Portable, non-destructive colorimetry and visible reflectance spectroscopy paired with machine learning can classify...

The objective of this study was to determine if visible reflectance spectroscopy and quantitative colorimetry represent viable approaches to classifying the heat treatment state of silcrete. Silcrete is a soil duricrust that has been used as toolstone since at least the Middle Stone Age. The ancient practice of heat treating silcrete prior to knapping is of considerable interest to paleolithic archaeologists because of its implications for early modern human complex cognition generally and the ability to manipulate the material properties of stone specifically. Here, we demonstrate that our quantitative, non-invasive, and portable approach to measuring color, used in conjunction with k-Nearest Neighbors "lazy" machine learning, is a highly promising method for heat treatment detection. Traditional, expert human analyst approaches typically rely upon subjective assessments of color and luster and comparison to experimental reference collections. This strongly visual method can prove quite accurate, but difficult to reproduce between different analysts. In this work, we measured percent reflectance for the visible spectrum (1018 variables) and standardized color values (CIEL*a*b*) in unheated and experimentally heat-treated silcrete specimens from three sources in South Africa. k-NN classification proved highly effective with both the spectroscopy and colorimetry data sets. An important innovation was using the heat treatment state predicted by the k-NN model for the majority of replicate observations of a single specimen to predict the heat treatment state for the specimen overall. When this majority voting approach was applied to the 746 individual observations in this study, associated with 94 discrete silcrete flakes, both spectroscopy and colorimetry k-NN models yielded 0% test set misclassification rates at the specimen level.