On direct extraction of scene component fractions and crown cover distribution in open forest canopies using high spatial resolution winter imagery

Spatial patterns of understory and overstory constitute an as-yet poorly understood obstacle to the derivation of boreal forest structural variables through remotely-sensed image data. The high resolution reflectance imagery collected seasonally as part of the Boreal Eco-system and Atmosphere Study (BOREAS) presents an opportunity to increase overall understanding of this complicated interplay. A research project is currently underway to compare and contrast two methods of separating understory from overstory in this imagery, one "directly" through classification and the other "indirectly" through end-member analysis of a high spectral resolution data set. Winter imagery is investigated to assess the best method(s) of classification into crown and snow understory. Simple thresholds are considered, followed by unsupervised and supervised classification tests by variation of scene type, tree-type, number of classes and number of wavelength channels. K-Means (Iterative Optimization) Clustering reveals that less than four classes is insufficient to effectively separate the land-cover classes for both tree-types investigated and that more than four classes leads to trivial division of existing classes. Results from this unsupervised classification are used to build training areas for Maximum-Likelihood supervised classifications. JM (Bhattacharya) distances are calculated, and it is shown that three channels provide optimum separability for both tree-types. A working methodology for winter image classification is proposed.