We find shark catch risk hotspots in all ocean basins, with notable high-risk areas off Southwest Africa and in the Eastern Tropical Pacific. These patterns are mostly driven by more common species such as blue sharks, though risk areas for less common, Endangered and Critically Endangered species are also identified. Clear spatial patterns of shark fishing risk identified here can be leveraged to develop spatial management strategies for threatened populations. Sharks are susceptible to industrial longline fishing due to their slow life histories and association with targeted tuna stocks. Identifying fished areas with high shark interaction risk is vital to protect threatened species. We harmonize shark catch records from global tuna Regional Fisheries Management Organizations (tRFMOs) from 2012–2020 and use machine learning to identify where sharks are most threatened by longline fishing. Most spatial patterns are driven by more common species such as blue sharks, though risk areas for less common, endangered and critically endangered species are also identified.

We built Random Forest (RF) machine learning models to estimate spatially explicit shark catch risk globally by longlines using a suite of catch and effort data from tRFMOs, additional effort datasets for fishing effort (Global Fishing Watch), environmental datasets (sea surface temperature, sea surface height, chlorophyll-A) and economic datasets (ex-vessel price). More information on the exact datasets used can be found in the associated software works. 

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.

We then conducted a second phase of parameter testing for combinations of the following variables: sea surface temperature (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.

The general formula for each of the models was: 

Component 1: Random Forest Classification Model

• presence or absence ~ species distribution model + species common name + mean SST + mean chl-a + effort (by flag if available) + (any combination Phase 2 predictors)

Component 2: Random Forest Regression Model

• catch ~ species distribution model + species common name + mean SST + mean chl-a + effort (by flag if available) + (any combination Phase 2 predictors)

Final Prediction: 

• Component 1 Result * Component 2 Result

Please refer to the associated software works for instructions on how to download the input dataset and set up your folder structure. 

The files saved here are the outputs for machine learning models that were run using publicly available tRFMO datasets. Please refer to the README files for metadata.

• <res>_count_all_rfmos_ll_effort_results.csv: where <res> refers to a 1x1 or 5x5 degree spatial scale; results from the first round of machine learning models to determine which effort (Global Fishing Watch or tRFMO reported effort) metric performed best at predicting shark catch (count only) for that resolution 
• <res>_mt_to_count_all_rfmos_ll_effort_results.csv: where <res> refers to a 1x1 or 5x5 degree spatial scale; results from the first round of machine learning models to determine which effort (Global Fishing Watch or tRFMO reported effort) metric performed best at predicting shark catch (count and metric tonnes converted to count) for that resolution 
• <rfmo>_ll_models_other_results.csv where <rfmo> refers to each tRFMO for which we run the models; results from the second round of machine learning models to determine which additional parameters (e.g., sea surface height, ex-vessel prices) result in the best performing model
• <rfmo>_ll_untuned_final_predict.csv: where <rfmo> refers to each tRFMO for which we run the models; the final, global prediction from the trained machine learning model for that RFMO
• <rfmo>_ll_untuned_model.rds: where <rfmo> refers to each tRFMO for which we run the models; the final fitted classification and regression model, the model prediction on the test dataset, the metrics for the prediction on the test dataset, the final global prediction on the novel dataset, the metrics for the final global prediction on the novel dataset, and the feature importances for the model