Tracing settlement dynamics in Latium vetus: Explainable machine learning perspectives from the Bronze to the Early iron age.

Abstract

Understanding the long-term drivers of settlement location requires disentangling ecological constraints from historically contingent choices. Here I apply a Random Forest machine-learning framework, coupled with explainable AI techniques, to investigate settlement dynamics in Latium vetus (central Italy) from the Early Bronze Age to the early Iron Age (c. 2100-725 BCE). Using 108 securely dated sites and a set of environmentally constrained pseudo-absence points, I model phase-specific patterns and evaluate predictor importance through both normalized split frequencies and SHAP values, capturing structural roles as well as the magnitude and direction of effects. The results reveal a coherent diachronic transformation in the environmental logic of settlement. Early phases are primarily structured by hydrological accessibility, whereas from MBA3 onwards topographic configuration and elevation progressively dominate, marking a shift towards morphologically distinctive and defensible locations. This transition culminates in highly canalised settlement signatures during the Final Bronze Age and RMCA phases, before the RMCAIII phase signals a landscape that is increasingly saturated, hierarchically organised and functionally diversified, despite overall political stability. While the reliance on pseudo-absences and uneven sample sizes constrain predictive robustness, the approach demonstrates the value of Random Forest and SHAP as exploratory tools to expose non-linear, threshold-like relationships and to formalise long-standing archaeological interpretations. The study shows how explainable machine learning can provide a quantitative backbone for reconstructing the emergence, consolidation and transformation of territorially structured landscapes in protohistoric central Italy.