Exploring the Spatiotemporal Influence of Climate on American Avian Migration with Random Forests

Abstract

North and South American Birds have adapted to climatic and ecological patterns to inform their Spring and Fall migration timings. Temperature and precipitation patterns are shifting under anthropogenic climate change, which has downstream effects on for instance plant flowering cycles, insect populations, and habitat availability. Understanding how these cues trigger migration could improve the effectiveness and timing of bird surveys, as well as informing habitat protection and creation efforts to lessen biodiversity loss due to climate change. In this work, we train an ensemble of random forest regressors on subsets of North and South American climate data to predict distributions of historical eBird occurrence probability for passerine bird species in a North American forested region using eBird citizen science surveys from 2008-2018. By running these ensembles with lagged climate data, we study spatiotemporal effects on bird migration through the resulting error and feature importance metrics. We find statistically significant decreases in error when using lagged climate features rather than the same-month climate to predict species communities in October, when many passerines have begun their southward winter migration, with longer temporal effects for precipitation than temperature. North American climate features were more important than South American features for prediction of species communities in Fall months, suggesting the model drew upon regional climatic cues for predicting when species would migrate southward. By predicting species occurrence using 2021-2040 climate projections, we predict that projected species occurrence probabilities will increase in April and May, agreeing with existing literature suggesting earlier Spring arrival of migratory birds. These results demonstrate that machine learning models have the potential to elucidate complex relationships between climate and bird behavior that linear models may miss. A deeper understanding of such spatiotemporal effects will be required to support biodiversity protection through climate change.