Abstract
American red squirrels (Tamiasciurus hudsonicus) are small mammals that are abundantly distributed throughout North America. Urbanization in the Anthropocene is now a global process, and squirrels live in affected landscapes. This leads to squirrels adjusting to human developments. Not much is known about the distribution of squirrels and squirrel middens near humans, especially not in the subarctic and sub-urbanized regions. Although this species is hunted, there are no real publicly available distribution and abundance estimates nor management plans and bag limits for squirrels in Alaska or in the United States known by us, except the endangered Mt. Graham squirrel. In general, insufficient squirrel conservation research is carried out; they are underrepresented in research and its literature. To further the science-based management for such species, this study aims to generate the first digital open access workflow as a generic research template for small mammal work including the latest machine learning of open source and high-resolution LIDAR data in an Open Source Geographic Information System (QGIS) and ArcGIS. Machine learning has proven to be less modeler biased and improve accuracy of the analysis outcome, therefore it is the preferred approach. This template is designed to be rapid, simple, robust, generic and effective for being used by a global audience. As a unique showcase, here a squirrel midden survey was carried out for two years (2016 and 2017). These squirrel middens were detected in a research area of 45,5 hectares (0,455 km(2)) in downtown Fairbanks, interior boreal forest of Alaska, U.S. Transect distances were geo-referenced with a GPS and adjusted to the visual conditions to count all squirrel middens within the survey area. Different layers of proximity to humans and habitat characteristics were assembled using aerial imagery and LIDAR data (3D data needed for an arboreal species like the red squirrels) consisting of a 3 × 3 m resolution. The layer data was used to train a predictive distribution model for red squirrel middens with machine learning. The model showed the relative index of occurrence (RIO) in a map and identified canopy height, distance to trails, canopy density and the distance to a lake, together, as the strongest predictors for squirrel midden distribution whereas open landscape and disturbed areas are avoided. It is concluded that squirrels select for high and dense forests for middens while avoiding human disturbance. This study is able to present a machine learning template to easily and rapidly produce an accurate abundance prediction which can be used for management implications.