Novel methods to improve predictions of alien plant species richness

Background/Question/Methods

Accurate and reliable predictions of spatial patterns of native and alien species richness are required to set priorities for conservation, management of natural resources, and monitoring and control of invasive species. We tested the hypothesis that relatively novel and sophisticated modeling methods such as boosted regression trees (BRT) might outperform commonly used least-square or log-likelihood methods in predicting the spatial patterns of native and alien plant species richness across the landscape. 

We collected native and non-native plant species data from 176 0.1-ha modified-Whittaker plots in Rocky Mountain National Park, Colorado, USA. We used topographic and remote sensing GIS layers as potential predictors of native and alien plant species richness. We compared three statistical methods: multiple linear regression, zero-inflated negative binomial (ZINB) regression (because of excessive zeros in alien species richness data), and boosted regression trees, for modeling and mapping of native and alien species richness in the Park. All analyses were conducted in R statistical software using the MASS, ‘pscl’, and ‘gbm’ packages. The best models for multiple linear regression were selected using ‘stepAIC’ function in MASS. Model performance was assessed based on correlations between predicted and observed species richness.

Results/Conclusions

We found that boosted regression trees performed far better (R² = 0.83) than the commonly used multiple linear regression (R² = 0.58), and zero-inflated negative binomial regression (R² = 0.71) methods for predicting alien species richness. Boosted regression trees also outperformed multiple regression and ZINB regression methods for predicting native species richness with an R² value of 0.86. All three modeling methods showed that elevation was the best predictor of alien species richness, followed by solar radiation and remotely sensed Enhanced Vegetation Index (EVI). Our results suggest that novel methods such as boosted regression trees, which are non-parametric and automatically include variable interactions, might be better suited for mapping and modeling of spatial patterns of native and alien species richness at landscape scales.