Learning Anomalies to Highlight Archaeological Structures in Cluttered 3D Point Clouds: Demonstration on Terrestrial Scans of Desert Kites

Abstract. Three-dimensional point clouds are becoming indispensable in archaeological studies. They are mostly used to document the site, to digitally visualize it, or to analyse its topographical context. Since many of the sites are embedded or semi-embedded within the terrain, they tend to be overlooked, misrepresented, or simply removed in the digital terrain modelling process. Therefore, a common practice is to manually mark them within the raw dataset and to add them to the finalized m odel. Here we propose a machine-learning approach to highlight regions that include archaeological features within 3D point clouds. It is based on the assumption that such features will present an anomaly within the surface. Therefore, the proposed method learns to reconstruct the surface from the acquired point cloud and then compares the reconstructed surface to the original one. In this way, a large error will signify an anomaly, i.e., a feature of interest within the point cloud. We demonstrate the proposed method on terrestrial laser scans of desert kites. These large ancient desert traps are found across the Middle-East and Central Asian arid and semi-arid regions. Their unique construction, made of two low long walls that converge into an enclosure, makes it difficult to distinguish them terrestrially. We show that using the proposed method, we can highlight the kites within the raw point cloud, without the need of an expert input and without labelled information.