COS 260-6 - CANCELLED - An easy automated solution to species distribution modeling

Thursday, August 10, 2023

11:15 AM – 11:30 AM PDT

Location: D140

Presenting Author(s)

BS

Benjamin J. Seliger

Center for Geospatial Analytics, North Carolina State University, United States

Coauthor(s)

CJ

Chris M. Jones

North Carolina State University Center for Geospatial Analytics, North Carolina, United States
BL

Brit Laginhas

North Carolina State University, United States
RM

Ross Meentemeyer

North Carolina State University, North Carolina, United States

Abstract: Managing the spread of invasive pests and pathogens depends upon accurate high-resolution maps of their hosts. Species distribution models (SDMs) are a commonly used technique for producing these maps; however, SDMs require informed inputs to produce accurate results. Our goal is to produce a tool that allows for the quick and easy production of SDMs with minimal user input and fully specified uncertainties, without a loss of accuracy or overfitting.

Here we present an R package (‘pops.sdm’) that automates variable and threshold selection, eliminating two key steps in the modeling process, and runs an ensemble of 8 models to map species presence with spatial uncertainty. Our package identifies the ideal minimum combination of variables from over 40 plausible ecologically-meaningful variables. This method works by first categorizing variables into clusters of correlated variables to account for collinearity and then running a series of models that rank each variable by their ability to predict the species’ range. Then ‘pops.sdm’ selects the best variable, and runs another series of models with each of the remaining variables, excluding variables within the same cluster as the previously selected best variable. This process repeats until subsequent variables do not significantly improve model performance or until all clusters have a variable included in the model. The ‘pops.sdm’ package also includes a new method for finding the ideal threshold to produce binary presence-absence maps. Our metric calculates the threshold that maximizes the proportion of presence points included while minimizing the overall predicted presence area. Another unique feature of ‘pops.sdm’ is that it visualizes spatial uncertainty by producing maps of standard deviation across the models used in the ensemble. Areas with high standard deviations indicate areas to focus future data collection, and standard deviation maps allow for the predictions to be used by forecasting systems that propagate and partition uncertainty in their drivers.

We demonstrate the ability of our approach to produce results as accurate as or more accurate than previous expert-informed approaches using two examples of invasive pest hosts: Tree-of-Heaven (Ailanthus altissima) and Tanoak (Notholithocarpus densiflorus). These species offer ideal cases to test the accuracy of our predictions because they each have multiple independent presence datasets and previous SDM predictions. By automating the approach to using SDMs, we can remove much of the expert-opinion required and allow for more users to produce accurate range maps.