Discussion- Accuracy of GNN Vegetation Maps

For the Washington and California study sites, species distributions were modeled equally well with the 250-m data and the 30-m data. Both the Kappa statistics and visual inspection of species maps indicated essentially the same pattern when moving from 30 to 250-m data. Because the gradient models were largely composed of climate variables, there is actually little loss in predictor data information when using the 250-m data. This is because the climate data for both the 250-m and 30-m studies were interpolated from the same 1-km resolution data. Species performance in Oregon was not as good, though it was also less accurate for the 30-m data. In both the 30-m and 250-m studies, we saw some definite differences between the results for Oregon and the results for Washington and California. Both California and Washington are precipitation limited, receive most of their precipitation during winter, and have large elevation gradients. The Coast Range in Oregon has much higher precipitation, milder temperatures, and lower overall topographic variation resulting in less orographic precipitation. Coastal Oregon has also had a long history of timber management and, therefore, has a large patchwork of even-aged stands.

Sources of Error in GNN

GNN and all nearest neighbor techniques, are particularly susceptible to errors introduced by natural variability at spectrally and environmentally similar sites. Whereas regression techniques model a trend and the departure from that trend, imputation retains the full range of variability within a dataset. As such, for a certain location, a regression model with little predictive capability will predict the mean plus some small departure based on predictor variables and coefficients, whereas imputation will find the most environmentally similar site and select it. The tendency for imputation to impute similar values to the actual target values is constrained by the strength of the relationship between available spatial predictor variables and the target response variables.

Other sources of error include: (1) residual spatial error of predictor data sets and plot locations as well as plot registration, (2) temporal mismatches between inventory dates and imagery dates, and (3) the lack of adequate disturbance and management history across large regions.

Advantages of GNN for Risk Assessment

There are several advantages to GNN for risk assessment. GNN retains the covariance structure for multiple attributes by imputing whole plots and provides mapped estimates of natural variability and sample sufficiency (Pierce and others, in review). Comparative risk assessment requires spatially explicit data with estimates of variability (Borchers 2005) in order to create probability surfaces for different management scenarios. For instance, what is the probability of the desired outcome given two different management choices, and are they statistically different? Without an estimation of uncertainty, this type of analysis can’t be performed. By using a set of multiple potential neighbors, the variability in potential neighbors for a selected attribute can be mapped. In addition, by using the frequency distribution of all interplot distances in gradient space, thresholds for closeness in gradient space can be assigned and the number of candidate plots within a threshold calculated. This gives an indication as to whether or not the inventory can provide adequate information for a certain pixel, and, as such, a map depicting the sampling support can be created.

Species Response Models in Multispecies Mapping

One of the key areas of interest in natural resource risk assessment is the interactions among species. The location of invasive species and the presence of host species are two data surfaces of interest to managers. Mapping with single neighbor imputation ensures that the assemblages of tree species mapped are consistent with actual inventoried assemblages. This has both benefits and limitations. The benefit is robust assemblages of species as currently exist. The limitation is that prediction for new interactions can not be inferred on the basis of these maps. Single species models are probably best suited for predicting suitable habitat for an individual species, or rather the present distribution of habitat consistent with currently occupied habitat.

Risk Assessment Applications of GNN Predictions

Single neighbor imputation using GNN provides a very flexible wall-to-wall data set that includes any variable that can be derived from those measured on all inventory plots. This includes the ability to derive new variables or vegetation classifications after the initial modeling. GNN imputation also provides a link to the full tree lists allowing for almost any kind of ecological modeling. The inclusion of multiple neighbors provides uncertainty data for Monte Carlo simulations or analyses seeking to show the uncertainty associated with different scenarios. As new risks or identification of new data needs arise, imputation maps are ready to adapt to new needs without the necessary production of a new model. However, at the 250-m scale, the variables, which are correlated with broad climate patterns, specifically species distributions, will probably be characterized the best. Structure attributes, such as coarse woody debris and quadratic mean diameter, can be mapped, but the mapped variability will likely overwhelm the utility of such products.