Application of Image Recognition Technology to Foraminiferal Assemblage Analyses

Analyses of foraminiferal assemblages involve time consuming microscopic assessment of sediment samples. Image recognition software, which systematically matches features within sample images against an image library, is widely used in contexts ranging from law enforcement to medical research. At present, scientific applications such as identification of specimens in plankton samples utilize flow through systems in which samples are suspended in liquid and pass through a beam of light where the images are captured using transmitted light. Identification of foraminifers generally utilizes reflected light, because most shells are relatively opaque. My goal was to design and test a protocol to directly image foraminiferal specimens using reflected light and then apply recognition software to those images. A library of high quality digital images was established by photographing foraminifers identified conventionally from sediment samples from the west Florida shelf. Recognition software, VisualSpreadsheet™ by Fluid Imaging Technologies, Inc., was then trained to improve automated assemblage counts and those results were compared to results from direct visual assessment. The auto classification feature produced composite accuracies of foraminiferal groups in the range of 60–70% compared to traditional visual identification by a researcher using a stereo microscope. Site SC34, the source of images for the original image library, had an initial accuracy of 75% that was improved slightly through an alteration to one of the software classes, but composite accuracy plateaued at 60% with the updated filters. Thus, image acquisition advancements and further development of image recognition software will be required to improve automated or semi automated foraminiferal classifications. However, other potential applications were noted. For example, an advantage of acquiring digital images of entire samples or subsamples is the ability to collect quantitative data such as diameter and length, allowing size-frequency assessments of foraminiferal populations while possibly automating grain size analyses without requiring separate processing. In addition, data files of library and sample specimens can be readily shared with other researchers