Potential of SEM and Deep Learning in Archaeobotanical Identification of Ancient Wheat Varieties

Abstract

This study investigated the morphological similarity between bread wheat landraces from the Van Lake Basin and ancient Urartian wheat seeds (9th century BCE) discovered at & Ccedil;avu & scedil;tepe Fortress, utilising a Convolutional Neural Network (CNN)-based framework. Scanning Electron Microscopy (SEM) datasets were created using 15 lines from 10 landraces and the ancient seeds; EfficientNetB0, ResNet18 and InceptionResNetV2 models were employed to extract discriminative surface texture features. The ancient wheat samples showed the highest surface-texture similarity to the Muradiye-1-1 line (Red Kirik wheat) across the tested models (39.8% to 44.1%). These results suggest a phenotypic convergence between ancient and modern landraces under similar agroecological conditions, demonstrating the utility of CNN models for archaeobotanical analysis.