Using species distribution models to predict new occurrences for rare plants

Aim  To evaluate a suite of species distribution models for their utility as predictors of suitable habitat and as tools for new population discovery of six rare plant species that have both narrow geographical ranges and specialized habitat requirements.
Location  The Rattlesnake Creek Terrane (RCT) of the Shasta-Trinity National Forest in the northern California Coast Range of the United States.
Methods  We used occurrence records from 25 years of US Forest Service botanical surveys, environmental and remotely sensed climate data to model the distributions of the target species across the RCT. The models included generalized linear models (GLM), artificial neural networks (ANN), random forests (RF) and maximum entropy (ME). From the results we generated predictive maps that were used to identify areas of high probability occurrence. We made field visits to the top-ranked sites to search for new populations of the target species.
Results  Random forests gave the best results according to area under the curve and Kappa statistics, although ME was in close agreement. While GLM and ANN also gave good results, they were less restrictive and more varied than RF and ME. Cross-model correlations were the highest for species with the most records and declined with record numbers. Model assessment using a separate dataset confirmed that RF provided the best predictions of appropriate habitat. Use of RF output to prioritize search areas resulted in the discovery of 16 new populations of the target species.
Main conclusions  Species distribution models, such as RF and ME, which use presence data and information about the background matrix where species do not occur, may be an effective tool for new population discovery of rare plant species, but there does appear to be a lower threshold in the number of occurrences required to build a good model.