Choosing appropriate subpopulations for modeling tree canopy cover nationwideThis article is part of a larger document. View the larger document here.
In prior national mapping efforts, the country has been divided into numerous ecologically similar mapping zones, and individual models have been constructed for each zone. Additionally, a hierarchical approach has been taken within zones to first mask out areas of nonforest, then target models of tree attributes within forested areas only. This results in many models nationwide, which reduces the number of training points per model, increases the cost of the process, results in numerous seam lines, and complicates validation efforts. Consequently, we use response data based on photo-interpreted aerial photography and spatially continuous predictor data (Landsat imagery, topographic and other ancillary data) in five pilot areas across the country to explore the effect of the choice of modeling subpopulation on models of tree canopy cover. Using Random Forests as our predictive tool, we explore the consequences of modeling pilot areas alone, modeling groups of pilot areas, and modeling hierarchically within each pilot area. Recommendations are made for appropriate modeling subpopulations to be used in a nationwide tree canopy cover map.