and 9.76%, respectively; second, sst, lat and DSH have an effect on tuna. The fishing effort of purse seine vessels also contributed significantly, accounting for 9.57%, 8.75%, and 7.11%, respectively. The cumulative contribution rate of 5 variables, including tem100, sst, latitude and longitude and DSH, reached 47.56%. Among them, sss, pp0, ssh, salt200, DO, and DO200 also have a certain influence, and the contribution rate is basically approximately 4%. The contribution rate of salt100, DPT, mld, and u is relatively small, basically approximately 3%; the contribution rate of DO100, pp200, pp100, chl200, v to the fishing effort of vessels is approximately 2%; and Depth, EKE, chl100, and sstf have the smallest contribution rates, basically approximately 1%.

Figure 4 is an image of the relationship between important environmental variables and predictive factors in each month. The figure reflects the range of environmental factors corresponding to the most likely fishing effort and the specific contribution rate of environmental factors. The first four important environmental variables in each month explain the contribution rate to the predictive factors. Figures (a) to (l) are, respectively January to December. Among them, the monthly results of Figures (a), (b), (c), (d), (e), (i), (j), (k), (l) show that the contribution of tem100 to the fishing effort of fishing vessels are all between 10% and 20%, tem100 is concentrated at 25-29°C, and the curve fluctuates greatly. Second, the explanation rate of sst and DSH is also high, sst is mainly concentrated in 29-31°C, and DSH is mainly concentrated in 100-300 km; in addition to Figure (i), lon is in the first four important environmental variables of each month, mainly concentrated in the sea west of 180°E. In addition to Figures (a), (j), (k), and (l), lat is also in the first four important environmental variables of each month; it is more concentrated in the 5°S-5°N region, and the contribution rate of lon and lat is relatively large.

Previous studies on purse seine tuna mostly used environmental variables at the surface of the ocean, but studies have shown that bonito, yellowfin tuna and bigeye tuna mostly caught by purse seine tuna live in the water layer of 50-100 m underwater. Purse seine tuna is mostly operated in economic exclusivity, and the distance from the vessel to the coastline has a certain impact on fishing options. This paper uses 24 variables, including dynamic variables of the ocean environment, spatial variables and static variables such as the distance to the coastline, and it uses the BRT model to clarify the impact of different variables on the spatial distribution of fishing operations.

Related research scholars have carried out many studies on the preferred habitat of tuna (Candan et al., 2014; Huang et al., 2020; Wang, 2020). These studies mainly used the collected fishery log data combined with environmental values to describe the tuna habitat. Few analyses of tuna fish habitats are based on satellite AIS data. Therefore, compared with previous work, this study uses AIS data to extract the fishing effort of the vessels as a predictor variable of the BRT model and GAM model, which is a more true and effective response to the behaviour of the vessels than traditional fishery data. The purpose of the tuna purse seiner on the sea is to fish, and every time it stays on the sea for a period of time, it will spend a corresponding or additional cost value.

Nicolas et al. (2011) used vessel monitoring system (VMS) data to study and quantify the spatial dynamics of Seschel tuna purse seine fishing activities in the Indian Ocean. The Bayesian state space model is used to divide movement on the sea into three situations: sailing, tracking and fishing. The spatial trajectory distribution described by the length of stay of the vessel is discussed from a statistical perspective. Zhang et al. (2021) analysed the spatial analysis of the fishing behaviour of tuna purse seiners in the western and central Pacific based on the AIS trajectory data of vessels and used the vessel’s residence time on the sea as the fishing effort, which was correlated with the data of the western and central Pacific tuna purse seine production and net times. The test results show that there is a high degree of positive correlation (r>0.8). Therefore, it is reasonable to define the time spent by vessels on the sea for catching tuna as fishing effort. Fishing effort is a measure of the degree of accumulation of tuna fish stocks and changes in the spatial behaviour of purse seining tuna vessels. Hsu et al. (2021) used the fishing effort of longline vessels to construct a habitat index model and successfully predicted the high-precision spatial and temporal distribution map of skipjack fisheries in the western and central Pacific, and the comparison results showed that fishing effort data were better than catch data. This indicates that the environmental factor response model constructed based on the fishing effort information of vessels can be used to understand the spatial distribution of fishery resources and environmental impacts.

In research on the prediction of marine environmental factors and fishery resource distribution, researchers mostly use the habitat suitability index (HSI), support vector machine (SVM) and other models to conduct research, which easily leads to overfitting, thereby reducing the prediction accuracy of the model. At present, research scholars (Chen et al., 2013; Megan et al., 2019;Hou et al., 2020) have applied methods such as random forests and gradient boosting trees in ensemble learning to fishery forecast research, but these methods are all combined with fishery data. Boosted regression trees (BRTs) are a self-learning method based on the classification and regression tree algorithm (CART) in the ensemble learning method, and the model reduces the risk of overfitting. Megan et al. (2019), in order to understand the temporal and spatial fishing trends of vessels, considered marine environmental factors and used the boosted regression tree model to analyse the internal relationship between vessel fishing behaviour and environmental factors and to identify the fishing behaviour of vessels and their dependence on the environment.

With the gradual increase in the number of training decision trees, the fitting of the boosted regression tree model to the training data will become increasingly better, but if the number of decision trees exceeds a certain number, overfitting may occur, resulting in a decrease in the prediction accuracy of the model. In this study, a 10-fold cross-validation method was used to establish a boosted regression tree model, and the number of decision trees with the smallest average estimation deviation was taken as the optimal number. We used all the training data to build a fishery forecast model that contains the optimal number of decision trees. Binary classification is carried out based on the presence or absence of fishing effort in the unit grid (recorded as “1” and nonrecorded as “0”); the predicted fishing effort ranges between 0 and 1, and the predicted fishing effort is calculated based on the average value. Two categories (≥average value is recorded as “1”, <average value is recorded as “0”), verify the accuracy of the model’s prediction.

Elith and Leathwick (2010) explained that among the parameter results returned by the “dismo” and “gbm” packages of model training, cv-AUC in the model prediction accuracy evaluation index is the most suitable parameter statistic for evaluating the quality of the model. Derville et al. (2016) stated that the explained deviance corresponds to the percentage of deviation shown by the fitted model, which to a large extent verified the accuracy of model training.

The results of the model training parameters all verify the high accuracy of model training. By comparing the parameters obtained from the BRT model and GAM model, a more appropriate BRT model was selected for prediction. The BRT model predicts the fishing effort of vessels in 2020, which is more consistent with the actual fishing effort of vessels in the spatial distribution area. The accuracy of the prediction of the binary classification reaches 77%, and the training and prediction results of the model have been verified.

The habitat of tuna is closely related to the marine environment. Huang et al. (2020) selected latitude, longitude, depth of mixed layer, month and sea surface temperature and used the second-order boosted regression tree model to establish a habitat model of a free-swimming school of bigeye tuna in the East Pacific to explore its temporal and spatial distribution characteristics. The result is that latitude, longitude, depth of the mixed layer, month and sea surface temperature are the main factors that affect the success rate of bigeye tuna. Nataniel et al. (2021) showed that the catchability of tuna biomass accumulation is related to factors such as sea temperature, productivity, sea level height, and geostrophic flow in addition to the interaction of temporal and spatial variables. This corresponds to several of the environmental factors used in this article.

The distribution of fishery resources changes in response to changes in the marine environment, as does the tracking of fish stocks by vessels. Therefore, vessel operating space changes also reveal information about the spatial changes in fishery resources. The research in this paper shows that tem100 and lon are the most important factors affecting the fishing effort of tuna purse seiners. Second, sst, lat and DSH have the greatest impact on tuna purse seine fishing operations. The importance of tem100, sst, latitude and longitude and DSH totalled 47.56%, and five variables contributed nearly half of the impact.

These important results indicate that temperature, spatial factors and the distance from the vessel to the coastline are the key environmental variables that affect the operation of tuna purse seiners in the western and central Pacific. Although tuna has a thermal regulation function, temperature still has an important influence on its distribution. Studies show that the active water layer where the tuna seine are captured is in the shallow area of 200 m underwater, and the daytime swimming depth mode is 100 m (Takayuki et al., 2016), so the shallow water temperature of the 100 m water layer has an important influence on the fish. This may be the reason why the water temperature of 100 metres and the temperature of the surface water have an important influence on the fishing operations of the purse seine.

Longitude and latitude have a high contribution rate, indicating that tuna purse seiner operations are also spatially selective, and fishing vessels tend to operate in certain areas. Zhang et al. (2021) found that the fishing behaviour of tuna purse seiners in the western and central Pacific Oceans showed a clustering pattern and clustered tightly. This spatial clustering of fishing vessels was related to the fishing objects. Yang et al. (2014) found that bonito resources in the western and central Pacific showed a strong aggregation and distribution, with strong local spatial autocorrelation.

The results of labelling tuna with DTs and ATs in the subtropical Pacific Ocean and equatorial regions showed that tuna is active in related areas (Kurt and Fuller, 2013). Most of the time, bonito is within 25 kilometres of favourable foraging habitat (Druon et al., 2017).

The distance from the purse seiner to the coastline also has a greater impact on the choice of fishing operations. The partial distribution map shows that most fishing vessels operate 100-300 km from the coast, indicating that purse seiners mostly operate within economic exclusives. Therefore, the management of purse seiners is of great significance to individual Pacific Island countries.

The primary productivity of the sea surface and the oxygen content at a depth of 200 metres also contribute more than 5%. Primary productivity is an important part of the ecosystem and has an important impact on tuna spawning habitat, survival rate, and supplementary biomass. Tuna has a dissolved oxygen demand. The purse seine catches small tuna fish, which have poor ability to withstand hypoxia. Salinity and sea surface height also have a certain degree of influence on tuna purse seine fishing operations, while ocean currents, chlorophyll, and the depth of the mixed layer have little effect, and temperature gradients have almost no effect on tuna purse seine fishing operations.

The partial distribution diagram of the influence of 100-metre water temperature and sea surface temperature on vessel operations shows that tuna purse seiners appear more in areas where the 100-metre water temperature is 25-29 °C and the sea surface temperature is 29-31 °C. The suitable area for sea surface temperature and 100-metre water temperature is consistent with the results obtained in the literature (Tang et al., 2014). The curve of the latitude and longitude distribution map shows that the tuna purse seiners are more likely to operate in the 5°S-5°N zonal area; in the longitude direction, the vessels are more likely to appear west of 180°E. This result is similar to the statistical results of Tang et al. (2014) and Kurt and Fuller (2013) based on purse seine data.

Changes in the marine environment have a greater impact on the distribution of tuna species and the spatial behaviour of fishing vessels. Tuna purse seine fishing aims to find and track fish stocks. The impact of the above environmental factors on the fishing operations of tuna purse fishing vessels can also reveal its impact on tuna purse seine fishery resources.