Epsilon Open Archive

Abstract

Little is known about the paleoecological histories of the three spruce species (white spruce, Picea glauca; black spruce, P. mariana;
and red spruce P. rubens) in eastern North America, largely because of the difficulty of separating the three species in the pollen
record. We describe a novel and effective classification method of distinguishing pollen grains on the basis of quantitative analysis of
grain attributes. The method is illustrated by an analysis of a large sample of modern pollen grains (522 grains from 38 collections)
of the three Picea species, collected from the region where the three species co-occur today. For each species X we computed a binary
regression tree that classified each grain either as X or as not-X; these three determinations for each grain were then combined as
Hamming codes in an error/uncertainty detection procedure. The use of Hamming codes to link multiple binary trees for error detection
allowed identification and exclusion of problematic specimens, with correspondingly greater classification certainty among the remaining
grains. We measured 13 attributes of 419 reference grains of the three species to construct the regression trees and classified
103 other reference grains by testing. Species-specific accuracies among the reliably classified grains were 100, 77, and 76% for P.
glauca, P. mariana, and P. rubens, respectively, and 21, 30, and 22% of the grains by species, respectively, were problematic. The
method is applicable to any multi-species classification problem for which a large reference sample is available.