Comparison of possible multispectral classification schemes for tree crowns individually delineated on high spatial resolution MEIS images

In the computer analysis of high spatial resolution multispectral aerial images for forest inven-tory purposes, it may be more useful to deal with individual tree crowns as the "objects " of our analysis than with forest stands or individual pixels. Starting from this tenet, it becomes important to be able to spectrally define these tree crowns as succinctly as possible. This paper proposes, describes, and compares seven different ways that tree crowns in high spatial resolution aerial images can be spectral-ly defined for species classification. In testing these seven types of multispectral signatures, it was found that five led to relatively similar classification accuracies (72 ± 3%) in differentiating five coniferous species. Additional classification accuracy improvements (to 76%) were possible with some of the signatures by using Canonical Analysis prior to classification. The "tree colour line"-based signatures, a newly introduced approach, gave comparable results to simpler signatures such as those based on the mean multispectral value of tree crowns. Red pine (Pinus resinosa Ait) crowns were consistently hard to separate spectral-ly from other species, whereas black spruce (Picea mariana (Mill.) B.S.P.) crowns were consistently easy to distinguish