OPS 1-2 - Development of the automated labeling tools for vegetation sample data and their application to national mapping

Monday, August 7, 2017
Exhibit Hall, Oregon Convention Center
Patrick J. Comer, Ecology Department, NatureServe, Boulder, CO and Donald Long Jr., LANDFIRE, U.S. Forest Service, Missoula, MT

The LANDFIRE Program produces national model and map products for vegetation, fuel, and fire regimes to address natural resource inventory, assessment, planning, and monitoring. LANDFIRE has produced new tools called “Auto-Keys” for processing vegetation sample plot data for subsequent modeling and mapping. This effort has processed over 500,000 vegetation plots; standardizing many sample attributes (e.g., species taxonomy, structural classes) and applied labels to each reflecting the LANDFIRE map legend. Regional Auto-Keys are now updated for labeling georeferenced vegetation samples to NatureServe Terrestrial Ecological System types and USNVC Group units in the conterminous USA. Each sample plot must meet all of the criteria for a particular classification unit; if the sample meets all the criteria, the auto-key attributes the plot with the type name and code. The program also generates a number of output tables to enable the user to evaluate the results. In the project area, 541 upland and wetland ecological system types, and 278 US NVC Groups are recommended for the LANDFIRE legend. Each Auto-key pertains to one of 16 auto-key regions (AKR), excluding the southern tip of Florida, which will be included with a future Caribbean effort.


A total of 427,309 plots were included in the key databases across all 16 AKRs; of these, 17,934 were reviewed and labeled by expert ecologists for use in assessment of Auto-Key performance. The total number for each AKR varied, from a high of 80,148 in the Rocky Mountains to 3,517 in the North Coast AKR. For most of the AKRs, the proportion of expert plots used in assessment was 4% to 8%. The overall validation statistic is a useful measure of how well the key is working across all types. It is the number of matches between expert and auto-key labels divided by the total number of expert plots x 100, and it was calculated for each Auto-Key. The overall agreement varies across the AKRs; with the overall agreement for ecological systems keys ranges from a high of 85% (Colorado Plateau) to a low of 36.5% (Coastal Plain). Overall agreement for the USNVC Group keys ranges from a high of 90% (Texas-Oklahoma Hill Prairie) to a low of 39.9% (Coastal Plain). In most cases, lower performance occurs where substantial proportions of the landscape are dominated by ruderal vegetation, and distinguishing among vegetation types using sample plots becomes much more challenging.