Understanding resource selection by elk (Cervus elaphus) at multiple spatial scales may provide information that will help resolve the increasing number of resource conflicts involving elk. We quantified vegetation at 412 sites where the precise location of elk was known by direct observation and 509 random sites in the Black Hills of South Dakota during 1998-2001. We obtained stand level variables from geographic information system coverages for a 250 m buffer around elk and random sites. We used classification and regression trees to simultaneously evaluate environmental thresholds of resources selected by feeding and bedded elk at the site and stand spatial scales. Length of roads (≤11.4 m) within 250 m differentiated elk feeding and bed sites from random sites. Only 24% of random sites were misclassified based on this criterion. Ninety-two percent of sites selected by bedded elk had little or no occurrence of the open forest shrub (western snowberry [Symphoricarpos occidentalis]) demonstrating general selection of medium to dense forest while elk were bedded. Whereas, 82% of elk feeding sites and 95% of the random sites with 2% cover of western snowberry. Feeding sites of elk not classified based on snowberry cover occurred in grassland and open forest vegetation types while random sites mostly occurred in pole-size forest with >40% overstory canopy cover and mature forest stands >70% overstory canopy cover. Overall, the estimated error of the classification and regression tree model was approximately 15%, with less clear separation occurring in the second and third nodes of the tree. Site characteristics were more important than stand characteristics determining sites selected by elk.