Starch granules identification and automatic classification based on an extended set of morphometric and optical measurements

Abstract

Starch granules have been found to be preserved in association with archaeological remains and their identification may provide direct botanical evidences of the plants used by ancient humans. However, subtle morphological differences between starch granules make their taxonomic identifications difficult. In order to improve the identification of these plant remains, we used an image analysis program that measures up to 123 different optical and morphological characters. With Random Forest tests we analyzed ~5000 starch granules extracted from underground storage organs (USO), seeds, and fruits of 20 different East African edible plant species. Our results show that correct identification rates are up to 74% for some species (Echinochloa colona, Cyperus rodundus), ~80% for some suprageneric taxa (Poaceae, Fabaceae), and 80% for underground storage organs. However, on average, success rates are just ~53% for species (up to 70% with a dataset reduced to herbaceous species), 60% for families, and 72% for plant parts. Yet, this automated system is not perfect, but it is still more powerful than the human eye, for which the average success rate is just of 25% for species level identifications. We evaluated the performance of our system and found that accuracy rates of identifications of starch granules are highly sensitive to the number of groups (species) to identify (r2=0.83) and, to a lesser extent to the number of characters used by the identification system (r2=0.87). It is therefore crucial to narrow down as much as possible the number of target species, by analyzing additional proxies. We conclude that better results can be achieved if the candidate field is narrowed. If not, the automated identification of starch granules will remain unsatisfactory to provide acceptable interpretations in archaeological contexts.