Analysis of image-based classification of foraminiferal tests

Two factors govern the performance of computerized systems for classifying foraminiferal tests (forams) from optical microscopy images: (1) the amount of information captured by the images; and (2) the ability of computer algorithms to exploit the information. This paper provides an analysis of the former, to understand the efficacy and limitations of digital images in foram representation. To this end, 500 images were captured to show default (unmanipulated) and alternative views of 250 variable-grade specimens taken from 25 core samples. The specimens (mainly planktic forams) were classified by a specialist three times: first, from the default views alone; second, from the default and alternative views together; and third, from the archived particles. Image and particle-based classifications were compared using statistical methods to establish significance. Default views alone were sufficient to realize low incorrect genus and species rates and high correct genus rates, irrespective of specimen quality, view side, and image quality. Availability of alternative views had only a minimal impact on performance, at least for the statistical sample. However, image quality, which was primarily limited by depth-of-field issues, had a significant impact on performance. To facilitate this work, an online database called the Microfossil Wiki was developed. All images and classifications referred to here are accessible via the wiki.