Forecasting fish recruitment using machine learning methods: A case study of arabesque greenling

Abstract

Fish recruitment prediction is one of the most challenging topics in fisheries science. The recruitment of arabesque greenling in northern Hokkaido, Japan, which has been annually assessed for population size, greatly fluctuates. Whether the cause of fluctuation is environment or overfishing is controversial. We use a machine learning method for predicting the recruitment of arabesque greenling. Biological, fisheries-related, and environmental factors were included in the predictive models as feature variables. A gradient boosting model (GBM) showed better predictive performance compared with a simple hockey-stick stock-recruitment curve (HS), linear regression model (LRM), generalized additive model (GAM), and a random forest model (RFM) in terms of relative bias and relative root mean square error for recruitment prediction in the last 5 years. The most influential feature for GBM was spawning stock biomass in the last year, followed by the fishing rate for older fish and recruitment at the last year. The sea temperatures (STs) at the depth of 0, 50, 100, and 200 m were unimportant predictors in GBM. The difference in important predictors among models suggests the importance of nonlinearity and incorporating multiple variables simultaneously. This study highlights the potential usefulness of GBM for fish recruitment forecast and thereby sustainable fisheries management.