Modeling species distribution dynamics with SpatioTemoral Exploratory Models: Discovering patterns and processes of broad-scale avian migrations

AbstractThe distributions of animal populations are not static. During regular migratory movements species exploit different habitats. This spatiotemporal variation needs to be accounted for when modeling a species’ distribution and is essential for developing conservation strategies for widespread species, and especially for migratory species. Attempts to design conservation landscapes across large regions based on models of distributions in a single season or a small region may not fully reflect the limiting factors that are driving population declines. Our goal is to predict and explore patterns of species’ occurrence and local habitat usage across broad landscapes. We use data from eBird (http://www.ebird.org), an online citizen science bird-monitoring project and environmental descriptions from continent-wide covariates linked through observation location and time. These covariates include remotely sensed habitat information from the National Land Cover Database and vegetation phenology from MODIS. We model species occurrence with the SpatioTemporal Exploratory Model (STEM), an ensemble model designed to adapt to non-stationary spatiotemporal processes. This is accomplished by creating a large ensemble of local models, each restricted to a local spatial and temporal region. Within each region a user specified predictive model associates the predictors with the response. Patterns modeled locally “scale up” via ensemble averaging to larger scales. Here we analyze eBird data to study broad-scale movements of bird populations throughout the year. We use STEM built with decision trees to adapt to a wide variety avian migration patterns without requiring a detailed understanding of the underlying dynamic local processes. We demonstrate how eBird data are capable of resolving the changing distributions of birds through their migrations. Then we illustrate how seasonal variation in habitat association can be identified and explored. These tools provide valuable information for generating hypotheses and making inference about the processes driving dynamic species distributional patterns