Field Campaign - September 2008
The area of interest is Baie-Comeau, Quebec, Canada because of historically documented Spruce Budworm infestations, available provincial forest inventory data, and aerial sketch mapping defoliation vectors (Figure 3). Plots were pre-selected based on data gathered from the available data sources, and targeted 2 stand-types (conifer and mixed), 3 age-classes, and 3 defoliation severity classes (Light 10-35%, Moderate 35-70%, and Severe >70%).
At each plot center a GPS point was taken using differential GPS, and a prism sweep was used to determine which trees were to be measured. The species was noted, height and diameter (DBH) of each tree was measured, and an ocular evaluation of current and cumulative defoliation recorded. Five trees were selected randomly for clipping branch samples (for counting egg masses) and increment coring (for age determination). Ancillary plot information include crown closure, percent cover of understory vegetation, and hemispherical and digital photos.
Imagery Processing
Imagery over the study area was acquired for Landsat 5, and SPOT was programmed for late summer 2008 (Table 1).
All imagery was corrected for top of atmospheric reflectance (accounting for sun angle) and orthorectified using the CDED 1:250,000 DEM (Figure 4). The SPOT imagery was resampled to 30m to enable comparison to Landsat. The imagery was then used to create a pseudo-invariant mask from which to derive the linear normalization equations. Pseudo-invariant pixels were selected using a thresholding technique over healthy forested areas, developed by the Canada Center for Remote Sensing. From these pixels, a linear regression equation was derived between T1 and T2 for each multispectral channel common to all sensors (i.e. green, red, near-infrared, and short-waved infrared). The regression equations were then applied to T2 and visually inspected prior to band difference and index calculation. When the validity of the normalization process was confirmed (i.e. through visual analysis and data exploration) vegetation indices were computed, and band and index differences and relative differences were calculated.
The plot GPS location was registered into the imagery and a 3x3 pixel window around plot center was created for sampling the resulting band values and calculations. The image channels selected for analysis include: green, red, near-infrared, short-waved infrared, NDVI, ISR, T1-T2 green difference, T1-T2 red difference, T1-T2 near-infrared difference, T1-T2 short-waved infrared difference, T1-T2 NDVI difference, T1-T2 ISR difference, T1-T2 green relative difference, T1-T2 red relative difference, T1-T2 near-infrared relative difference, T1-T2 short-waved infrared relative difference, T1-T2 NDVI relative difference, and T1-T2 ISR relative difference. The values for every pixel falling under the plot window was output and related back to the plot for comparing to defoliation estimates. Averages were taken of all defoliation estimates and image band values and summarized into tables.
Statistical Analyses
The proposed method to determine if defoliation severity can be detected using multi-date, multispectral satellite imagery, is to use the CANDISC procedure in SAS. Field plots were classified into 3 severity classes as defined by Quebec provincial defoliation ranks. Nil, which are plots having less than 10% cumulative defoliation, light plots have between 10 and 35% defoliation, moderate plots have 35 to 70% defoliation and, severe plots are over 70% defoliated. Because only 1 plot was considered Nil, it was removed from the analyses. Six datasets were used which were subsets of the original field data. The first dataset contained all plots. The second dataset contained only conifer plots (>80% conifer dominant). The third dataset contained mixed deciduous and coniferous plots. The fourth dataset contained fir dominant plots (>80% fir). The fifth dataset contained spruce dominant plots (>80% spruce), and the sixth dataset containing conifer-mixed plots (where individual conifer species occur <80%). The input predictor variables include: all band and vegetation indice differences (i.e. T1 - T2 for R(red), G(green), NIR(near-infrared), SWIR(short-waved infrared), NDVI(normalized difference vegetation index), and ISR(infrared simple ratio)), and relative differences (i.e.(T1-T2)/T1).
The data was first explored by producing multiple scatter plots between the predictor and response variables. This aided in the identification of outliers, and assessing the distribution of the data. Once normality was ascertained (or not rejected as in this case) and the outliers removed, the following analyses was conducted.
Multiple output datasets were specified, but a generalized sample of the input syntax is as follows:
PROC CANDISC DATA=input OUT=output;
CLASS DEFOLIATION_CLASS;
VAR T1TXG T1TXR T1TXNIR T1TXSWIR T1TXNDVI T1TXISR T1TXNDVIrd T1TXISRrd T1TXGrd T1TXRrd T1TXNIRrd T1TXSWIRrd;
RUN;
The output results were analyzed to determine if:
- there are significant predictors of the defoliation severity classes
- the influential predictor variables change between datasets
- there are redundant predictor variables, and if so which are best to keep.
Depending on the dimensionality of the data, and the discriminatory power of the predictor variables, different methods may need to be employed to derive an equation for modelling defoliation severity across the T2 image.
The GLM (MANOVA) procedure was determined to be the optimal method of determining statistically significant predictor variables in the case of two-dimensional data. Excessively redundant predictor variables (i.e. highly correlated such as between the band difference and band relative difference) were removed from the datasets before running this procedure. Below represents the syntax used:
PROC GLM DATA=input;
MODEL T1TXNDVIrd T1TXISRrd T1TXGrd T1TXRrd T1TXNIRrd T1TXSWIRrd = DEFOL;
MANOVA H=_ALL_;
RUN;
Once significant variables were determined for each dataset from the GLM output, a multiple regression could be run using the continuous cumulative defoliation variable against the significant predictor variables. The GUI SAS interface was used to run a multiple linear regression with: cumulative defoliation as the dependent variable, and the significant predictor variables as the explanatory variables.
Forest cover classification can be possible using the EOSD (Earth Observation for Sustainable Development of Forests) landcover product, therefore seperate models for different forest cover types can be applied to the T2 imagery. Significant multiple regression models were applied to their respective cover classes (i.e. Conifer and other) resulting in a image map of cumulative defoliation.
To assess the adequacy of the model residuals can be compared between the input cumulative defoliation, and the defoliation values resulting from the model.