Incidental take of nontargeted marine fauna, known as bycatch, has negative long-term effects on marine ecosystems and populations by accidentally killing top predators, fauna with slow reproductive capacity, spawning individuals, and endangered species. Predicting spatial areas of concentrated bycatch and variables with bycatch can inform mitigation strategies. Random Forest (RF) models are a powerful tool in classification and regression for forecasting problems and its capability in ensemble nonparametric learning. Whereas, bycatch analyses often use generative additive models (GAMs) to model nonlinear relationships, a random forest is more flexible to incorporate a larger range of predictor variables with complex interactions, including correlations among species. This study aims to identify bycatch hotspots in the US Gulf of Mexico reef fish longline fishery by building a random forest to predict bycatch per unit effort of multiple species across time and space using biological, environmental, and anthropogenic variables. Based on analyses of observer data for this fishery, RF models have better predictive power for this dataset compared to other algorithms including Support Vector Machines, Boosted Regression Trees, and Neural Networks. For the species red snapper (Lutjanus campechanus), night shark (Carcharhinus signatus), and loggerhead sea turtle (Caretta caretta), tree based machine learning algorithms resulted in better diagnostics for Random Forest and Boosted Regression Tree algorithms. Results show that species similar in niche and habitat selection are likely to be statistically significant predictors for bycatch. Red grouper (Epinephelus morio) was the 3rd strongest predictor for red snappers, with overall 61% out of bag accuracy. For night sharks, the species that were statistically significant predictors were prey species with the exception of the spiny dogfish (Squalus cubensis). In rarely observed species like loggerhead sea turtles, random forest had an 87% out of bag accuracy although the year was the largest predictor suggesting that bycatch of loggerhead sea turtles was not consistent and was most likely outlier events. This suggests that although machine learning provides stronger computing power, observer datasets still need to accumulate sufficient data in order for algorithms to analyze patterns accurately. The additional components of bycatch forecast modeling such as a multispecies co-catch variable could result in more predictive forecasts for species that may be abundantly caught as bycatch but do not have any stock estimation. By describing underlying patterns in bycatch such as trophic interactions for specific species, fisheries management can use geostatistical machine learning forecasts to inform bycatch mitigation and fishing area selection.
