Semi-automatic detection of building stones and wall segments of archaeological ruins

This study presents a semi-automatic methodology for detecting building stones and wall segments in archaeological research, using drone imagery and deep learning algorithms. The immediate outputs of the methodology are a georeferenced stones layer, with each stone detected as a separate instance, and a site plan layer, composed of stones considered part of detected wall segments. We developed this model via nine sites of varying size with different vegetation coverage, ground color, and material composition, exemplifying the model's ability to perform successfully even in challenging conditions. The digital layers, along with additional attributes associated with each shape, provide a foundation for further analysis, such as identifying multiple construction styles and site organization patterns, with significant potential for large-scale multi-site studies. Evaluation results demonstrate good model performance under varied conditions. We also provide trained models, trained on data from multiple sites, for immediate use and further refinement. • Detect stones and walls from drone imagery with deep learning. • Analyze archaeological sites using digitized stone and wall layers. • Provide open-source models and code for stone and wall detection. • Deliver an end-to-end workflow from drone images to GIS layers.