Getting our ducks in a row, thousands of years later

Ducks (Anatidae) are a speciose and ecologically diverse family, thus their fossil remains hold significant palaeoecological value. However, a major challenge is their conservative morphology, which hampers specific identification. An underexplored approach is identification to an ecologically-informative level, which is not necessarily the species level. Also, recent advances in Machine Learning (particularly Convolutional Neural Networks) could expand the toolkit in the identification process. Here, we use these approaches to investigate the Quaternary duck diversity of Doggerland. Using private and museum collections, we identified 304 specimens that were retrieved from Dutch beaches after sand nourishments from the southern part of the North Sea. This resulted in at least 15 identified species, of which 12 had not previously been confirmed in the Dutch fossil record. Mergini (marine ducks) are far more abundant than any other. Notably, long-tailed ducks Clangula hyemalis (Linnaeus, 1758) dominate despite currently being rare in the area, whereas the now-abundant common eider Somateria mollissima (Linnaeus, 1758) is absent entirely, which is congruent with a previous hypothesis on post-glacial expansion of the eider. The training of several Convolutional Neural Networks shows that artificial intelligence image recognition has the potential to aid in duck bone identification, but the models are not consistent enough yet. Despite being difficult to identify, fossil duck bones are valuable remains that hold significant information that has applications in fields ranging from palaeontology to conservation ecology. Our approaches could well be applied to other fossil assemblages.