Conclusions and Future Work
The inital null hypothesis that defoliation severity classes can be discriminated using linear combinations of spectral bands, was modified by modelling the continuous variable directly. The new null hypothesis that cumulative defoliation can not be modelled from multispectral satellite imagery is rejected. There are significant relationships between cumulative defoliation and its spectral response in different wavelengths (or combinations there of). Defoliation can be confidently detected over all potentially affected stand types (i.e. conifer and mixed), with better predictability occuring over homogenous stands. Modelling cumulative defoliation directly provides a more flexible approach to classifying severity by allowing the user to determine the number of classes suitable to his/her needs.
From a scientific standpoint, it would be desirable to refine the model by stand compostion such as species dominance (i.e. fir, spruce, mixed conifer, and mixed forest), and across age classes. This may provide a better understanding of the patterns of spruce budworm infestations, and allow for a better estimate of impact for carbon accounting puposes and/or when assessing volume loss of merchantable stands. However, taking this approach may defeat the purpose of using remote sensing as a feasible means to assess cumulative defoliation over large areas (i.e. on a provincial or national scale). The ability to distinguish species dominance and/or age class from a multispectral image requires high resolution both spatially and spectrally, and would negate the overall purpose: to use remote sensing to assess cumulative defoliation over large scales in a cost-effective and timely manner. It is optimal therefore, to derive models that do not require external datasets. It would be useful to perform this analyses using only freely available imagery such as Landsat. Pre-defoliation images can be used to classify forest into broad categories, such as conifer, deciduous and mixed stands, which in turn can be used to apply the cumulative defoliation models to their respective stand types.
Although the results found to date are encouraging, there is much left to be done. The data needs to be revisited and tested for normality, to determine what types of transformations are required to provide a more accurate model. To model severity across stand types requires more field sampling of mixed stands across severity levels. A more rigourous method of selecting input variables is required; different methods (i.e. MANOVA, Pearson's correlation, multiple regression) identified either different or varying degrees of significant variables. It would also be useful to compare to within-platform multi-date images (i.e. Landsat). Perhaps the variability of the results is in part attributable to the cross-platform analysis.
In conlcusion, multi-date, multispectral satellite has proven useful for mapping defoliation severity across landscapes.