Publicly available shark catch data from longlines between 2012 and 2020 were collected from four tRFMOs (Supplementary Material, Table S1 ): Western and Central Pacific Fisheries Commission (WCPFC), Inter-American Tropical Tuna Commission (IATTC), International Commission for the Conservation of Atlantic Tunas (ICCAT) and Indian Ocean Tuna Commission (IOTC). This time frame was chosen to represent a current snapshot of fishing effort and ensure compatibility among tRFMO datasets and other datasets used in our modeling efforts (see section 2.2 Model parameters). We define “sharks” based on the species present in the tRFMO longline datasets, which includes species or species groups belonging to orders Carcharhiniformes, Lamniformes, and Orectolobiformes, though catches for the latter in longlines are rare ( Table 1 ). We use the term shark “species” generally to refer to all species and species groups reported by tRFMOs ( Table 2 ).

Shark catch data were self-reported by fishers for all tRFMOs except WCPFC, where data were reported by onboard observers. The fisher self-reported and observer catch data only contain non-zero values. We therefore assume that if any shark catch is reported in a cell, sharks that were not reported have a catch value of zero (Supplementary Material, Tables S2, S3 ). This assumption is founded on the belief that if fishers or observers are reporting catch, they will report all species or species groups that are caught given regulations to do so (Supplementary Material, Table S3 ), and will not selectively report some species and not others. The spatiotemporal resolution of data varied by tRFMO. Data that were reported at a 1x1 degree resolution or a 5x5 degree resolution were retained. While IATTC, ICCAT, and IOTC data were reported at monthly scales, WCPFC data were reported on an annual scale (Supplementary Material, Table S1 ). To maintain consistency across tRFMOs and predict shark catch risk at a global scale, data reported at an annual scale were used for this analysis; data that were reported at a monthly scale were summed to an annual scale. The reporting units for catch also varied by tRFMO, with shark catch reported in either counts or metric tonnes and effort reported in number of hooks (Supplementary Material, Table S1 ).

Fishing effort is known to have at least some influence on catch. We tested three different fishing effort parameters (Supplementary Material, Table S4 ). Publicly available fishing effort data associated with target catch (tuna and tuna-like species; “target effort”; Supplementary Material, Figure S1 ) and shark catch (“shark effort”; Supplementary Material, Figure S2 ) specifically were collected from WCPFC, IATTC, ICCAT, and IOTC for longline vessels through 2020. All fishing effort data except WCPFC are self-reported by fishers and included vessel flag information; WCPFC fishing effort data are reported by observers and do not include vessel flag information. The procedure used to parse data on a spatiotemporal scale for shark catch was also applied to the fishing effort data. Due to potential spatial or reporting biases in observer and self-reported datasets, respectively, independent effort data were also collected from the Global Fishing Watch (GFW) database (Kroodsma et al., 2018; Supplementary Material, Figure S3 ). GFW compiles and analyzes automatic identification systems (AIS) data, which are transmitted through a network of terrestrial and satellite receivers, and uses observed changes to vessel movement to infer fishing activity, which they report using various fishing effort metrics (Kroodsma et al., 2018). We downloaded GFW fishing effort reported in kilowatt hours (“GFW effort”), a measure of fishing intensity, at 5x5 and 1x1 degree resolutions from 2012 (the first year data are available) to 2020. All data used to parameterize the shark catch risk model were subsequently subset or queried to represent a current snapshot of shark catch and match the temporal range of the GFW fishing effort dataset (2012–2020).

We assembled key environmental parameters known to influence shark occurrence (Supplementary Material, Table S4 ). Sea surface temperature (SST; °C; Supplementary Material, Figure S4 ) and chlorophyll-A (milligram m-3; Supplementary Material, Figure S5 ) data were compiled from the NOAA Coastwatch ERDDAP data server (dataset IDs: erdHadISST and pmlEsaCCI42OceanColorMonthly, respectively) from 2012–2020. Sea surface height (meters; Supplementary Material, Figure S6 ) data were available from 2012–2019 (Zlotnicki et al., 2019) and 2020 values were interpolated using values from the three nearest neighbors. All environmental data were aggregated to 5x5 and 1x1 degree resolutions and the mean and coefficient of variation (CV) per cell were calculated.

Sharks cannot be caught where they are not present. We included a measure of species occurrence using species distribution model (SDM) polygons from IUCN for all shark species reported as longline catch by tRFMOs (IUCN, 2022). The original probability categories in the IUCN dataset include: 1 (extant), 2 (probably extant), 3 (possibly extant), 4 (possibly extinct), 5 (extinct), 6 (presence uncertain; excluded from the dataset, IUCN SSC Red List Technical Working Group, 2021). We converted these to values between 0 (extinct) and 1 (extant) to represent probability of occurrence in a cell (Supplementary Material, Figures S7–S21 ). For shark catch reported to the family or group (e.g., “Species nei”) level, the mean of SDMs for all relevant species was calculated. Data were aggregated to 5x5 and 1x1 degree resolutions.

We assume that shark species with higher monetary value are more likely to be targeted by longline fisheries. Species-specific ex-vessel prices were generated using methods from Melnychuk et al. (2017) for 2012–2020 (Supplementary Material, Tables S5, S6 ). For species with catch data, but no ex-vessel price information, a median ex-vessel price was calculated by year for all shark species collectively. Values for the year 2020 were inferred by multiplying the 2019 ex-vessel price by the estimated inflation rate for the US in 2020 (1.2%).

We built Random Forest (RF) machine learning models to estimate spatially explicit shark catch risk. The RF modeling approach consists of a series of “trees” that are trained independently using a combination of predictor variables and rows of data that are randomly resampled from a training dataset (Breiman, 2001). Combining results from all trees within a “forest” allows for an efficient, accurate, and robust prediction of the response variable with low risk of overfitting (Breiman, 2001). A testing dataset is reserved to evaluate the predictive power of the RF model. RF models are increasingly applied to ecological and fisheries data for their proven ability to provide robust predictions and superior performance in the face of common challenges associated with sparse data collection efforts, low sample sizes, and noisy data (Luan et al., 2020). The RF modeling approach does not require assumptions about data properties, which is a constraining feature of traditional parametric regression, and model performance is not impeded by interactions between predictor variables (Qi, 2012). RF modeling is used for both classification and regression problems, and has proven to be particularly useful in marine research to build species distribution models (Luan et al., 2020), estimate small-scale fishing effort (Behivoke et al., 2021), and estimate bycatch risk via echosounder data (Mannocci et al., 2021). Our motivation for using this RF approach to predict shark catch risk is threefold: (1) it does not bind us to parametric or non-parametric approaches like traditional regression analyses, (2) it allows us to leverage our sample to predict over areas in which no catch was originally reported, which is particularly useful for overcoming spatiotemporal observation biases, and (3) it provides a robust prediction that can be easily refined as more data become available (Breiman, 2001; Siroky, 2009).

Individual RF models were created to predict shark catch risk (catch in counts) for each tRFMO using the tidymodels R package (Kuhn and Wickham, 2020). All reported shark catch was retained in the model, though we note that not all caught sharks were landed (e.g., 12.37% landed, 50.88% discarded, and 36.76% unknown fate in WCPFC from 2012–2020). Shark catch risk is therefore an estimate of the likelihood that a shark will be caught, but not necessarily be killed, by a longline vessel in a spatial cell.

For each tRFMO, we tested various spatial resolutions and shark catch units to determine the most appropriate dataset for future model runs, identified by the highest R² for each tRFMO. Once a resolution and unit were selected for a tRFMO, the same resolution was used in future model runs (Supplementary Material, Table S7 ). The resolutions of data that were tested include:

We employed a two-phase approach to conduct a feature selection procedure via the “wrapper” method described by Kuhn and Johnson (2019). The wrapper method uses different combinations of predictor variables to evaluate model performance (Supplementary Material, Table S7 ). In Phase 1, we tested our three different fishing effort metrics: target effort (number of hooks, by flag when available), shark effort (number of hooks, by flag when available), and GFW effort (kilowatt hours). Other predictors were held constant, including mean SST, mean chlorophyll-A, and species suitability. We used R² to determine which dataset and effort metric best predicted shark catch risk. In Phase 2, we used the best performing dataset and metric found in Phase 1 to conduct a second wrapper-based feature selection procedure to test the combinations of the following variables: SST (mean or mean and coefficient of variation), chlorophyll-A (mean or mean and coefficient of variation), sea surface height (mean or mean and coefficient of variation), species-specific ex-vessel prices, and group-wide ex-vessel prices.

For each model, we grouped shark catch data into five spatial clusters using spatial dissimilarity matrices and then randomly split catch across combinations of years and spatial clusters into training (75% of all data; n = 19,393 [WCPFC], 254,475 [IATTC], 10,125 [IOTC], 6,444 [ICCAT] rows) and testing datasets (25% of all data; n = 6,464 [WCPFC], 84,825 [IATTC], 3,375 [IOTC], 2,148 [ICCAT] cells) to ensure that model performance was not based purely on spatiotemporal autocorrelation. Shark catch records and reported fishing effort are incomplete due to low observer coverage or partial self-reporting. To accommodate this and not limit our spatial risk estimation to cells with reported shark catch only, we allowed our model to predict shark catch over the entire spatial footprint of longline fishing effort, even if no sharks were reported in a fished cell. For each year, we gathered all cells in which shark catch or fishing effort by longlines was reported. For cells in which effort was reported, but shark catch was not, we recorded catch as 0 for each shark species and species group in our dataset. Similarly, for cells in which shark catch for at least one species was reported, we recorded catch as 0 for all other shark species and species group in our dataset (Supplementary Material, Tables S2, S3 ).

Due to the large number of 0 catch values within the dataset, we used a hurdle model approach (Mullahy, 1986). First, shark catch within the training dataset was classified as either “present” (catch > 0) or “absent” (catch = 0). To ensure a balanced training dataset, the minority classification (shark catch present) was randomly upsampled until there was an equal number of both classification groups (using the themis R package; Hvitfeldt, 2022). Numerical predictor variables were then centered and scaled and missing numerical data were imputed using data from three nearest neighbors. Missing categorical values were assigned as “unknown” factors and dummy variables were created for each categorical variable. Finally, all predictors that contained only a single value were removed (e.g., if a particular flag only reported effort once throughout the time series, the column for that flag was removed from the model).

The resulting dataset was used to train an RF classification model (Luan et al., 2020) to predict either shark catch presence or absence. We then trained an RF regression model on the “presence” data to predict shark catch in each cell. Due to potential reporting biases in observer or self-reported data, model results are not intended to be viewed as actual catch estimates, but rather hotspots of potential catch risk. The final prediction of shark catch risk was the product of the outcome from the classification model and the regression model. Models were run with 500 trees.

The general formula for each of the models was:

Component 1: Random Forest Classification Model

Component 2: Random Forest Regression Model

Final Prediction:

To avoid overfitting, we generated a new dataset in which fishing effort and environmental variables in each cell were assigned a random value from a normal distribution using the mean and standard deviation of the predictor in the cell across all years (2012–2020). We then used the same hurdle model procedure to identify spatial hotspots of shark interactions with longline fisheries globally.

All code and data required to reproduce this analysis are available on Zenodo (Burns et al., 2022a) and Dryad (Burns et al., 2022b), respectively.

Collation and harmonization of publicly available tRFMO data provide a geospatial estimate of areas with high reported shark catch (count), effort associated with shark catch (hooks), and shark CPUE (count/hook, Figure 1 ). Due to differences in reporting requirements across tRFMOs and the general scale of data reported, values were scaled on a tRFMO basis for visualization. Areas in each tRFMO with the highest reported catch, effort, or CPUE for sharks often do not overlap with each other (Supplementary Material, Table S8 ). WCPFC has the highest degree of overlap among high value cells, with the overlap of high shark catch and shark CPUE (21.1% of cells), and the overlap of high shark catch and effort associated with shark catch (20.7% of cells). The remaining 58.2% of high value cells do not overlap with another metric. Comparatively, IATTC has the lowest degree of overlap, with 87.3% of high value cells having no overlap with other metrics. The low degree of overlap among shark catch, effort associated with shark catch, and shark CPUE indicates the need for a modeling framework that does not assume a simple, linear relationship between these metrics.

Areas with the highest shark CPUE (count/hook) are typically associated with catches of blue shark (Prionace glauca, 78.23% of total reported tRFMO shark catch globally; Supplementary Material, Table S9 ) or shortfin mako shark (Isurus oxyrinchus, 4.99% of total reported tRFMO shark catch globally; Supplementary Material, Table S9 ), but species-level CPUE varied spatially. In WCPFC, areas with high CPUE are offshore of western New Zealand, dominated by blue sharks, porbeagle sharks (Lamna nasus), and shortfin mako sharks. Comparatively, areas of high CPUE in IATTC correspond to silky sharks (Carcharhinus falciformis) along the coast of Mexico and blue and shortfin mako sharks off Chile. Areas of high CPUE in ICCAT extend from coastal Brazil to Argentina for blue, shortfin mako, and porbeagle sharks. Areas of high CPUE in IOTC are offshore of South Africa and southwestern Australia for blue sharks. The highest diversity of shark species is reported by the WCPFC (n = 12 true species [i.e., not “sharks nei” or other grouped categories]; compared to n = 4, 4, 3, species in IATTC, IOTC, and ICCAT, respectively), which is more likely a reflection of differences in reporting requirements as opposed to differences in true catch rates of different species.

Areas of high tuna catch in each tRFMO were identified for the three most frequently reported tuna species: bigeye tuna (Thunnus obesus), albacore tuna (T. alalunga), and yellowfin tuna (T. albacares, Figure 2 ). Bigeye and yellowfin tuna catch is highest in the central oceans whereas albacore tuna catch is distributed across the southern oceans, consistent with each species' habitat preferences (Arrizabalaga et al., 2015). Within areas of high tuna catch, mean CPUE for shark species is consistently low across tRFMOs and years. The summed mean CPUE for all shark species does not exceed 10 individuals per 10,000 hooks and summed mean CPUE for threatened shark species' (IUCN categories of Vulnerable, Endangered, or Critically Endangered) does not exceed 0.05 individuals per 10,000 hooks.

The best performing RF models vary slightly among tRFMOs and result in a more comprehensive picture of shark catch risk than tRFMO catch reports alone (Supplementary Material, Table S10 , Figures S22-S25 ). Sea surface temperature, chlorophyll-A, species suitability, and species identity serve as important predictors across all tRFMOs ( Table 3 ; Supplementary Material, Figures S26-S29 ). Sea surface height improved model performance for IATTC only, and ex-vessel prices improved model performance for IOTC and WCPFC. Remotely sensed GFW fishing effort, which was tested in lieu of effort reported with target catch or shark catch, did not improve model performance and, therefore, was not used in any of the final model runs. For all shark species, areas of high catch risk are primarily located within the ICCAT jurisdiction, off the coast of Southwestern Africa and in the Central Atlantic Ocean ( Figure 3A ; Supplementary Material, Figures S30, S31 ). Areas of high catch risk for threatened species are concentrated offshore of Central America in the Eastern Tropical Pacific, partially abutting the Costa Rica Thermal Dome ( Figure 3B ). Overall, areas of high catch risk for sharks appear to correspond to areas of high risk for the three most prominently reported shark species: blue (Near Threatened; Supplementary Material, Figure S32 ), shortfin mako (Endangered; Supplementary Material, Figure S33 ), and silky sharks (Vulnerable; Supplementary Material, Figure S34 ). Shark catch risk areas were consistent across various “risk” thresholds examined, with higher upper bounds and smaller lower bounds serving to spatially constrain identified hotspots but not affect their location (Supplementary Material, Figure S31 ). Catch risk for all sharks reported at the species level are shown in the Supplementary Material Figures S32–S46 .

In addition to blue, shortfin mako, and silky sharks being the most prominently reported shark species, they are also most evenly represented across three or more tRFMOs ( Figure 4 ). We calculated the number of high catch risk cells within each tRFMO independently. The relative percentage of high-risk cells within a particular tRFMO for blue sharks is similar across tRFMOs, while a higher percentage of cells are high-risk for both shortfin mako and silky sharks within WCPFC boundaries. The majority of other species are represented across fewer tRFMOs (e.g., longfin mako shark [Isurus paucus], oceanic whitetip shark [C. longimanus]) or are predicted in small quantities, rarely exceeding the threshold required to be considered high risk (e.g., porbeagle sharks, common thresher sharks [Alopias vulpinus]).

We also investigated the overlap of areas of high catch risk for shark species across different IUCN Red List categories ( Figure 5 ). Several species categorized as Critically Endangered (1 of 5 species reported), Endangered (2 of 3 species reported) and Vulnerable (2 of 5 species reported) were found to have high-risk cells offshore of the Hawaiian Islands ( Figure 6 ). Although great hammerhead (Sphyrna mokarran; Critically Endangered), whale (Rhincodon typus; Endangered), and blacktip sharks (C. limbatus; Vulnerable) appear in catch reports, interactions with these species in longline tRFMOs were reported so rarely that they resulted in a global predicted catch risk of 0. In addition, areas of high-risk for both blue and shortfin mako sharks show moderate overlap. Blue, shortfin mako, and silky sharks are the only species that exhibit high catch risk areas covering large expanses of oceans across at least two tRFMOs, which could be a reflection of relatively high global abundances, a high degree of association with target stocks, and/or greater consistency in reporting standards across tRFMOs for these species.