Southward re‐distribution of tropical tuna fisheries activity can be explained by technological and management change

Abstract There is broad evidence of climate change causing shifts in fish distribution worldwide, but less is known about the response of fisheries to these changes. Responses to climate‐driven shifts in a fishery may be constrained by existing management or institutional arrangements and technological settings. In order to understand how fisheries are responding to ocean warming, we investigate purse seine fleets targeting tropical tunas in the east Atlantic Ocean using effort and sea surface temperature anomaly (SSTA) data from 1991 to 2017. An analysis of the spatial change in effort using a centre of gravity approach and empirical orthogonal functions is used to assess the spatiotemporal changes in effort anomalies and investigate links to SSTA. Both analyses indicate that effort shifts southward from the equator, while no clear pattern is seen northward from the equator. Random forest models show that while technology and institutional settings better explain total effort, SSTA is playing a role when explaining the spatiotemporal changes of effort, together with management and international agreements. These results show the potential of management to minimize the impacts of climate change on fisheries activity. Our results provide guidance for improved understanding about how climate, management and governance interact in tropical tuna fisheries, with methods that are replicable and transferable. Future actions should take into account all these elements in order to plan successful adaptation.

additional burden on fisheries management by affecting abundance, phenology and causing distribution shifts of fish stocks across jurisdictional boundaries (Free et al., 2019;Poloczanska et al., 2016;Young et al., 2019). These impacts have a range of implications for fisheries management . As transboundary fisheries become more common with climate change, fisheries can face access problems due to the relatively slow speed of negotiations and changes to management agreements (McIlgorm et al., 2010).
Existing projections of marine biodiversity, global catch potential and fisheries revenues under climate change predict significant impacts, with biodiversity of marine species decreasing in the tropics (García Molinos et al., 2015), a re-distribution of the catch potential globally (Cheung et al., 2010) and increases in revenues in northern latitudes as lower latitudes face reduction in profits (Lam, Cheung, Reygondeau, & Sumaila, 2016). These impacts are expected to affect heavily the people depending on marine resources for their livelihoods (e.g. Barange et al., 2018;Bell et al., 2013;Young et al., 2019). These species are expected to be affected by anthropogenic climate change (Muhling et al., 2015), shifting their distribution, migration times, physiological rates and abundance with consequences for catchability to fisheries (Báez, Pascual-Alayón, Ramos, & Abascal, 2018;Brill & Hobday, 2017). Due to the migratory nature of these species, some tuna fisheries operate over large spatial scales and multiple jurisdictions, utilizing fishing agreements to access EEZs, and with significant fishing activity in the high seas (Mullon et al., 2017). The importance of these fisheries for economies and livelihoods, together with the transboundary nature of the stocks and management, and the threat of climate change to tropical tuna species make it crucial to understand the challenges regarding future sustainability. Moreover, the interactions between management institutions, international agreements and climate-driven changes in the tropical tuna fisheries need to be understood to plan effective adaptation to climate change.
As with other food production sectors, integrating biophysical information together with social and economic factors is important when developing management response options to climate change.
Here, we investigate evidence of climate change affecting the recent distribution of tropical tuna fisheries using time series analyses of effort data. Then, we explore the role of technological and institutional actions associated with the distribution shifts, including the existing international agreements and management regulations.
We focus on the east Atlantic Ocean, where a number of fleets operate. Our goal is to improve the understanding about how climate, management and governance interact in tropical tuna fisheries, with methods that are replicable and transferable.

| MATERIAL S AND ME THODS
Effort data from the ICCAT and sea surface temperature anomaly (SSTA) data have been used to evaluate whether climate change is impacting the effort distribution of purse seine (PS) fisheries operating in the east Atlantic Ocean. An analysis of the spatial change in effort distribution is conducted using a centre of gravity (COG) approach. Empirical orthogonal functions (EOFs) are used to assess the spatiotemporal changes in effort anomaly (effortA) and SSTA.
These functions generate spatial patterns and time series of effort and temperature that are correlated. Finally, random forest models are used to explore the influence of management, institutional agreements and technological changes on the effort. All analyses are performed using R software (version 3.5.1; R Core Team, 2018). R scripts for replication are available in GitHub/irrubio/tropituna_fish-ery_change (see the workflow of scripts in Figure S1).

| Tropical tuna fisheries data
Monthly effort (fishing hours) and catch data of all PS fisheries targeting tropical tunas from 1991 to 2017 in the Atlantic Ocean were downloaded from the ICCAT website (ICCAT, 2018). This PS database is at a 1 by 1 degree resolution. For our analysis, PS data were limited to the eastern Atlantic Ocean (29°N, 30°S, 19°E, 35°W).
Effort data were aggregated by summing to 5 by 5 degrees and also by quarter to keep the representativeness of the data in the study area for the EOF analysis. Then, effortA was calculated by subtracting the quarterly mean over the entire data period from the data (Benestad, Hanssen-Bauer, & Chen, 2008;Bjornsson & Venegas, 1997). Quarterly effort was also calculated at 1 by 1 resolution by summing for the COG analysis.
In this study, we assume that the location of the reported fishing effort represents the distribution of the fleet activity (Davies, Mees, & Milner-Gulland, 2014). Our main objective is to understand any fleet activity distribution response to climate change. Previous studies have considered catch and catch per unit effort (CPUE) spatial changes, using these variables as proxies for abundance. These studies are, however, not free from caveats, especially for PS fisheries (see in Kaplan et al., 2014 andTidd, Brouwer, &Pilling, 2017). In a preliminary analysis, we found evidence of distributional changes in the fishery based on catch and CPUE data (calculated as the weight [t] caught by operation mode per fishing hour), but PS CPUE is not considered a good proxy for abundance (Kaplan et al., 2014;Maunder et al., 2006). To test for any relationship between catch/ CPUE and abundance, we correlated estimated yearly biomass by the ICCAT Standing Committee on Research and Statistics (SCRS) with yearly catch/CPUE on fishing aggregation devices (FADs) and free schools of YFT and BET (SKJ biomass data are nowadays unavailable from the SCRS). Only YFT catch data on free schools were significantly correlated with abundance, as catches or CPUE on FADs did not follow (and were decoupled from) biomass trends (Figures S2 and S3); and BET correlations between catch/CPUE and estimated biomass were non-significant. However, we were unable to use YFT catch location when fishing on free schools as a proxy for YFT distribution since the data were too patchy to allow the EOF analysis. For all these reasons, we conducted the final analysis with a focus on the distribution of fishing effort.

| Sea surface temperature anomaly data
To describe environmental change, we used monthly Kaplan SST V2 anomaly data from 1856 to 2017 with a spatial resolution of 5 by 5 degrees (NOAA/OAR/ESRL.PSD, 2018), which were aggregated by quarter by averaging for the study period and area. These SST anomalies are based on the 1951-1980 time period. We used 5 by 5 degree resolution to match effort resolution for the EOF analysis.
While this coarse resolution does not permit analysis of environmental features (e.g. eddies or fronts), we are seeking to understand large scale shifts linked to highly mobile fleets targeting highly migratory species. We used the data from 1856 to 2017 to show the long-term change in SSTA in the study region ( Figure 1); however, for the rest of the analysis, we matched SSTA and effort data temporal scales, that is 1991-2017.

| Institutional and technological data
In order to investigate the role of technological and institutional actions in the effortA distribution shifts, we collected information on PS institutional arrangements, management regulations and technological change in the study region over the period 1991-2017.

The European Union has public Sustainable Fisheries Partnership
Agreements (SFPAs) with African countries that allow European vessels to enter EEZs of those countries. These agreements could have had an impact on the fishery distribution since European PS represents 56% of the PS tropical tuna catch in the study area (calculated from the database used in 2.1). We focused on two variables, the number of vessels allowed to enter a specific EEZ by SFPA and Unfortunately, information on private agreements is not available publicly and we were not able to include it in the analysis.
The ICCAT is the management body that establishes fishing regulations, including Total Allowable Catches (TACs) that apply in the region. We compiled information on the different management and conservation policies taken by the ICCAT to preserve tuna fisher-

| Distribution change analysis
To capture the spatial patterns of the PS tropical tuna fishery and to investigate year-to-year variations, the latitudinal COG of the effort is calculated every year. The COG represents the mean location of the effort (Saraux et al., 2014). We consider annual COG north and south of the equator separately to check for poleward expansion in each hemisphere. The COG is calculated using Equation (1), where n is the number of fishing sets, effort i is the effort in the ith set and latitude i is the latitude of the ith set. A linear regression is used to evaluate trends over time in COG in each hemisphere, and COG is also correlated (Pearson) to SSTA.

| Temperature and effort distribution analysis
To further assess the variability of the tropical tuna fishery distribution over time and its relationship with SSTA, we apply empirical orthogonal functions (EOFs) to the quarterly PS effort anomaly data (effortA) (Bjornsson & Venegas, 1997;Preisendorfer, 1988) and the quarterly SSTA. We use the "sinkr" package in R (version 0.6; Taylor, 2017). The EOFs are found by computing the eigenvalues and eigenvectors of the effortA or SSTA covariance matrix. The derived eigenvalues provide a measure of the per cent variance explained by each mode of variability. Then, the most informative EOFs are identified, and temporal and spatial structures investigated. However, the variance of effortA or SSTA is not equal at all gridpoints; therefore, the "local" explained variance (LC) is also calculated using Equation (2).
which is equivalent to the fraction of variance expressed by the ith EOF at each jth grid point over the total variability reconstructed by the leading k max EOF (Ganzedo, Alvera-Azcárate, Esnaola, Ezcurra, & Sáenz, 2011). λ i represents the eigenvalue of the covariance matrix associated with the ith eigenvector e i . This analysis does not allow for missing data, thus, effortA is reconstructed by means of the Data Interpolating Empirical Orthogonal Functions method (DINEOF) (Beckers, Barth, & Alvera-Azcárate, 2006), which has already been applied to the ICCAT dataset and validated by Ganzedo, Erdaide, Trujillo-Santana, Alvera-Azcárate, and Castro (2013). Spatiotemporal time series with <25% of missing data were selected for reconstruction (Ganzedo et al., 2013). In total, the dataset for EOF analysis has 26 spatial series each with 108 time values (quarters from 1991 to 2017).
The advantage of performing an EOF analysis is that a small number of leading EOF can explain a large fraction of the total variance of the whole dataset, as other studies applied to fish ecology and fisheries science have found (Marshall et al., 2016;Petitgas, Doray, Huret, Masse, & Woillez, 2014;Saraux et al., 2014). Then, posterior testing can be done to see whether a relationship exists between effortA and SSTA by performing Pearson correlations between the EOF temporal structures of effortA versus SSTA. This methodology takes into account the entire spatiotemporal structure of the datasets (including latitude and longitude) which is not possible with the previous COG analysis.

| Institutional and technological analysis
Two random forest models are used to evaluate the influence of management, international agreements and technology on the effort and the temporal structure resulting from the EOF analysis of the effortA compared with other variables (e.g. SSTA) using the "randomForest" package in R (version 4.6-14; Liaw & Wiener, 2002).
This method has been previously applied in fisheries research (e.g. Melnychuk, Banobi, & Hilborn, 2013;Pons et al., 2017) and allows for non-linear relationships between predictors and a response variable without making any parametric assumptions about the distribution of the response variable. Random forests (Breiman, 2001) are an ensemble method, which build a selected number of regression trees (m) from a boostrap sample of the original data set. Kuhn and Johnson (2013) suggest using at least 1,000 trees that in our case are adequate to stabilize the mean squared error (MSE) of the model.
For each regression tree, a set of predictors (mtry) are randomly selected from the original predictors at a given node. We use the default value of mtry, which is equal to a third of the predictor variables (Liaw & Wiener, 2002). The best predictor is then determined, and the data split in two groups such that the overall sum of squared errors is minimized (Kuhn & Johnson, 2013). This process continues until a tree is built, and multiple predictor variables can be shown to influence the response variable. We present from the analysis a variable importance plot for visualizing the percentage of increase in the MSE (%IncMSE) when variables are randomly permuted, which gives a measure of how influential each considered predictor variable is on predicting the response variable. We also include the results of partial dependence plots that provide insight into the directionality of the effect for a given predictor (Berk, 2008). Before conducting random forest, predictors were tested for collinearity using generalized variance inflation factors (Fox & Monette, 1992), which were <6.0 suggesting little possibility of confounding among the predictor variables (Zuur, Ieno, & Elphick, 2010).
In our two random forest models, the response variables are the quarterly effort in fishing hours and the quarterly temporal struc-

| RE SULTS
Records of SSTA in the study region show a statistically significant in- and correlates with SSTA with a Pearson r coefficient of −0.3 (p < .05).
When the SSTA is higher, the COG shifts south. The northern COG of effort does not follow any clear distribution trend over the study period and area ( Figure 2) and does not correlate with SSTA.
In order to further explore changes in distribution and any relationship with temperature, we performed an EOF analysis on SSTA and effortA and then examined the correlation between the resulting temporal structures (Figure 3; see uncorrelated structures of second and third leading EOF in Figure S4). The first leading EOF of the SSTA accounts for 70% of the data variance and the first leading EOF of the effortA for 19% (Figure 3). Correlation between the EOF temporal structure of SSTA and effortA is significant (p < .05), with a Pearson r coefficient of 0.5.     (Deary et al., 2015;Dell et al., 2015;Erauskin-Extramiana et al., 2019;Evans et al., 2015). Here, we did not take into account changes in the depth of fishing effort, but PS fishing may be more affected than other methods such as longline fisheries that can set hooks at greater depths (Marsac, 2017).
Regarding latitudinal shifts, PS vessels in the east Atlantic Ocean are already going further south and north from the equator due to technological advances among other factors, but this also leads to increased fuel costs (see also Michael et al., 2017). Thus, home port location could become limiting if species shift further away.
Other implications of shifting species concern the distribution of tropical tunas in EEZs of the coastal countries (Báez et al., 2018).
Some countries in the Guinean Gulf might be losing resources and others gaining, which adds a challenge for the sustainability of the resources (Bell et al., 2013;Dubik et al., 2019;Pinsky et al., 2018).
New public or private agreements and conflicts (e.g. Spijkers & Boonstra, 2017) might also arise between countries from tropical tunas' re-distribution.
The IPCC (2019)  Finally, identifying impacts of climate change on tropical tunas and their associated fisheries and management is of great importance to determine whether it will affect societies dependent on these resources. The method developed here to explore the effect of ocean temperature on the effort of tropical tunas in the east Atlantic Ocean over the period 1991-2017 showed that PS fisheries have shifted southward from the equator, which can be explained by climate change, institutional, management and technological factors. However, our results suggest that management can be a powerful tool to overcome climate change distribution impacts on fisheries activity. As fisheries management organizations have a crucial role to maintain resource sustainability, adaptation to climate change needs to be incorporated in their agendas, which must span environmental, institutional and socioeconomic considerations.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are openly available in "tropituna_fishery_change" at http://githu b/irrub io/tropi tuna_fishe ry_change and https ://doi.org/10.5281/zenodo.3574095.