Abstract

The Forests of North America play a critical role in the global carbon cycle and biodiversity but they are threatened by the rapidly changing climate. Assisted migration is one tool we have to mitigate the effects of climate change. However, there are multiple climate variables that are changing, so it is not obvious how far and in which direction to move species or seed sources. Statistical analysis and modeling is needed to translate multivariate climate change into geographic movements. For this purpose, we evaluated three machine learning approaches (Linear discriminate analysis, random forest, and a neural network). The random forest model had the highest accuracy against an independent test set, at the ecosystem level and for all of North America. The random forest also had good certainty indicators which could be used to rate site suitability of sites for assisted migration. However precipitation values have a high feature importance so errors in precipitation projections could compromise the models predictions.  Overall the random forest model could be a valuable decision support tool for assisted migration in North America.