Assessment of ZooImage as a tool for the classification of zooplankton

Abstract

ZooImage, image analysis software, was evaluated to determine its ability to differentiate between zooplankton groups in preserved zooplankton samples collected in Prince William Sound, Alaska. A training set of 53 categories were established to train the software for automatic recognition. Using the Random forest algorithm, ZooImage identified particles in the training set with less than 13% error. Despite reasonable results with the training set, however, ZooImage was less effective when this training set was used to identify particles from field-collected zooplankton samples. When all particles were examined, ZooImage had an accuracy of 81.7% but this dropped to 63.3% when discard particles (e.g. marine snow and fibers) were removed from total particles. Copepods, the numerically dominant organisms in most samples, were examined separately and were correctly identified 67.8% of the time. Further investigation suggested size was effective in determining identifications; medium size copepods (e.g. Pseudocalanus sp., Acartia sp.) were accurately identified 73.3% of the time. ZooImage can provide a coarse level of taxonomic classification and we anticipate continued improvement to this software should further enhance automatic identification of preserved zooplankton samples.