Predictive Habitat Modeling of Rare Plant Species in Pacific Northwest Forests
Abstract Certification requirements associated with the Sustainable Forestry Initiative include efforts to identify and protect occurrences of endangered plant species. Habitat models were constructed in this study using maximum entropy and random forest algorithms to generate independent predictions for four selected rare plants, Castilleja chambersii, Erythronium elegans, Filipendula occidentalis, and Sidalcea nelsoniana, associated with divergent physical environments. Explanatory variables used to model rare plant occurrence included Landsat Enhanced Thematic Mapper Plus spectral imagery, spectral-based vegetation indices, climatic data, and several terrain variables derived from a digital elevation model. Models were trained with known occurrence records obtained from the Oregon Biodiversity Information Center. Subsequent field surveys were conducted to acquire randomly located test data for comparative model evaluation. A range of accuracy statistics was computed that indicated generally high prediction accuracy for both methods. Model performance was highest for species with narrow, well-defined ecological requirements at scales comparable to the resolution of the calibration data. Species with relatively broad ecological distributions or with extremely specific habitat requirements were less accurately predicted. Random forest-based models generally produced higher rates of prediction success than maximum entropy when model performance was limited by the ecology of a species. Field surveys identified 22 previously unknown populations of the target rare plants, suggesting the efficacy of habitat models for predicting rare species occurrence and their utility as a prescriptive tool for land management planning.