Grain morphometric analysis on mudbrick thin sections

This repository contains the dataset, code, results, and computational workflow associated with the peer-reviewed article: Kouki, S. et al. (2026). "Against the grain: Leveraging machine learning to analyze mudbrick structures." PLOS ONE. https://doi.org/10.1371/journal.pone.0349295 Overview: This dataset provides a complete, reproducible Python-based pipeline for automated grain morphometry of archaeological mudbrick samples using petrographic thin sections. It evaluates how unsupervised machine learning can be used to extract objective grain size, sorting, and shape descriptors to understand ancient raw material selection, labor organization, and construction techniques. Contents: 1. Raw images: 45 cross-polarized light (XPL) microscopic images (RGB, .tif file, and 2280 x 2160 pixels; spatial calibration of 890 pixels = 1 mm and scale factor of 1.124 μm-pixel). 2. Segmentation masks: processed binary masks isolating the clusters (pores, grains, and clay matrix). 3. Morphometric data: extracted feature measurements (grain size, sorting, and shape descriptors) 4. Code: documented Python workflow including K-means clustering segmentation, feature extraction, and statistical/multivariate analysis scripts. Sampling content: The dataset includes samples from two geographically and chronologically distinct archaeological sites, which have been individually studied and published: 1. Artaxata, Armenia (Urartian/Hellenistic period, $n=27$ samples). See previous fabric analysis: https://doi.org/10.1371/journal.pone.0292361 2. Los Villares de la Encarnación, Spain (Early Iron Age, $n=18$ samples). See previous fabric analysis: https://doi.org/10.1515/opar-2022-0304