Environmentally Driven Seasonal Forecasts of Pacific Hake Distribution

Abstract

Changing ecosystem conditions present a challenge for the monitoring and management of living marine resources, where decisions often require lead-times of weeks to months. Consistent improvement in the skill of regional ocean models to predict physical ocean states at seasonal time scales provides opportunities to forecast biological responses to changing ecosystem conditions that impact fishery management practices. In this study, we used 8-month lead-time predictions of temperature at 250 m depth from the J-SCOPE regional ocean model, along with stationary habitat conditions (e.g., distance to shelf break), to forecast Pacific hake (Merluccius productus) distribution in the northern California Current Ecosystem (CCE). Using retrospective skill assessments, we found strong agreement between hake distribution forecasts and historical observations. The top performing models [based on out-of-sample skill assessments using the area-under-the-curve (AUC) skill metric] were a generalized additive model (GAM) that included shelf-break distance (i.e., distance to the 200 m isobath) (AUC = 0.813) and a boosted regression tree (BRT) that included temperature at 250 m depth and shelf-break distance (AUC = 0.830). An ensemble forecast of the top performing GAM and BRT models only improved out-of-sample forecast skill slightly (AUC = 0.838) due to strongly correlated forecast errors between models (r = 0.88). Collectively, our results demonstrate that seasonal lead-time ocean predictions have predictive skill for important ecological processes in the northern CCE and can be used to provide early detection of impending distribution shifts of ecologically and economically important marine species.