Combining local‐ and large‐scale models to predict the distributions of invasive plant species

Abstract

Habitat distribution models are increasingly used to predict the potential distributions of invasive species and to inform monitoring. However, these models assume that species are in equilibrium with the environment, which is clearly not true for most invasive species. Although this assumption is frequently acknowledged, solutions have not been adequately addressed. There are several potential methods for improving habitat distribution models. Models that require only presence data may be more effective for invasive species, but this assumption has rarely been tested. In addition, combining modeling types to form "ensemble" models may improve the accuracy of predictions. However, even with these improvements, models developed for recently invaded areas are greatly influenced by the current distributions of species and thus reflect near- rather than long-term potential for invasion. Larger scale models from species' native and invaded ranges may better reflect long-term invasion potential, but they lack finer scale resolution. We compared logistic regression (which uses presence/absence data) and two presence-only methods for modeling the potential distributions of three invasive plant species on the Olympic Peninsula in Washington, USA. We then combined the three methods to create ensemble models. We also developed climate envelope models for the same species based on larger scale distributions and combined models from multiple scales to create an index of near- and long-term invasion risk to inform monitoring in Olympic National Park (ONP). Neither presence-only nor ensemble models were more accurate than logistic regression for any of the species. Larger scale models predicted much greater areas at risk of invasion. Our index of near- and long-term invasion risk indicates that < 4% of ONP is at high near-term risk of invasion while 67-99% of the Park is at moderate or high long-term risk of invasion. We demonstrate how modeling results can be used to guide the design of monitoring protocols and monitoring results can in turn be used to refine models. We propose that, by using models from multiple scales to predict invasion risk and by explicitly linking model development to monitoring, it may be possible to overcome some of the limitations of habitat distribution models.