Predicting Bycatch in Space: Comparison of Different Approaches

Nearly all marine fisheries have at least some bycatch, which concerns commercial and recreational fishermen, resource managers, conservationists, and the public. High bycatch rates reduce the efficiency and sustainability of fisheries, but even extremely low bycatch rates can be a problem for protected or rebuilding species. Spatial fishing practices affect bycatch rates, and understanding spatiotemporal patterns in bycatch offers a possibility to improve fisheries management. We demonstrate the ability of powerful new tools, integrated nested Laplace approximations (INLA) and stochastic partial differential equations (SPDE), to spatially model bycatch in two large U.S. fisheries observer datasets (West Coast Groundfish and Hawaii Longline Observer Programs). We compare the INLA-SPDE approach with other spatial modeling frameworks (GAMs, MaxEnt, random forests, and boosted regression trees), and show how the models' performance differs across a broad range of bycatch rates, from loggerhead sea turtles (0.4%) to blue sharks (89%) in the Hawaii longline fishery, and yelloweye rockfish (0.4%) to Pacific halibut (28%) in the West Coast groundfish fishery. We use area under the ROC curve (AUC) to determine the best model in each case, and use it to produce spatial bycatch prediction maps to help management reduce bycatch.