Reproducibility of species recognition in modern planktonic foraminifera and its implications for analyses of community structure

Applications of planktonic foraminifera in Quaternary palaeoceanographic and palaeobiological studies require consistency in species identification. Yet the degree of taxonomic consistency among the practitioners and the effects of any potential deviations on community structure metrics have never been quantitatively assessed. Here we present the results of an experiment in taxonomic consistency involving 21 researchers representing a range of experience and taxonomic schools from around the world. Participants were asked to identify the same two sets of 300 specimens from a modern subtropical North Atlantic sample, one sieved at > 125 µm and one at > 150 µm. The identification was carried out either on actual specimens (slide test) or their digital images (digital test). The specimens were fixed so the identifications could be directly compared. In all tests, only between one-quarter and one-eighth of the specimens achieved absolute agreement. Therefore, the identifications across the participants were used to determine a consensus ID for each specimen. Since no strict consensus ( > 50 % agreement) could be achieved for 20–30 % of the specimens, we used a soft consensus based on the most common identification. The average percentage agreement relative to the consensus of the slide test was 77 % in the > 150 µm and 69 % in the > 125 µm test. These values were 7 % lower for the digital analyses. We find that taxonomic consistency is enhanced when researchers have been trained within a taxonomic school and when they regularly perform community analyses. There is an almost negligible effect of taxonomic inconsistency on sea surface temperature estimates based on transfer function conversion of the census counts, indicating the temperature signal in foraminiferal assemblages is correctly represented even if only two-thirds of the assemblage is consistently identified. The same does not apply to measures of diversity and community structure within the assemblage, and here we advise caution in using compound datasets for such studies. The decrease in the level of consistency when specimens are identified from digital images is significant and species-specific, with implications for the development of training sets for automated identification systems.