Large Scale ML-driven Detection and Orientation Analysis of Remote Heritage Sites: Case Study in Al-Hayit, Saudi Arabia

Abstract

In the past decade, machine learning-based large-scale archaeological analysis of sites (e.g., burial mounds and megalithic structures) in remote areas has been gaining traction with object detection models such as YOLO used to detect sites. However, most studies do not go further with the analysis. One underexplored dimension is the study of site orientations, which numerous cultural and historical studies have shown to play an important role in reflecting ritual practice, cosmological knowledge, and landscape integration. In this study, we propose a platform integrated within QGIS as a plugin for multidimensional site orientation analysis. The platform is validated on a real-life case study from Saudi Arabia. The case study shows how the analysis using our platform can be performed. Object detection using a YOLOv11 model resulted in an F1 score of 0.972 during validation. The orientation analysis provided good results using (1) image processing techniques (Mean Square Error of 0.936 \(\text{deg}^{2}\) ) and (2) promising ones using a ResNetv2 model (F1 score of 0.676 for validation for classifying the orientation in 72 classes of 5 degrees each and a Mean Square Error of 1.294 \(\text{deg}^{2}\) ). Curvigram analysis of the results from a domain specific perspective showed that the ResNetv2 model provides the closest statistical results to the baseline.

The authors' abstract, as published at the source. Journal on Computing and Cultural Heritage, 2026 · DOI ↗