Identifying patterns of plant “waste accumulation” in House 169, Elephantine Island, Egypt (1773–1650 BCE) using Machine Learning

Abstract

The exceptionally well-preserved multi-period settlement, cemeteries, and temples on Elephantine Island (Aswan, Egypt – 1st Nile cataract) have been under investigation by the German Institute for over fifty years. The location of the settlement site on an island well above the highest Nile flood level in one of the most arid regions in Africa yields near-perfect preservation of organic remains. Between 2013-2019, the “Realities of Life” project focussed excavations on House 169 which was identified as a single house unit via recognition of abandonment layers / phases separating it from earlier / later buildings. H169 dates to the mid-late 13th dynasty (latter part of 1773-1650 BCE). Because the house did not experience any kind of catastrophic destruction, and was abandoned very slowly, with new houses constructed on top, we assumed that the remains found in the “occupation” deposits would be limited to waste gathered in corners, trampled into floors, and the ashes and debris left following final use of cooking / baking installations. For this reason, we anticipated that the archaeobotanical assemblage from H169 was likely a homogenous mix of debris from household activities and plants blown / brought into the house. However, because this is largely a desiccated assemblage, we were also interested in identifying spatial differences, either between rooms, or between different types of features. To do that and following the broader aims of the project—to explore as many different archaeological methods as possible—we decided to assess the potential of machine learning techniques for identification of patterns in the data. Traditional PCA analysis indicated that the assemblage is almost totally homogeneous, with no real differences in the composition of the samples other than a few easily identified outliers. However, the results of our two-step approach in data analysis revealed some patterns. A hierarchical clustering strategy applied to the composition of the samples yielded 2 to 4 reasonably equally sized clusters among the collected samples; subsequently, evolutionary prediction trees – our additional step in Machine Learning – served to relate spatial, temporal, and functional variables to cluster membership. The fit metrics associated with the predictive trees indicate a robust tie between the site characteristics and the samples associated with each cluster. In this way, the resulting tie between spatial/temporal/functional properties of the site and sample clusters serves to achieve our stated goal of finding patterns even within a largely homogenous assemblage.