Page 47 - GIS for Science, Volume 3 Preview

IMPROVING MODELS THROUGH COLLABORATION
The machine learning tools used to produce predictive models are powerful. However, AI methods are not foolproof. When a child studying the American Revolution asks a smart device whether the patriots won, she just might be informed that the National Football League team lost 20 to 14 in overtime. When a modeler asks a machine learning algorithm where a species is likely to occur, the results can be similarly misleading. Dozens of choices made regarding modeling parameters can influence the final results, and it’s not always clear which approach is “right.”
To address this issue, NatureServe developed an interactive online Model Review Tool to facilitate expert evaluation of species habitat models developed for MoBI and other conservation applications. The information received allows modelers to identify where machine learning methods have fallen short and gain insight that can support iterative model refinement. These tools help bring together the promise of machine learning and the expertise of scientists who understand these species and their habitats.
The Model Review Tool is a web application built on an ArcGIS Online platform that allows species experts to view model outputs and associated metadata and provide feedback on how well the model performs. Reviewers can import their own
data to evaluate the results, indicate areas where results are questionable, provide comments on how the model might be improved, and rate model performance. The Model Review Tool for the MoBI project helped determine which models performed well, which required adjustments, and for which species we needed to pursue an alternative approach to mapping habitat. The reviews also provide insight about the promise of these predictive modeling approaches, including lessons about which groups of species resulted in the most defensible models. Models for plants tended to perform well, which is unsurprising given their lack of mobility and simpler relationships to the environmental factors represented in the predictor data. Successfully modeling pollinators proved to be more difficult, partly reflecting the complicated relationship between pollinators and their food plants.
Successful modeling of habitat for many pollinators may depend on first developing models for nectar plants. Rarer species, and those with small range sizes, tend to model best, reflecting the geophysical limitations often driving their rarity. While habitat modeling has many applications, these findings highlight its promising application for management of imperiled species.
Map showing at-risk species in protected areas using modeled distributions of the most imperiled plants and animals in the United States.
Preventing Species Extinctions 35