Methods
Between the spring of 2003 and the winter of 2006, a total of 307 P. alni infested and 127 healthy/noninfested alder tree locations were sampled in forested areas in Bavaria (Figure 1). Among the 307 infested sites, there were 232 points where alder trees had been planted, and 75 points where alders were naturally occurring. Of the 127 healthy sample points, 38 were planted and 89 had natural alder growth. A GIS sample point theme of the dependent variable was created containing all 434 sample point locations for analysis in the classification tree.
Thirteen independent variable raster datasets were used in the Bavarian classification tree analysis (Table: Bavarian Independent Variables). Specifically, these were 12 93-m physiographic datasets including: nine soil texture components (minimum, mean and maximum percentage values for sand, silt, and clay polygons), aspect, slope, and landform an index of concavity and convexity. The 13th dataset was the Normalized Difference Vegetation Index (NDVI) calculated from the Moderate Resolution Imaging Spectroradiometer (MODIS) satellite imagery at 250m. Numerical values were extracted from each of the independent variable grids at each of the sample locations from the sample point theme. These values were then used to compose the spatial information database that was exported to S-PLUS© statistical software (S-PLUS©, Statistical Sciences 2000) for analysis. The independent variables selected by the classification tree were: silt, sand, slope, aspects that were Southeast, South, Southwest, and West, and the landform index.
The default S-PLUS© validation technique, tenfold cross validation, was used to prune the tree to avoid overfitting the classification tree model to the Spatial Information Database. The tenfold cross validation was used, as it does not rely on an independent dataset and can identify the optimum tree size for minimizing prediction errors.
Based on the results of the classification tree analysis (Figure 2), conditional statements (CON statements; ESRI ArcView, 2000), were used to create a binary P. alni potential distribution, (ie., presence and absence) surface for the forested areas in Bavaria. The significant independent variables selected by the classification tree, as well as the decision tree rules, (e.g., threshold values taken at the tree nodes), were the input for the CON statements.
Only three of the independent variables that were selected by the Bavarian classification tree were available globally. These were slope, aspect, and the landform index (Table: Global Independent Variables). To see how the rules would change given only the three independent variables, a second classification tree was developed for Bavaria using only those three datasets. This second model had limited utility as it overpredicted the presence of P. alni, predicting more than 90 percent of the study area to have P. alni present. Still, the rules from the second model did provide some additional insight regarding the broader range of conditions within which P. alni might be present. Therefore, the rules from both the original as well as from the second model were combined, along with additional expert knowledge, in a final multicriteria model to create a global susceptibility surface for P. alni.
To develop the multicriteria susceptibility model for the globe (Figure 3), the unique numerical values from each criterion had to be standardized. Therefore, each dataset was reclassified using a hazard ranking of 0 – 10 (Table: Global criteria re-classifications to hazard rankings). The decision rules from both classification trees as well as additional expert knowledge were used as a guide in setting the hazard rankings.
Areas where alder and P. alni could not grow were eliminated from the global analysis by creating masks from climate and landcover data (Table: Global Mask). To determine temperature thresholds for the climate mask, an investigation of Alnus species was conducted. It was determined from frost hardiness and heat/drought hardiness zones for all alder species that alder does not survive temperatures +/- 40 degrees Celsius.
In addition, lab results performed by Dr. Jung indicated that soil temperatures greater than 32 degrees Celsius prevent the survival of P. alni. Although soil temperature data is not available, a regression formula (Temperature MAX threshold value = (Soil Temperature MAX threshold value - intercept estimate) / Regression Coefficient Estimate), was applied to determine that 32 degrees Celsius equates to air temperatures of 34 degrees Celsius.
Consequently, areas with temperatures less than -40 degrees Celsius and greater than +34 degrees Celsius, as well as areas that could not support alder such as tundra, bare ground, and bodies of water, were removed from further analysis. The binary climate and landcover masks were combined by multiplying both surfaces together to create a combined binary temperature and landcover mask. The resulting mask was combined again in a weighted overlay with the re-classified criteria to produce a potential global distribution.
Because slope predicted most of the variability in the classification tree, it was weighted at 50 percent; aspect and the landform index were both weighted at 25 percent (Table: Global Criteria and Arithmetic Weights).
To produce the final global susceptibility surface, areas that were identified in the global distribution as having a potential for a P. alni infestation were classified according to hazard. Biome and stream data (Table: Hazard: Global datasets used to assign hazard) were combined in an equal weighted overlay (Table: Global Masks and Arithmetic Weights) to assign a hazard ranking. The hazard ranking was based on each pixel’s occurrence within ecological biomes similar to the biomes where P. alni presently occurs, as well as its proximity to streams. Thus, pixels within the selected biomes were assigned hazard rankings based on their proximity to streams. A set of three global stream buffers at distances of 1km were used to assign the hazard rankings. Pixels that had the potential for infestation were given a high hazard ranking if they fell within 1 km of the stream. Those pixels between 1 and 2 km were assigned a medium potential hazard, and pixels between 2 and 3 km from the stream were assigned a low potential hazard. Pixels that were found greater than 3 km from a stream or outside the selected biomes were given a hazard ranking of little or no potential hazard (Table: Phytophthora alni susceptibility).
Encyclopedia ID: p3330
