The Virginia Department of Conservation and Recreation – Natural Heritage Program (DCRDNH) and the Florida Natural Areas Inventory (FNAI) at Florida State University (collectively, Project Partners) were funded by the South Atlantic Landscape Conservation Cooperative (SALCC) in April 2015 to develop ten species distribution models (SDM) of priority at-risk andrange-restricted species (Ambystoma cingulatum, Echinacea laevigata, Heterodon simus, Lindera melissifolia, Lythrum curtissii, Notophthalmus perstriatus, Phemeranthus piedmontanus, Rhus michauxii, and Schwalbea americana) for the purposes of incorporating the models and supporting information on the conservation and management needs of the species into the SALCC’s Conservation Blueprint.
Species location data were taken from the Biotics databases maintained by each Natural Heritage program in all 6 states within the SALCC. Data from additional sources were also obtained and utilized. We reviewed each location record for each species to evaluate its value for use in this project. Species observations not meeting a determined set of criteria were not included. Accepted locations were further reviewed and edited if necessary to ensure the inclusion of appropriate habitats. Presence points were randomly generated from within the final location polygons.
Because true absence data are rarely available, we generated a random set of background points (pseudo-absences) to represent locations where a species is not known to occur. Environmental variables (n = 88) were developed at a 30m resolution for the entire region plus a 5km buffer. Our methods for developing these variables are provided as metadata. These variables represented various gradients associated with temperature, precipitation, geology, land cover, hydrological features, and topography.
SDMs were built using Random Forest, a machine-learning approach, implemented with the R statistical package. Random Forest is an ensemble modeling method, creating thousands of classification trees from randomly sampled presence points, background points, and environmental variables. The result is an output raster of probability values depicting where suitable habitat for a species may occur. For each model we calculated validation statistics such as the true skill statistic (TSS) and several statistically-derived threshold values used to convert a continuous probability raster to a binary raster representing suitable vs. unsuitable habitat. This information and other metadata are provided for each SDM.
Overall, the models performed well with TSS scores, a measure of model performance, ranging from 0.82 – 0.98. Choice of threshold value with which to depict suitable habitat may vary depending upon the intended purpose. For example, identifying new areas to survey for a species or for potential species reintroductions could argue for a high threshold thus keeping areas with a high probability of being suitable. Performing an initial recommendation for project review may use a lower threshold to ensure capturing potentially suitable habitat for additional scrutiny.
Short profiles for each species are reported which summarize life history and known threats, and recommendations for potential conservation and monitoring protocols are suggested. Further, these recommendations are placed into context with the model outputs.
