Paleolithic settlement patterns in the Central Balkans: A predictive approach on the territory of Serbia

Abstract

Predictive spatial modelling in archaeology has emerged as a valuable analytical tool for research and cultural resource management over the last decades. By combining GIS techniques and geostatistical methods, this tool examines the distribution of archaeological sites within their environmental context, aiming to understand settlement patterns and explain past human behavior through their link with the local environment. In this study, we compare three predictive modelling methods, Probabilistic Frequency Ratio, Generalized Additive Model, and MaxEnt applied to the case of the Paleolithic cave sites in the Central Balkans. The models were evaluated firstly through a threshold-independent measure using the Area Under the Curve (AUC), and secondly as binary models to assess their classification accuracy. While all three methods demonstrated good performances (AUC>0.85), MaxEnt proved to be the most efficient, as evidenced by both binary and threshold-independent evaluations. MaxEnt identified several high-probability areas based on geomorphological and ecological associations with the known archaeological cave sites. The reliability of MaxEnt's predictions was further reinforced by the substantial overlap with the results from the other two models. With each model built for two nested regions, we examined their respective interpolation and extrapolation performances to explore the potential of such approaches in cases where the archaeological data might be too scarce. The results suggest that such procedures might be applicable/warranted for investigating broader or underexplored regions with limited datasets.