Products
Two kinds of GNN models for Pacific Coast States mapping
The use of multivariate statistics and imputation by GNN results in unique spatial modeling properties. Alternative model forms can be specified to optimize for different objectives and outcomes. For the Pacific Coast States mapping, we are providing two GNN models and associated map products for each modeling region, a GNN ‘species model’ and a GNN ‘species-size’ model:
Species model: Response variables used in model development are basal area by tree species. Landsat, disturbance, and ownership variables are not included as explanatory variables. This model provides the most accurate spatial predictions of distributions of individual species and of community types that are defined based on species composition. Stand structure variables are not attached to this grid.
Species-size model: Response variables used in model development are basal area by species and size-class. This model weighs information on both species composition and stand structure, and resulting maps are useful for applications where elements of both are important – especially where it is important to maintain the covariance among these vegetation elements (e.g., if tree lists are to be input into simulation models such as the Forest Vegetation Simulator).
For the recently completed GNNFire
project, we provided additional GNN models based on varying the spatial
resolution of (i.e., median filtering) the Landsat data used in model
development, and additional models that place greater emphasis on forest
structure.
Map products for each modeling region include the following:
- GNN species model (forest only, with nonforest mask). Joined variables are: Ecological System (where available), individual tree species abundances.
- Nearest-neighbor distance grid for GNN species model.
- GNN species-size model (forest only, with nonforest mask). A standard set of vegetation attributes are joined to the grid. All Oregon maps are based on 2000-2001 Landsat imagery.
- Nearest-neighbor distance grid for GNN species-size model.
- A grid of nonforest vegetation and land cover for use with GNN maps (nonforest mask). Where available, this will be developed from maps of Ecological Systems.
To download GNN map products, visit the modeling regions and schedule page and then click on your region of interest.
GNN imputation maps
Digital GNN imputation maps are provided as 30-m-resolution ArcGIS grids,
where the grid value is a unique plot number that links to the plot database.
Selected vegetation variables from the plot database are joined as items
in the grid to facilitate viewing and exploratory spatial analysis. Metadata
for the vegetation variables are included with the grids and in the plot
database. Dates for maps developed from GNN species-size models are determined
by the vintage of the satellite imagery used in their development. For
Oregon, all GNN species-size maps are based on 2000-2001 imagery.
We are distributing ‘masked’ versions of the GNN maps, where
areas of nonforest land cover developed from ancillary data sources have
been embedded in the GNN grids. The GNN models apply only to forest land
(areas currently or with the potential to support at least 10% tree cover),
because a consistent regional plot sample of nonforest areas is unavailable.
For our masks in Oregon and Washington we are using either 2001 National
Land Cover Data (NLCD), or the Ecological Systems grids where available.
Unmasked versions of the GNN maps are available on request, for users who
would like to apply a nonforest mask of their own choosing.
GNN nearest-neighbor-distance maps
Digital map where the grid value is the distance to the nearest-neighbor plot that was imputed to the pixel. Distance is unit-less Euclidean distance in eight-dimensional gradient space for the first eight axes from canonical correspondence analysis (CCA), with axes weighted by their explanatory power in the model (eigenvalues), and converted to integer grids (see Ohmann and Gregory 2002). The nearest-neighbor distance gives a spatially explicit indication of plot support for the GNN model. Areas with shorter distances are more likely to have better prediction accuracy than those with longer distances.