Conclusions

Creating an acceptable model of a complex, dynamic system is always challenging. Landscape-scale analyses increase the level of complexity by limiting the variables that are available for consideration. Here we applied knowledge- and data-driven approaches to model hemlock decline following HWA infestation. An initial review of existing literature identified potential variables for model inclusion and helped direct field measurement efforts. We used a mixed stepwise linear regression model based on plots that have been infested for at least 4 years to identify the set of landscape and chemical variables that could best predict the average rate of hemlock decline since HWA infestation. Using a continuous output variable instead of a simple tolerant/susceptible classification allows for flexibility in interpretation, depending on the needs of the user. For example, a research scientist with a limited number of HWA predator beetles to release, may choose a conservative approach, selecting the stands with the lowest anticipated rate of decline and, therefore, highest probability of continued health in spite of HWA infestation. Conversely, a forester who wants to preserve a strong genetic pool of hemlock may decide not to cut any hemlock in stands that have even a marginal probability of sustaining long-term infestation with minimal health impacts.

Hemlock vulnerability to HWA is complex and likely results from a combination of landscape and chemical factors. Because the ability to map relative risk on a landscape scale could prove to be a useful tool for managers faced with HWA, we limited ourselves to variables available in digital, raster format for inclusion in a GIS model. A model based only on topographic variables derived from a 10-m DEM was able to account for almost one-third of the variability in hemlock decline rates from infested plots across the northeast. This is consistent with previous studies that link variables related to soil-moisture availability with hemlock vulnerability. By adding foliar nitrogen concentrations to the model, over two-thirds of the variability in hemlock decline rate following infestation can be accounted for. This is also consistent with previous fertilization and foliar chemistry studies, which identify a palatability-based relationship between foliar nitrogen and HWA population levels.

The significant improvement in model accuracy with the inclusion of chemical data highlights the value of hyperspectral data-derived coverages in risk modeling. In addition to improved predictive accuracy, hyperspectral imagery can provide spatially continuous maps of host species abundance and detailed decline assessments for model validation. This will allow land managers to better locate the host resource, identify stands to target management activities, and monitor forest health.

It is likely that the inclusion of other organic compounds, such as phenolics or other defensive chemicals would further improve this model (Bi and others 1997, Zucker and others 1982). However, the ability to use remote sensing platforms to assess secondary compound concentration has not been attempted to date. Other factors such as duration of infestation, climatic variables, and mineral nutrition likely interact, and these factors may exhibit different influence under different situations (Pontius and others 2006). In areas with available hyperspectral imagery, digital soil maps, and climate data, more complex models may soon be available to land managers. The addition of such data layers, which are not typically available for risk modeling, can be incorporated for more detailed and accurate risk maps. This type of spatially continuous information could be used by integrated pest management plans to help target specific areas on the ground where management efforts may be most effective.

This work will continue to be validated and improved in our future research efforts. By tracking infestation as it progresses through these stands and monitoring changes in hemlock health on newly infested plots, we will be better able to test the accuracy of this model. Finalized coverages of foliar nitrogen concentration will also be added to this model so that parameter estimates can be used to predict actual rates of decline instead of relative vulnerability to HWA.