APPLICATION OF AN AUTOMATED PALYNOLOGY SYSTEM TO QUATERNARY FOSSIL POLLEN ANALYSIS: PROGRESS AND PITFALLS (Invited Presentation)

The automation of pollen identification and counting has the potential to deliver benefits to virtually all areas of palynology. Of these areas, analysis of fossil pollen perhaps stands to gain the most through savings in time spent at the microscope counting pollen samples. Yet out of all the potential palynological sub-disciplines fossil pollen analysis perhaps poses some of the greatest challenges to automated palynology. Broken, deformed, and clumped pollen are common in Quaternary fossil pollen samples, as is non-pollen debris. These factors make image-processing based recognition and classification of fossil pollen inherently problematic, and to date the majority of published automated palynology systems have not attempted counting and classification of fossil material, focusing instead on fresh or live pollen.

Here we report the results of preliminary tests which apply an automated palynology system to examples of Quaternary fossil pollen samples from New Zealand. This system (known as ‘Classifynder’) employs robotics and image processing to locate and image pollen on slides, and is coupled with a neural network-based classifier to identify the pollen in the captured images. Slides of pollen processed using standard pollen extraction procedures were counted repeatedly by experienced palynologists and by the Classifynder. Raw counts from the Classifynder were scrutinized by a human palynologist, with any incorrectly classified images reassigned to their correct taxon, and debris images deleted. The degree of similarity/difference between the human and user-adjusted Classifynder counts is used as the basis for assessing the accuracy of the Classifynder system at counting fossil pollen. Results demonstrate that the accuracy of the neural network-based classifier can be quite variable, caused partly by misclassification of deformed or broken grains. However final Classifynder counts of the fossil samples matched very closely with the human counts. This indicates that although the system is misclassifying some pollen grains, it is still recognizing the majority of pollen objects on the slide, and with checking by a human palynologist, can produce counts of fossil pollen to the same degree of accuracy as a human palynologist, but with much less time and effort required from the palynologist.