Machine learning matched forest plot data with biophysical characteristics of the landscape to produce a seamless tree-level forest map.
A map of the location, size, and species of every tree in the forests of the United States would be useful for any number of applications, ranging from habitat mapping to estimation of carbon resources. No such map currently exists, but we have used machine learning to make a spatial model of the trees in continental US forests. Detailed maps and descriptions of trees are available where the Forest Inventory Analysis program measures their annual forest plots, however, these plots are widely spaced (approximately one plot per 24.3 km). We based our TreeMap on this detailed forest plot data measured by FIA (https://www.fia.fs.fed.us/) and national gridded maps of forest cover, height, and vegetation type provided by the LANDFIRE project (https://www.landfire.gov/). Using random forests (a type of machine learning algorithm) we match the forest plot data to the gridded vegetation maps, producing a seamless map of the trees of the forests of the US. Specifically, the result is a map of plot ID numbers, which identify the best-matching forest plot for each 30x30 pixel in the map. The map of plot ID numbers can be linked back to the FIA databases (https://apps.fs.usda.gov/fia/datamart/) to generate maps of any number of forest characteristics, ranging from basal area to biomass to species types. Some current uses of the dataset include: 1) inventory of habitat types at regional scales by land managers, 2) research on evaluation of tradeoffs between fuel management and timber harvest targets, and 3) modeling hydrologic effects of fuel treatment.
Click on photos for larger view.