Predicting walleye recruitment as a tool for prioritizing management actions

Abstract

We classified walleye (Sander vitreus) recruitment with 81% accuracy (recruitment success and failure predicted correctly in 84% and 78% of lake-years, respectively) using a random forest model. Models were constructed using 2779 surveys collected from 541 Wisconsin lakes between 1989 and 2013 and predictor variables related to lake morphometry, thermal habitat, land use, and fishing pressure. We selected predictors to minimize collinearity while maximizing classification accuracy and data availability. The final model classified recruitment success based on lake surface area, water temperature degree-days, shoreline development factor, and conductivity. On average, recruitment was most likely in lakes larger than 225 ha. Low degree-days also increased the probability of successful recruitment, but primarily in lakes smaller than 150 ha. We forecasted the probability of walleye recruitment in 343 lakes considered for walleye stocking; lakes with high probability of natural reproduction but recent history of recruitment failure were prioritized for restoration stocking. Our results highlight the utility of models designed to predict recruitment for guiding management decisions, provided models are validated appropriately.