Predicting Archaeological Site Vandalism Risk in Pakistan Using GIS-Based Accessibility Indicators and Socio-Demographic Machine Learning Models

Archaeological site vandalism is a growing threat to cultural heritage in Pakistan, yet predictive, data-driven risk assessments remain limited. This study develops an integrated socio-spatial machine learning framework to predict vandalism risk in Khyber Pakhtunkhwa (KP). Field surveys at 50 sites using a Vandalism Observation Index (VOI) and 350 community questionnaires captured site conditions, awareness, drivers, protection, education, and visitation. GIS-based accessibility indicators were calculated for each site. Multivariate linear regression and Random Forest regression examined relationships between vandalism risk and social-spatial predictors. Linear regression explained 28% of variance in vandalism risk, while Random Forest improved performance (R2 = 0.398). Feature importance analysis showed that spatial accessibility variables, especially proximity to water sources, road networks, and populated places, were the dominant contributors. The results demonstrate the potential of combining GIS and machine learning to enable proactive, spatially informed heritage management in data-scarce regions.