Abstract
Incidental take of nontargeted marine fauna, known as bycatch, has negative long-term effects on marine ecosystems. This is especially concerning for top predators, slow-reproducing fauna, spawning individuals, and endangered species. Predicting spatial distributions of bycatch and the variables that influence those patterns can inform mitigation strategies for fisheries management. This study aims to identify bycatch hotspots in the U.S. Gulf of Mexico reef fish longline fishery with three methods of analysis: a random forest model, a geostatistical model, and boosted regression tree model to predict bycatch per unit effort of each species across time and space using environmental and fishery-specific variables. Random Forest (RF) and Gradient Boosted Tree (GBM) models are powerful tools in classification and regression for forecasting based on ensemble nonparametric learning where each differs technically in methodology on how the model’s learner structure analyzes patterns in data. Spatiotemporal Gaussian Random Field generalized linear mixed effects models (GLMMs) model spatiotemporal patterns based on spatial random fields. These three methods were used to capture bycatch hotspots for a commonly caught bycatch red snapper (Lutjanus campechanus), less commonly caught night shark (Carcharhinus signatus), and rarely caught loggerhead sea turtle (Caretta caretta). Fisheries variables such as soak time and fishing depth were important for predictive power. For example, in night sharks, fishing depth was one of the most important variables for prediction. Environmental variables associated with productivity or nutrient richness such as apparent oxygen utilization, nitrates, and phosphates were also relevant for many species While not as flexible, GLMM showed to be the best model for extremely rare species bycatch prediction. For species with more bycatch incidences, GBM was a model that had higher or similar probabilities with GLMM but with the variable selection and flexibility that RF had.