Methods
We analyzed vegetation data from 629 field plots established across all
forest lands (Table 1, Fig.
1). Nonforested areas were excluded. We calculated summary measures
of vegetation for each plot from data on tree species, diameter, and age.
Values from mapped topography, geology, climate, and 1996 Landsat TM satellite
imagery (Table 2, Fig.
2) were assigned to each plot in GIS.
We evaluated model performance by comparing mapped predictions to ground
observations on field plots reserved from model development.
We developed Gradient Nearest Neighbor method (GNN) as follows (Fig.
3):
- We quantified relations between ground (response) data and mapped (explanatory)
data using direct gradient analysis (stepwise canonical correspondence
analysis, CCA). We developed two models: in the species model, response
variables were basal area of tree species; in the structure model, response
variables were basal area of broad species groups and size-classes.
- For each pixel, scores on the first eight CCA axes were predicted from
the mapped explanatory variables.
- For each pixel, the single plot was identified that was nearest in
eight-dimensional gradient space. Distances were Euclidean, with axis
scores weighted by their eigenvalues.
- Ground attributes of the nearest-neighbor plot were imputed to the mapped pixel.