Predicting bycatch of Chinook salmon in the Pacific hake fishery using spatiotemporal models

Abstract

Fisheries bycatch is a global problem, and the ability to avoid incidental catch of non-target species is important to fishermen, managers, and conservationists. In areas with sufficient data, spatiotemporal models have been used to identify times and locations with high bycatch risk, potentially enabling fishing operations to shift their effort in response to the dynamic ocean landscape. Here, we use 18 years of observer data from the Pacific hake (Merluccius productus) fishery, the largest by tonnage on the US West Coast, to evaluate our ability to predict bycatch of the commercially, culturally, and ecologically important Chinook salmon (Oncorhynchus tshawytscha). Using multiple approaches (regression models, tree-based methods, and model averages), we tested our ability to predict bycatch at weekly and yearly timescales and found that spatiotemporal models can have good predictive ability. Gradient boosting trees (GBTs) and model averages typically had higher performance, while generalized linear models and generalized additive models (without interaction terms) did less well. Using a GBT model to remove 1% of hauls with the highest predicted bycatch reduced the bycatch-to-hake ratio by 20%. Our results indicate that spatiotemporal models may be a useful forecasting tool that can help fishing operations avoid bycatch while minimizing losses from target catches.