A Machine Learning–Based Predictive Model

Commissioned by the NEOM Heritage Department, we conducted an extensive interdisciplinary programme to document the rock-art, epigraphic and archaeological heritage of the Bajdah area, on the Ḥismā Plateau. Across a vast area —approximately 105 kilometers surveyed on foot and 24 sites visited, 7 of them newly identified— a machine learning–based predictive model was used to identify areas most likely to contain previously unknown evidence, directing resources where the likelihood of discovery was highest. All of this was achieved through a replicable methodology combining traditional archaeology, digital technologies and predictive modelling—now a benchmark for regional-scale documentation and conservation.

Predictive Model Results

Trained on known evidence and a geodatabase of environmental and morphological covariates, the model produced two operational tools: a continuous probability map assigning each grid cell an archaeological-potential value, and a binary map of high-potential areas generated by applying an optimised statistical threshold (TSS). These tools made it possible to direct field surveys towards the areas with the highest probability, concentrating resources where the likelihood of discovery was greatest.

Reliability is measured, not merely claimed: the model was validated using spatial block cross-validation, and only algorithms exceeding a predictive-quality threshold (continuous Boyce index ≥ 0.7) were included in the final ensemble. The result is also interpretable: the variable-importance analysis reveals the factors driving predictions in this area—primarily geology, topographic wetness (TWI), distance from watercourses and solar radiation—transforming the map from a black box into a genuine archaeological hypothesis that can be read and verified.