Predictive spatial niche and biodiversity hotspot models for small mammal communities in Alaska: applying machine-learning to conservation planning

Context Changing global environmental conditions, especially at northern latitudes, are threatening to shift species distributions and alter wildlife communities. Objective We aimed to establish current distributions and community arrangements of small mammals to provide important baselines for monitoring and conserving biodiversity into the future. Methods We used 4,408 archived museum and open-access records and the machine learning algorithm, RandomForests, to create high-resolution spatial niche models for 17 species of rodents and shrews in Alaska. Models were validated using independent trapping results from 20 locations stratified along statewide mega-transects, and an average species richness curve was calculated for field samples. Community cluster analyses (varclus) identified geographic patterns of sympatry among species. Species models were summed to create the first small-mammal species richness map for Alaska. Results Species richness increased logarithmically to a mean of 3.3 species per location over 1,500 trapnights. Distribution models yielded mean accuracies of 71 % (45-90 %), and maps correctly predicted a mean of 75 % (60-95 %) of occurrences correctly in the field. Top predictors included Soil Type, Eco-region, Landfire Land-cover, December Sea Ice, and July Temperature at the geographic scale. Cluster analysis delineated five community groups (3-4 species/group), and species richness was highest (11-13 species) over the Yukon-Tanana Uplands. Conclusions Models presented here provide spatial predictions of current small mammal biodiversity in Alaska and an initial framework for mapping and monitoring wildlife distributions across broad landscapes into the future.