Species Distribution Models for Pectis imberbis, a Rare Plant Species in Southeastern Arizona

Species distribution models (SDMs) can be an important tool in rare species conservation. Specifically, SDMs have been used to location previously unknown populations and identify sites for reintroduction or translocation. With these goals in mind, we applied SDM to a recently listed plant species, Pectis imberbis, which is found in the Madrean Archipelago region of southern Arizona, USA, and northern Mexico. We used presence-pseudoabsence data and applied 10 replicates of 5 modeling algorithms, generalized linear model (GLM), generalized additive model (GAM, generalized boosted model (GBM, aka boosted regression trees), random forests (RF), and classification tree analysis (CTA) to 4 different predictor datasets which were developed with correlation analysis, feature selection, and variable importance analysis. The resulting models were evaluated based on k-fold cross validation using 4 different metrics: Cohen’s kappa statistic (K), the area under the curve of the receiver operating characteristic curve (ROC), the True Skill Statistic (TSS), and the Boyce Index (BI). High performing models were then included in ensemble model building using the ensemble mean, ensemble median, and committee averaging methods. We applied optimized threshold values to transform continuous species presence probability rasters into binary presence/absence rasters. We also computed the coefficient of variation for the model components of each predictor dataset.
Based on calibration metrics, coefficient of variation between component models, consistency across known populations, and data coverage, the best model from these analyses for this application is the ensemble committee averaging based on the ROC metric using the 5-predictor dataset with USGS geologic data and a threshold based on optimizing the TSS (USGS-5v_EMcaByROC_binTSS.tif). However, we present all 100 modeling outputs rasters (24 continuous rasters, 72 binary rasters, and 4 coefficient of variation rasters) in this data release.