Unsupervised classification and automated shape recognition as tool for computer-assisted reproducible typology

Abstract

Since 2016, in the SNFS-project ‘Mobilities, entanglements and transformations in Neolithic societies on the Swiss Plateau (3900-3500 BC)’, we have been using a computer-aided classification of vessel shapes to investigate completely preserved Neolithic pottery of Swiss wetland sites and neighbouring regions from the 4th millennium BC beyond classical typologies and cultural assignments. Our strategy is essentially very simple: We follow a holistic approach in which the entire vessel body is understood as a rotational body (comparable with Mom 2005, Chapman et al., 2006, Keogh et al., 2009). One side of the rasterized, filled and equally scaled profile is extracted, and via a simple transfer of the image information into a matrix, the profile line is converted into 400 measuring points. The profile information obtained in this way, enriched with nominal values such as rim shape or decoration as well as other metric values such as absolute height, can be evaluated using various multivariate methods. We opted for a combination of t-sne as ordination and dimension-reducing method and HDBSCAN as cluster algorithm for this analysis. This computer-aided method was accompanied by a impressionistic classification by hand. In our analyses it became clear that both methods complement each other meaningfully: While computer-aided classification was able to work out more general, cross-cultural trends, which can be interpreted with regard to the function of the vessels and a consumer perspective, impressionistic classification leads to an identification of different styles, which rather open up a producer perspective. The juxtaposition of both methods on the same material can be used to overcome existing stereotypes and topoi in the ceramic classification. Furthermore the results also serve to identify yet unresolved shortcomings of the computer-based approach, which will be addressed in the future.