Impacts of humans on ecosystems in western United States have necessitated ecological restoration, which includes the development of native seed that can be used for revegetation efforts. Development of such seed sources are costly and time consuming. This study describes the use of non-parametric multiplicative regression analysis (NPMR) to develop a predictive model for occurrence of sulfur-flower buckwheat (Eriogonum umbellatum Torrey) population seed collection. This perennial forb species is of interest for seed source development in the western United States. Presence and absence data for E. umbellatum was taken from the Utah Division of Wildlife Resources Big Game Range Trend project as well as herbarium specimens across Utah, U.S.A. NPMR, a statistical niche modeling system that selects the best predictor variables and develops probability of occurrence estimates multiplicatively, was used to select predictor variables from spatially explicit data made available in a Geographic Information System (GIS). Two models were created using NPMR, one with a suggested default minimum average neighborhood size and the other with a less-restricted minimum average neighborhood size. GIS maps of models were created, artificially classified into low, medium, and high probability areas, and validated in the field in Tooele County, Utah. Of 68 possible physiographic, climatic, and soil variables provided for analysis, NPMR selected 4 variables for the default minimum average neighborhood model and 10 variables for the less restricted neighborhood model. The default model had a higher descriptive statistic (log β value) and mapped a larger area than the less restrictive neighborhood model. When increased minimum neighborhood sizes were selected during the development of the probability maps, the resulting areas of probability prediction decreased. The presence rates of E. umbellatum in field-validated test sites were 7.4%, 12.0%, and 28.6% for the low, medium, and high probability sites, respectively. Although presence rates of field validated data were lower than the predicted probability ranges for those same sites, presence rates increased with increased probability ranges. Using the generated model can reduce the cost and time necessary to locate plants compared to searching for species populations using an undirected approach.