Predicting habitat models of six rare montane plant taxa of northern California

We used 25 years of US Forest Service occurrence data on six rare plant species in the northern California Coast Range to compare predictions of four habitat modeling techniques. These statistical models include: multiple logistic regression (MLR), artificial neural networks (ANN), random forests (RF-a classification and regression type model) and maximum entropy (ME). A secondary objective of the study was to provide the best possible map to predict possible additional population locations. This goal was challenging because our focal species are serpentine soil endemics with narrow geographic ranges and few known populations. Thus, this study challenged model performance with minimal data sets. We found that model fit was, in general good despite small sample sizes and that ME outperformed the other models using the area-under-the-curve (AUC) criterion of the receiver operating characteristic (ROC), but that ME was restrictive in its classification of appropriate habitat—predicting occurrences primarily in the southern portion of the study area amongst  the existing known populations . Model assessment, using both field reconnaissance and a separate US Forest Service data set showed that MLR, ANN and RF provided consistent predictions of appropriate habitat across models and resulted in high habitat suitability scores over a wider geographic area. Best model predictions varied among species in model validation. An average score using these three models provided the best overall predictive results among all species.