Spruce Budworm
Spruce budworm (SBW) is widespread across Canada with Choristoneura fumiferana being the most prevalent. SBW is a native defoliator and is part of the natural cycle of spruce-fir forests 2. Outbreaks typically last 10 years and repeat every 35 years on average 3. Adults can be dispersed hundreds of kilometres depending on prevailing wind velocity potentially becoming a widespread infestation.
Mating usually occurs late July, and females begin to lay eggs on the needle undersurfaces of host trees 2. Although SBW prefer Balsam Fir and Spruce species, they will colonize other conifer species such as hemlock and larch when infestations are severe. Larva emerge from the egg masses late summer and overwinter during their second larval stage 4. Feeding usually occurs late spring and early summer 1. The larva can kill most of new needle growth, and sometimes all of the new buds and shoots 2.
Current defoliation appears reddish due to foliage senescence and typically leads to foliage loss. Recovery generally occurs the following growing season unless the infestation is severe for 3 or more years. Severe outbreaks that last longer than 3 years (i.e. cumulative defoliation), may lead to dead tops, stunting, or mortality of infested trees 3.
STANDARD DEFOLIATION SURVEYS
In Canada, defoliation has typically been assessed through aerial sketch mapping. Generally, an area of infestation is targeted and the aerial survey carried out by regional jurisdictions. In these surveys the observer estimates defoliation severity over a given stand and delineates the extent of the infestation. Errors may arise through geolocation inaccuracies, flight attitude variability, subjective and/or inconsistent severity calls, and between regional surveys. Because of these potential sources of error, and the costly, time-consuming nature of aerial surveys, cost-effective estimates of defoliation impact are required on a national scale.
Multispectral Satellite Imagery
Multispectral satellite imagery (SPOT, Landsat TM) captures the reflectance of the Eath's surface in different parts of the electromagnetic spectrum. The sensitivity of red, near-infrared and short-waved infrared wavelengths to chlorophyll and water content can aid in identifying changes in vegetation phenology (Figure 1). Vegetation indices, which compare the spectral responses of vegetation in different parts of the spectrum, are often used to optimize the extraction of biophysical parameters. The normalized difference vegetation index (NDVI) and the infrared simple ratio (ISR) are indices that capitalize on properties such as chlorophyll and water content, and can be used to identify areas experiencing vegetation stress. The spatial, spectral and temporal coverage of Landsat and SPOT at high resolutions facilitate the analysis of forest stand characteristics, and may prove useful in detecting and mapping the extent of spruce budworm defoliation.
The reflective properties of vegetation is highly variable; spatially (i.e. local vs. regional), temporally (i.e. inter vs. intra-annual), structurally (i.e. dense vs. open canopy and/or understory), and radiometrically (i.e. sensor differences between the times of image acquisition). Spectral variability presents limitations in the confident detection of change in forest health. It is essential that temporal images are acquired at similar phenological stages and spatial extents, and that radiometric differences between dates are normalized. The normalization of post-defoliation to pre-defoliation images is a critical component in detecting change using satellite imagery.
MULTISPECTRAL SENSORS
Landsat 5 Thematic Mapper (TM) is a multispectral sensor following a polar, sun-synchronous orbit approximately 705km above the Earths surface. It orbits Earth roughly 14.5 times a day with repeat coverage of the same area every 16 days. the sensor has a spatial resolution of 30m and captures Earth observation images in the blue, green, red, near-infrared, short-wave infrared, and long-wave infrared bandwidths. Landsat 5 has been in operation since 1984, thus can provide historical earth observation data 5.
SPOT 4 is a programmable multispectral sensor allowing the programmer to obtain coverage at the time and location of interest. The SPOT 4 sensor captures data in the green, red, near-infrared and short-waved infrared wavelengths at a 20m spatial resolution. SPOT 4 also has the ability to obtain panchromatic data at a 10m resolution.
Statistical Methods
DISCRIMINANT ANALYSIS
Canonical discriminant analysis is a dimension-reducing technique that maximizes the separability of input classes based on linear combinations of the input predictor variables 11. The resulting canonical discriminant functions summarize the between class variation, which can be used to distinguish classes, and/or predict class membership of unclassified data. This method assumes that within group class is distributed multivariate normal and functions are derived using squared Mahalanobis distances. The first resulting canonical discriminant represents the maximum correlation of the variables and the groups. The second canonical discriminant represents the maximum correlation of the variables and the groups that is uncorrelated to the first discriminant. Discriminants are produced until most of the variation within the data is explained. The correlation coefficients between the predictor variables can be plotted for each canonical discriminant for each class, to identify suitable predictor variables and class separability. Erroneous results may be produced when within class distrubution violates the normal distribution and/or sample size is too small.
MANOVA
Manova is used to understand the effects a response variable has on multiple predictor variables. MANOVA tests the differences in the centroid means of the multiple predictors, for various categories of the iresponse 11. The GLM procedure in SAS was used because it can handle unnormal between-class distributions. It has been documented that Manova can be used to identify the predictor variables which differentiate a set of response variables the most 12. In this case it was used to determine which predictor variables can explain a significant proportion of variation in the response variable (i.e. cumulative defoliation). The resulting F-tests and probabilities can be used to determine which predictor variables can be used to drive a multiple regression equation between predictor variables and the response variable.
MULTIPLE REGRESSION
Multiple regression works like linear regression except with multiple predictor variables and takes on the form:
Y = a + b1*X1 + b2*X2 + ... + bn*Xn
where each subscript denotes a separate predictor variable. This equation illustrates the contribution that each predictor variable has in explaining the response variable. However, the X and Y variables are correlated after controlling for all predictor variables. This method assumes a linear relationship between predictor and response variables, that variables follow a normal distribution, that there are more observations than predictor variables, and that outliers and redundant variables are identified and removed 13.
These methods will be used as a first attempt at modelling the dependent variable from the predictor variables.