THE DISCRIMINATION OF MORPHOLOGICALLY SIMILAR POLLEN USING MACHINE LEARNING METHODS

Pollen is a critical data source for many paleontological fields, including paleoecology, paleoclimatology and biostratigraphy, because it is among the most ubiquitous of terrestrial fossils and often preserves an extended, continuous record of vegetation change. However, this temporal continuity has come with a taxonomic trade-off. Most pollen grains are often only identified to genus, with some morphologically homogenous clades only identified to family. Species identifications are often only possible in low-diversity Quaternary samples, where the pool of potential species is already known. Improving the taxonomic precision of pollen identifications would expand the research questions that could be addressed by the palynological record. Our proposed solution is the development of machine learning algorithms sensitive to subtle variations in pollen morphology. Our test case is the classic Quaternary example of black and white spruce.

White spruce (Picea glauca) occupies predominately drier upland soils while black spruce (Picea mariana) populates poorly drained lowlands. Changes in their relative abundance through the Quaternary reflect changing environmental conditions in North American forests. Given their paleoclimatic importance, there is an extensive literature on Picea morphology, but despite two decades of analysis, inconsistency in identification remained a persistent problem. However, our results using instance-based learning (n-nearest neighbor) algorithms demonstrate that consistent discrimination of black and white spruce is possible, with >85% accuracy in the identification of black and white spruce species, and ~100% accuracy in the discrimination of spruce from other saccate genera (Pinus, Abies). The algorithms used are robust to some folding, tearing, and compression of the grains, suggesting that these methods would be applicable to the identification of other palynomorphs in both Quaternary and pre-Quaternary studies. By identifying and developing shape and texture measures sensitive to more subtle differences in morphology, our research complements other research programs seeking to fully automate palynological analysis, while providing identification tools that should be of immediate utility to many palynologists.