Moderate Resolution Data and Gradient Nearest Neighbor Imputation for Regional-National Risk Assessment

Ken Pierce, Ken Brewer and Janet Ohmann

USDA Forest Service Pacific Northwest Research Station (1 and 3) and Remote Sensing Applications Center (2)

One of the most fundamental informational needs in environmental risk assessment is consistent and continuous existing vegetation data.  These data must be of sufficient accuracy and precision to address the complex and often uncertain ecological relationships necessary to understand risk factors and management options.  Where these data do exist they are normally based on a sampling inference procedure rather than “wall to wall” inventory data.  This is particularly true when the risk assessment covers large geographic extents and multiple ownerships. Despite the capability of simulation models and decision support tools comprehensive large area risk assessment is still difficult to implement because the inventory data are rarely complete and/or current. It would be convenient to be able to operate as if detailed inventory information were available for all units in the risk assessment area. 

As an alternative to historically common statistical approaches (e.g., regression estimates or stratum averages) to populating unsampled units with data, imputation can be used.  Imputation involves estimating values for variables of interest (Y variables) by supplying realistic measurements from one or more sampled units to unsampled units with similar characteristics in auxiliary (X) variable-space.  Imputation of inventory data from sampled areas to similar unsampled areas produce datasets that function like “wall to wall” data for risk assessment purposes.  The Gradient Nearest Neighbor (GNN) imputation method developed by Ohmann and Gregory has been implemented successfully over large geographic areas to characterize existing vegetation and fire and fuels characteristics.  As currently implemented, this method develops 30 meter raster surfaces of inventory data that can then be readily analyzed for characterizing risks.

The USFS Forest Inventory and Analysis (FIA) program has shown that it is possible to derive national mapping products using moderate resolution remote sensing images and GIS data layers (e.g. forest biomass, forest types/type groups).  These maps were produced for the continuous US and Alaska at a spatial resolution of 250 meters using FIA plot data and a geospatial predictor database, consisting of approximately 300 data layers.  Having this consistent national-scale forest type information is important for modeling forest areas at risk of mortality due to insects and diseases as well as other environmental threats.

The current work (sponsored by the Western Wildland Environmental Threat Assessment Center) utilizes the 250 meter geospatial predictor data for the Eastern Washington GNN project area to explore the feasibility of using moderate resolution data and GNN imputation over regional-national geographic extents.  The utility of the 250 meter data surface is evaluated using the 30 meter data surface from the previous Joint Fire Science project work.  While the use of moderate resolution data is a promising approach its utility is both a function of the geographic extent and specific analysis objectives.  The 250 meter data surface is more appropriate for regional- and national-level analyses.

Statistical Methods Session - Wednesday Afternoon

corresponding author:

Ken Pierce
USDA Forest Service
3200 SW Jefferson Way
Corvallis, OR 97331
541-750-7393
kpierce@fs.fed.us