A tale of two wildfires; testing detection and prediction of invasive species distributions using models fit with topographic and spectral indices

Abstract

Developing species distribution models (SDMs) to detect invasive species cover and evaluate habitat suitability are high priorities for land managers. We tested SDMs fit with different variable combinations to provide guidelines for future invasive species model development based on transferability between landscapes. Generalized linear model, boosted regression trees, multivariate adaptive regression splines, and Random Forests were fit with location data for high cheatgrass (Bromus tectorum) cover in situ for two post-burn sites independently using topographic indices, spectral indices derived from multiple dates of Landsat 8 satellite imagery, or both. Models developed for one site were applied to the other, using independent cheatgrass cover data from the respective ex situ site to test model transferability. Fitted models were statistically robust and comparable when fit with at least 200 cover plots in situ and transferred to the ex situ site. Only the Random Forests models were robust when fit with a small number of cover plots in situ. Our study indicated spectral indices can be used in SDMs to estimate species cover across landscapes (e.g., both within the same Landsat scene and in an adjacent Landsat scene). Important considerations for transferability include the model employed, quantity of cover data used to train/test the models, and phenology of the species coupled with the timing of imagery. The results also suggest that when cover data are limited, SDMs fit with topographic indices are sufficient for evaluating cheatgrass habitat suitability in new post-disturbance landscapes; however, spectral indices can provide a more robust estimate for detection based on local phenology.