Effective conservation planning requires reliable information on the distribution of species, which is generally incomplete for many species, especially for rare ones, due to limited observation data. Species distribution models (SDM) are highly valuable in determining critical remaining and potential habitats of at-risk species for conservation planning. Despite the proliferation of SDM and tools in the past two decades, management programs have not fully adopted them to inform species surveys and other monitoring efforts; instead, many rely on expert knowledge and other traditional methods to locate extant populations. One important framework that would benefit from SDM is the Species Status Assessment (SSA) developed by the U.S. Fish and Wildlife Service. While a SSA considers multiple elements associated with species condition, including distributions, there are no standard requirements to estimate species distributions. Our objective was to find an optimal SDM approach for at-risk species that can be considered for SSA and similar species monitoring efforts. We applied four modeling approaches (regression, machine learning, boosting modeling, and weighted ensemble modeling) to recent monitoring data for three endangered species (Papaipema eryngii, Scutellaria ocmulgee, and Balduina atropurpurea) in the Southeastern U.S.
Results/Conclusions
Among three individual modeling approaches calculated for Papaipema eryngii, Scutellaria ocmulgee, and Balduina atropurpurea, machine learning (AUCs=0.88, 0.9, 0.84, respectively) and boosting modeling (AUCs=0.85, 0.92, 0.94, respectively) had better predicting power relative to regression-based approaches (AUCs= 0.71, 0.83, 0.92, respectively). The weighted ensemble model, however, had the highest model accuracy for each of the three species (AUCs=0.92, 0.94, 0.98, respectively). As such, the ensemble models reduced the predictive uncertainty caused by differences among the three modeling approaches. We conclude that ensemble SDMs provide the best predictive power for our three species, as indicated by both testing AUC and independent ground-truthed surveys for some species. We suggest that this approach could be adopted into the SSA framework to improve monitoring efforts and contribute to more robust assessments of at-risk species.