Assessing the potential responses of ten important fisheries species to a changing climate with machine learning and observational data across the province of Québec

Abstract

Models are needed to predict changes in game fish abundances with respect to climatic factors undergoing change, but such models are often limited by data availability and the capacity of statistical methods to fit challenging ecological datasets. We use current methods in machine learning to describe the responses of ten fish species to climatic factors across Québec. We assembled a new province-wide, synthetic dataset of fish catches spanning almost 50 years and 6000 sites. Extreme Gradient Boosting (XGBoost) models revealed that climatic factors are more important predictors of trends in game fish catches than nuisance factors (sampling gear, time), lending support to collating other heterogeneous datasets for analyses. Mean annual temperature and precipitation were the most important drivers of species catches. Fish thermal preference guilds predicted primarily species responses to temperature, suggesting that warmer and wetter climates may not favour the same species. Despite the challenging nature of these datasets, XGBoost models provided excellent fit, predictive capacity, and interpretability, thereby illustrating that large, heterogeneous datasets can be used to inform freshwater fisheries management in a changing climate.