A multivariate approach to vegetation mapping of Manitoba’s Hudson Bay Lowlands

The Hudson Bay Lowlands of Manitoba contain a wide range of vegetation types that re ect local variations in climate, geological history, permafrost, Ž re, wildlife grazing and human use. This study, in Wapusk National Park and the Cape Churchill Wildlife Management Area, uses a Landsat-5 TM image mosaic to examine landscape-level vegetation classes. Field data from 600 sites were Ž rst classiŽ ed into 14 vegetation classes and three unvegetated classes. Principal component analysis was used to examine the spectral properties of these classes and identify outliers. Multiple discriminant analysis was then applied to determine the statistical signiŽ cance of the vegetation classes in spectral space. Finally, redundancy analysis was used to determine the amount of vegetation variance explained by the spectral re ectance data. We advocate this adaptive learning approach to vegetation mapping, by which the researcher employs an iterative strategy to carefully examine the relationship between ground and spectral data. This approach is labour intensive, but has the advantage of producing vegetation classes that are spectrally separable, decreasing the likelihood of errors in classiŽ cation caused by overlap between classes.