Regression tree and ratio estimates of marine mammal, sea turtle, and seabird bycatch in the California drift gillnet fishery, 1990-2015

Marine mammal, sea turtle, and seabird bycatch was estimated for the California swordfish drift gillnet fishery during a 26-year period (1990-2015), using random forest regression trees. Tree estimates were compared with traditional annual ratio estimates generated from the same observer data. Ratio estimates suffer from systematic bias (under- and overestimation of bycatch) when observed bycatch is rare, bycatch rates are inferred only from same-year data, and observer coverage is low. Model-based approaches result in more stable annual bycatch estimates with better precision, because estimates are informed by all available data. Even in years with zero observed bycatch, expected values from regression trees are usually positive (sometimes fractions of animals) and have error estimates (thus acknowledging the possibility that animals may be caught even when none are observed), whereas corresponding ratio estimates would be zero and have no error estimates. Regression tree bycatch models include a suite of oceanographic, location, and gear variables used as predictors to estimate bycatch at the fishing-set level. Variables that significantly influenced bycatch rates were identified with a variable selection approach using the R-package rfPermute and validated with a simulated bycatch dataset. Results indicate that rfPermute can succeed in identifying significant predictor variables for rare bycatch events, even when these events represent <1% of all data. [doi:10.7289/V5/TM-SWFSC-568(http://dx.doi.org/10.7289/V5/TM-SWFSC-568)] James V. Carretta, Jeffrey E. Moore, Karin A. Forney. "January 2017." System requirements: Adobe Acrobat Reader. Includes bibliographical references.