Maps of the number, size, and species of trees in forests across the western United States are desirable for a number of applications including estimating terrestrial carbon resources, tree mortality following wildfires, and for forest inventory. However, detailed mapping of trees for large areas is not feasible with current technologies. We used a statistical method called random forests for matching forest plot data with biophysical characteristics of the landscape in order to populate entire landscapes with a limited set of forest plot inventory data.
Our objective was to match plot data collected by the USDA Forest Service’s Forest Inventory Analysis (FIA) with characteristics of the landscape, as mapped on a 30mx30m grid by the LANDFIRE project. The result is a map of plot identifiers, with the identifier of the best-matching plot assigned to each grid cell. We used a modified random forests approach, which utilizes a “forest” of decision trees in order to identify the best-matching plot for each grid cell. One of the strengths of the random forests method is that it can model complex nonlinear relationships.
The map of plot identifiers can be linked to FIA’s databases to produce tree-level maps (called a “tree list”) or to map a number of other plot attributes. For example, we used the map of plot IDs to generate maps of forest cover, forest height, and existing vegetation group at 30mx30m resolution for all forested pixels in the western United States. We compared these maps with LANDFIRE maps as a means of assessing the accuracy of our product. Our data closely corresponded with the LANDFIRE data, with an overall within-class agreement of 79 percent for forest cover, 96 percent for forest height, and 92 percent for existing vegetation group.
High levels of agreement between our dataset and LANDFIRE data indicate our modified random forests model was able to identify forest plots that closely matched the landscape characteristics of 30mx30m grid cells
Within-class agreement between the tree list and LANDFIRE data was 79 percent for forest cover, 96 percent for forest height, and 92 percent for existing vegetation group
Since the dataset in essence provides a tree-level model of forests in the western U.S., it greatly augments the information available to researchers and managers who previously relied on data from sparse forest plots or stand inventories
Riley, Karin L., Isaac C. Grenfell, and Mark A. Finney. In Press. Mapping forest vegetation for the western US using modified random forests imputation of FIA forest plots. Ecosphere.