LEMMA | The Gradient Nearest Neighbor (GNN) Method

Results

Species gradients were most strongly associated with climate, and structure gradients with Landsat TM data (Table 3). Location, geology, topography, and ownership also were important in both models.
The primary species gradient followed a climatic gradient from coastal areas with frequent summer fog and rainfall to inland areas with high summer moisture stress and less maritime influence (Fig. 4a). The second axis was associated with elevation.
The primary structure gradients were tree size and density, which varied with Landsat TM band 4 and ownership (Fig. 4b). Low scores were in dense stands of large trees on public lands. High scores were younger, more open stands on recently disturbed private lands. Axis 2 differentiated species groups and was associated with the maritime climatic gradient.

At the aggregate, regional level, the mapped predictions captured the means and ranges of variability present in the plot data (Table 4) and portrayal of spatial heterogeneity appeared quite reasonable (Fig. 5).
Because GNN assigns a single nearest-neighbor plot to each pixel, mapped predictions retained the covariance structure among response variables.

Prediction accuracy for occurrence of six tree species (Fig. 6) was 56-93%, or 12-51% better than chance (Table 5). There were more errors of commission than of omission for all species. The species model more accurately predicted species occurrence than the structure model.
Species whose distributions are geographically limited and controlled by climate (Picea sitchensis and Quercus garryana) were most accurately predicted. Widely distributed species that occur in locally low abundances (Acer macrophyllum and Thuja plicata) were more difficult to predict.
Prediction accuracy for selected measures of vegetation structure was moderate to low for specific sites (Fig. 7). For most variables the models slightly over-predicted at low values and under-predicted at high values.
Vegetation classification accuracy was similar to other published image classification methods in our region.
Classification accuracy in a 10-class system was 41%, and “fuzzy” accuracy (+/- one class) was 85% (Table 6). Tree density and composition were more difficult to predict than tree size. Open stands of large trees were especially problematic.

Home