Use Of GIS To Assess The Risk Of Swiss Needle Cast Disease Of Coastal Douglas-Fir In Oregon

Rosso, Pablo H.*, Hansen, Everett M.*, and Kanaskie, Alan**.

* Dept. Botany and Plant Pathology, Oregon State University, Cordley Hall 2082, , Corvallis, OR 97331. e-mail: rossop@bcc.orst.edu.
** Oregon Department of Forestry, 2600 State st., Salem, OR 97310.

Swiss needle cast of Douglas-fir, a fungal disease, is producing severe defoliation and growth reduction in forests and plantations along the coastal area of Oregon. Presently, planting of tree species other than Douglas-fir in highly susceptible stands seems to be the only disease management option. A GIS-based predictive model is being built in order to understand important aspects of the ecology of the disease and to determine the areas of higher disease risk.

A ground-based survey of the disease (about 200 plots) was done to generate the dependent variable. GIS was used to obtain, adapt and incorporate climatic and stand history independent variables. These variables included: temperature, precipitation , fog-low cloud occurrence, past stand composition, etc. Topographic characteristics of the stands measured in situ were used to account for local-scale variability. Variables were used directly, combined or modified to represent other factors such as ambient vapor pressure deficit, solar radiation incidence, etc. After a series of exploratory analyses, variables were incorporated into a multiple regression model. Preliminary results suggest a stronger association of temperature and precipitation with the disease. Locally, the stand slope aspect seems to influence the distribution of the disease. Whether or not Douglas-fir dominated the previous stand composition was also found to be correlated with low and high disease severity, respectively. Results are in accordance with recent findings about the physiology of the disease and the biology of the fungus, which demonstrates the utility of GIS on large-scale epidemiological modeling.

Swiss needle cast disease of Douglas-fir has been expanding and intensifying in severity since the first stands showing symptoms called the attention of forest managers about 15 years ago. The causes of this out-brake are not clear, considering that its causal agent, the fungus Phaeocryptopus gaeumaniae, has been always present in Douglas-fir forests and plantations. Results of the research presented here suggest that the cause could be found in environmental changes or in the effect of relatively recent forest practices. This model also provides an insight on the influence of the environment on the disease, which can guide not only the decision-making process of forest managers, but also future research on microclimatology of the disease and biology of the causal agent. The final model will be transferred to a GIS environment to automatically produce disease risk maps.

Swiss needle cast of Douglas-fir has been know for decades as a disease that affects Christmas tree plantations or forests outside of Douglas-fir's natural habitat (Chastagner 1996, Hood 1996). However, since the early '90s an extensive and continuing disease outbreak started to be noticed by growers in coastal Washington and Oregon. Recent aerial and ground surveys show a regional expansion of symptoms (now from southern Oregon to Puget Sound, Washington) and an increase in disease severity, indicated by crown chlorosis an defoliation, in some areas (for more information see: SNC website ).

The question of why Swiss needle cast (SNC) is expanding in the Pacific Northwest has a complex answer. Since its causal agent, the fungus Phaeocryptopus gaeumanii is known to be native to Douglas-fir forests and widespread, the question has implicit the fact that something has changed in the normal Douglas-fir-Phaeocryptopus interaction that resulted in disease. This change could have been genetic and/or environmental. But in both cases the environment is a fundamental aspect of the disease development. The ecology of the disease, or epidemiology, has the role of conceptually placing the host and the pathogen within the framework of the environment. Having this framework not only helps in understanding the cause of the disease, it makes the disease more predictable, and hence, manageable.

SNC occurs in some areas and not in others. It is also evident that the sole presence of the pathogen does not determine the presence of the disease, because the pathogen is everywhere. In terms of understanding this, we could ask: "where is SNC now?" and "where next?". These questions can be re-formulated in terms of "what are the risks of SNC occurring in a certain area?"

The answer can be approached through the construction of a model that: a) understands the conditions for disease development and expression, and, b) simulates these conditions and assigns a level of risk at any point in space.

The question about the risks of SNC is a question posed at a regional scale. The presence of symptomatic trees does not seem to depend on the location of another tree in the same stand, or on the proximity of other stands containing symptomatic trees, etc. The disease seems to affect whole areas that are not necessarily interconnected.

Since the infection of a needle by a fungus is a process that occurs at a microscopic scale, and this process determines the presence of the disease at the regional scale, a risk assessment model needs to understand the connection between several scales. There are two ways of doing that. One is the "bottom-up" approach, which consists of a gradual scaling up (from needle to branch, to tree, to stand, etc.) of the factors that allow the infection process to occur. The problem with this approach is that it needs a detailed understanding of the microscopic process, and how to translate this process from scale to scale. This approach cannot be used at the present level of knowledge of SNC in the Pacific Northwest.

The "top-down" approach starts from factors operating at the regional or landscape scale and tries to establish a connection between these factors and the "end product" of the infection process, the severity of the disease in a stand. Although this approach does not take primarily into account the mechanics of the infection process, it is more realistic in the sense that it does not need (at least, initially) a detailed knowledge, which is often not available. The "bottom-up" and the "top-down" approaches are not mutually exclusive, and it is possible to combine them as new knowledge of the processes involved is acquired.

The objective of the present study is to build a model to represent the conditions under which SNC develops and expresses itself. Given the available information, a "top-down" approach was chosen, using environmental variables at the regional, landscape and stand scale. GIS was the basic tool used for preparing, combining and manipulating all the information.

Research area:

The area of study, the northern coast region of Oregon, extends from the city of Astoria to the north (46°10’ lat. N), to Coos Bay to the south (43°25’ lat. N); and from the coast inland to a line that coincides with Hwy 99 W, at approximately 123°10’ long. W. The size of this area is approximately 4,500 mi² (11,500 km²). The Oregon coast topography is determined by the Coast Range, a moderately high range with most ridge tops within 450-750 m of altitude. Steep slopes and soils with relatively little development characterize the area. A wet and mild climate allows for the growth of dense and productive conifer forests, among which Douglas-fir (Pseudotsuga menziesii), has been extensively harvested and re-planted by the timber industry.

Disease survey (Table 1):

The stand SNC disease severity variable was based on a 1998 SNC disease survey. A total of 220 10-30 yr-old stands in the Oregon coastal area were selected according to accessibility and with no other evident bias. Each stands was surveyed and a SNC rating value from 1 to 6 was assigned. Each value of the SNC rating system represents a composite of the stand’s average tree foliar retention, height growth of the last main stem internodes and the degree of crown discoloration. "1" represents a healthy looking stand, and "6" a highly symptomatic stand.

The central point of each stands was GPS located and the resulting point coverage was displayed on a Landsat TM image to assess the accuracy of the location by comparison to detailed stand maps and aerial photos.

Independent variables (Table 1):

Temperature and precipitation data were obtained from raster-based outputs from the climate prediction model PRISM (Daly et al. 1994). Grids have an approximate cell resolution of 4x4 km (2.5 arc-minute). Temperature and precipitation grids of the months considered more relevant from the epidemiological point of view, January, April-July and November (Capitano 1999), were selected from the 5 most recent available years (1989-1993). Each of the 6 months was averaged over the 5 years. Ambient vapor pressure deficit (VPD), a measure of the water content of the atmosphere, was derived from temperature rasters following Monteith and Unsworth (1990).

Degree-days base 0°C were calculated using an interpolation model that uses PRISM climate grids as inputs (Degree-Day Calculator ). Total radiation was calculated using SolarImg, a solar input model (Harmon and Marks 1995), that uses geographic location and topography data.

Fog/low cloud occurrence was estimated using daily images obtained from GOES satellite (NOAA). This satellite produces an 8-km resolution raster image in which low clouds and high clouds are represented by different pixel values. Images from1999 were downloaded from the Internet and processed to obtain a cumulative fog occurrence value for the spring-summer and winter-fall periods.

Topographic variables of the stands were both measured in situ and obtained from D.E.M.s. A 60-m resolution D.E.M. was "degraded" to 90- and 140-m resolution D.E.M.s. Slope aspect and inclination were calculated in ERDAS-Imagine for each of the 3 D.E.M.s. Slope aspect was converted from azimuth angle values to 8 cardinal directions: N, NE, E, SE, S, SW, W, NW, and finally made into a binary variable grouping SE, S, SW and W on one side, and NW, N, NE and E, on the other side.

Stand slope position was obtained from field measurements. This variable describes the location of the stand with respect to the bottom of the slope, and it has 4 possible values: low, medium, high or mountain ridge/top. Stand characteristics, age, site index, tree density and seed source were obtained from stand records. Shortest distance to the coast was measured on a map.

The 1936 stand vegetation variables were extracted from a polygon coverage produced by the Forest Service based on a map by H.J. Andrews and Cowlins (US Forest Service 1936). Coverage was transformed to raster and a 1936 vegetation category was assigned to each stand. Stands were classified according to whether or not they had Douglas-fir or Spruce-hemlock as the dominant species in 1936.

Soil variables were obtained from the State Soil Geographic Data Base (USDA). Stands were geographically matched with the corresponding map units, and the dominant soil types (sequum number) were selected from each map unit. The following soil characteristics were obtained: upper layer depth, calcium carbonate contents (% by weight), cation exchange capacity, organic matter (% by weight), particle size (sandy, loamy, silty, etc.) and order. Calcium carbonate and cation exchange capacity were later discarded due to lack of variability and missing information.

Variables in raster format were overlaid on the SNC survey point coverage and an ASCII table was generated using the "Convert Pixels to ASCII" function in ERDAS-Imagine v. 8.3. The statistical analysis was performed in SAS v. 8.

Analysis:

Before the regression analysis was carried out, the total number of original variables (64) was reduced to a manageable group. Climate variables corresponding to months within a group (ex.: maximum temperature) were correlated by pairs and only one variable of each highly correlated subgroup was taken for further analysis. Other variables, such as site index were excluded a priori because of excessive number of missing values. Categorical variables that could not be reduced to binary (ex.: soil order) were also excluded because of the high numbers of indicator variables to be created if they had to be included.

The resulting variables were screened for fulfillment of the regression assumptions (Ramsey and Schafer 1997). Correlation coefficients were calculated to check for high correlation between any two variables. A PCA analysis was done to check for multivariate correlation (Tabachnik and Fidell 1989). Normality assumption was also tested and variables (including the response variable) were transformed as needed.

For the selection of the best regression model, two variable selection techniques called "Cp statistic" and "BIC" were used (Ramsey and Schafer 1997). These techniques assign the best value to those regression models that better explain the response variable with the least number of independent variables. In the search for the best model, quadratic terms and interaction effects were also investigated.

Once a set of candidate variables was obtained, a series of regression analyses were used to include or exclude individual variables that were of special interest. The contribution of these individual variables to the general model was assessed with the corresponding F-statistic. A final model refinement was attempted by plotting the residuals against variables that could not be used at the time due to excessive missing values.

The best model found was:

Mean { log(SNC rating +1)} =

The resulting model is a good indicator of the kind of factors that may be related to Swiss needle cast disease expression. According to the model, there is a strong association between the disease and the climate and topography of the stands. Keeping in mind that higher values of the response correspond to higher disease severity, then, higher November minimum temperatures seem to be associated with higher disease levels, whereas higher June degree days (or temperatures) appear to decrease the disease severity. Higher precipitation in July is related to higher disease levels. These results appear to correspond with basic biological principles. Higher temperatures in colder months may favor the development of organisms, such as Phaeocryptopus gaeumannii. In contrast, high temperatures in summer months could be detrimental to the fungus and the host, especially when it is combined with dry periods. From the point of view of fungal growth, this also explains why higher precipitation and fog in a dry season (or month) could be favoring the disease development (Hood 1996).

These climate variables have a very coarse resolution (cells sizes of more than 1 km), and although they can determine a general trend, they have to be complemented with more detailed information, probably at the stand scale or smaller, in order to be related to processes that show high variability at lower scales such as fungal growth, canopy transpiration, infection, etc. Quite possibly topography would adequately represent this scale.

The fact that the west-southeast facing stands are associated with higher disease severity may be related to the fact that surfaces facing south and (to some degree) west tend to receive more radiation because they are more directly exposed to the sun than the surfaces facing north and east (Jones 1992, Rosenberg et al. 1983). However, this is contradicted by the fact that solar radiation did not show a significant influence.

Although not included in the final model due to marginal statistical significance (at p=0.05), the 1936 vegetation composition showed some effect on disease distribution. Particularly, stands that were converted to Douglas-fir after 1936 seemed to be more prone to have higher disease severity. This suggests that relatively recent some forest practices might be producing better conditions for disease development.

Although in this discussion some attempts were made to make biological sense out of the regression results, interpretations should be taken with caution. As mentioned before, possible effects and mechanisms involved in disease expression may be numerous and interactions can be complex. A mechanistic, complete explanation of why these variables may be associated with the disease is not an objective of this research. The purpose of this model is to sort out the main variables that influence the disease expression. Once these variables have been found, then, the generation of likely hypotheses follows.

Table 3 presents a theoretical summary of all the variables that could potentially be involved in the disease expression, and some suggestions of how these variables could be represented or estimated. Many of them have been some way or another considered in this study. It is arguable, however, that indirect assessment of some variables might produce the same effect than using the variables directly. For example, wind could be indirectly assessed using topography and stand density, but it that does not mean that the effect of wind is being clearly taken into account. Mainly because it is impossible to sort out other effects of topography on disease expression.

This problem can be better addressed with a bottom-up approach, that is, trying to model the components of the model first, and then putting them all together to predict the final outcome. In the case of the wind (supposing wind is an important factor in disease development) it would mean to model wind occurrence first and then incorporate it to a more comprehensive model that predicts disease development and expression. Unfortunately, for this kind of approach more knowledge about the biology of the disease is needed. The regression model presented in this paper could have a key role in the design of a future process-based model, because it points at the main factors that may be influencing disease development.

This regression model can be considered a static GIS model (Brady and Whysong 1999, Johnston 1998). Time is a dimension that has not been so far addressed and that also represents a potential source of model inaccuracy. There are at least three main factors that may make time relevant in the case of SNC disease development: changes in inoculum load (for example, fungal spore density in air), changes in climate from year to year, and genetic changes in both the host and/or the fungus. The model, so far has been built assuming that none of these factors exists.

In terms of inoculum load, it has been the assumption that since the fungus is always present everywhere, local changes in fungal abundance do not substantially impact regional distribution of inocula. Climate is known to change at different time scales (years, decades, centuries, etc.) and any model that depends on climate variables has to be adjusted accordingly. This means that a SNC model should be continuously tested against these changes and, ideally, these changes have to be incorporated in the model itself. Genetic changes are also assumed not to occur during the development of the model, although this a factor that has to be addressed one way or another.

Despite its static nature, the regression model may give some clue of where to look for when dynamics aspects of the disease are to be incorporated. For example, the model suggests that ambient moisture in certain time of the year might be associated with disease expression. A likely explanation to the increasing disease severity is in the recent years, then, could be a measurable increase of precipitation or fog occurrence in that period. In fact, Taylor (1999) suggests that 1996 might be the starting point of a "wet climate" period in Oregon.

Another main objective of this work is to obtain a model to estimate the probability of a certain area or stand to express SNC symptoms. The r-square of the regression (Table 2) indicates that approximately 40% of the variability observed cannot be explained with the model. 60% of the variability explained by the model is a good percentage for biological systems, in which typically, the complexity and uncertainties are too high to be captured by a simple set of variables.

To assess the predictability of the model, the observed SNC disease values were compared with the predicted values of the regression model (Figure 1). Points in Fig. 1 were "jittered" around each SNC rating class to facilitate their visualization. The short horizontal red lines represent the average predicted values within each SNC rating class. Even when the model did not show a one-to-one relationship between true and predicted ratings, it seems to represent the trend in which higher SNC rating values correspond to higher predicted values (Figure 1). When predicted values of each SNC rating class were averaged, the means of the classes 2 and 3 would coincide with a one-to-one diagonal, class 1 would be slightly above, and 4 and 5, markedly below. This displacement of the predictions and the vertical dispersion of the points around each average line indicates the non-explained variability already mentioned above.

The SNC disease regression model, seems capable of predicting whether or not rather extreme rating values, 1 or 4-5, are likely to occur at a certain location. Although, its accuracy may not be high enough for the development of an adequate disease management strategy, it can provide general guidelines for what to expect in extreme conditions. The application of alternative statistical analysis and model validation techniques are currently being carried out. The final product will include a Swiss needle cast risk prediction map.

Conclusions:

• November temperatures, July precipitation and fog, June degree-days, and slope aspect, are the variables that showed closest association with SNC disease expression.
• The regression-based model accounts for 60% of the disease variability present in the stands.
• The resulting model can be used as a starting point for future disease models and research.
• The model also represents an important step towards a predictive model for disease management purposes.

- Brady, W.W. and Whysong, G.L. 1999. Ch.7: Modeling. In: Morain, S., ed. Gis Solutions in Natural Resource Management. Onword Press, Santa Fe, NM. 364 pp.

- Capitano, B. 1999. The infection and colonization of Douglas-fir by P. gaeumannii . Oregon State University thesis, 81 pp.

- Chastagner, G.A.1996. Diseases of Christmas trees. In: Hansen and Lewis, eds. Compendium of Conifer Diseases. APS Press, St. Paul, MN.

- Daly, C., Neilson, R.P. and Phillips, D. 1994. A statistical-topographic model for mapping climatological precipitation over montainous terrain. J. Appl. Meteorology 33(2):140-158

- Harmon, M.E. and Marks, B. 1995. Programs to estimate the solar radiation for ecosystem models. http://www.fsl.orst.edu/lter/datafr.htm.

- Hood, I.A. 1996. Swiss needle cast. In: Hansen and Lewis, eds. Compendium of Conifer Diseases. APS Press, St. Paul, MN.

- Jones, H. 1992. Plants and microclimate. 2^nd ed., Cambridge University Press, 427 pp.

- Johnston, C.A. 1998. Geographic information systems in ecology. Blackwell Science. 239 pp.

- Monteith, J.L. and Unsworth, M.H. 1990. Principles of environmental physics. 2nd ed., London: Edward Arnold.

- Parker, A.J. 1982. The topographic relative moisture index: an approach to soil-moisture assessment in mountain terrain. Physical Geography, 3, 2: 160-168

- Ramsey, F.L. and Schafer, D.W. 1997. The statistical sleuth. Duxbury Press. 742 pp.

- Rosenberg, N.J.; Blad, B.L. and Verma, S.B. 1983. Microclimate. The biological environment. 2^nd ed., J. Wiley & Sons, 495 pp.

- Tabachnik, B.G. and Fidell, L.S. 1989. Using multivariate statistics. 2nd ed. Harper Collins, NY. 746 pp.

- Taylor, G. 1999. Long term wet-dry cycles in Oregon. http://www.ocs.orst.edu/reports/wet-dry.html .

- US Forest Service, 1936. Forest type map, State of Oregon. NW quarter. US Pacific NW Forest Experimental Station.

Table 2: Regression analysis

Table 3 : List of variables and related data potentially involved in disease expression.