Detecting desert kites in 3D point clouds by learning anomalies

Desert kites are large prehistoric hunting traps typically composed of two long, low stone walls that converge toward an enclosure.  These structures are widely distributed across the arid and semi-arid margins of the Middle East and Central Asia, exhibiting substantial variability in size, geometry, construction techniques, and topographic setting. To better understand their functionality from the Neolithic to sub-contemporaneous times, terrestrial laser scanning has increasingly been used to capture high-resolution three-dimensional representations of desert kites, enabling detailed characterization of their construction and local terrain setting. However, the kites’ subtle expression, their large spatial extent, and their progressive blending into the natural surface complicate their detection. These difficulties are further exacerbated by variable point density resulting from the alignment of multiple terrestrial scans, unavoidable occlusions caused by topography or vegetation, and the sheer volume of data produced by high-resolution ground-based surveys.  Together, these factors make the reliable identification and analysis of desert kite features within raw terrestrial point clouds a challenge, which requires extensive manual intervention and expert interpretation.

In this study, I present an automated, machine-learning-based approach for highlighting desert kite features directly within 3D point clouds derived from terrestrial laser scanning, without the need for manual annotation or labelled training data. The proposed method is based on the premise that the kites' structures introduce geometric irregularities (anomalies) relative to the surrounding natural surface. Rather than explicitly modelling the kite's form  or imposing predefined shape descriptors, the method learns a representation of the underlying terrain surface directly from the point cloud. This learned representation is then used to reconstruct the surface, which is subsequently compared to the original terrestrial measurements. Local deviations between the reconstructed surface and the original point cloud are quantified, with larger reconstruction errors interpreted as potential surface anomalies indicative of the kite's features. 

The proposed workflow is fully data-driven and unsupervised. It does not rely on prior knowledge of kite geometry, site-specific heuristics, or expert-defined thresholds. Instead, the learning process adapts to the local surface characteristics captured in the input dataset, making it robust to variations in resolution, occlusions, and terrain complexity commonly encountered in terrestrial laser scanning surveys. 

The findings demonstrate that surface-reconstruction-based anomaly detection offers a promising pathway for the automated identification of desert kite features in terrestrial 3D point clouds. More broadly, the approach is applicable to archaeological structures that exhibit weak or subtle geometric signatures. By reducing dependence on manual interpretation and labelled datasets, the method supports more objective, scalable, and reproducible analyses of archaeological landscapes, particularly in complex terrain where anthropogenic features are embedded within natural surfaces.