Modelling the predictability of Upper Palaeolithic sites in the Bistrița Basin (Eastern Carpathians) through GIS and machine learning

Predictive modelling has become an important tool in archaeological research, providing objective methods for identifying areas with high archaeological potential and investigating the relationship between site distribution and environmental variables. Although predictive approaches have been widely applied across many European regions, their use in Romanian Palaeolithic research remains limited. This study develops and evaluates three independent predictive models (Combine, Weighted Overlay Analysis (WOA), and Maximum Entropy (MaxEnt)) to identify areas of high archaeological potential within the Bistrița Basin (Eastern Carpathians, Romania). The models were developed using the spatial distribution of known Upper Palaeolithic sites together with selected geomorphological and environmental variables and were applied at two spatial scales: the entire Bistrița Basin and its central sector, the Ceahlău Basin. The three models revealed consistent spatial patterns, identifying fluvial terraces and gently sloping surfaces as the areas with the highest archaeological potential. Statistical validation using Receiver Operating Characteristic (ROC) curves and Area Under the Curve (AUC) values demonstrated acceptable to excellent predictive performance for all three models. MaxEnt achieved the highest predictive accuracy (AUC = 0.934 for the Bistrița Basin and 0.915 for the Ceahlău Basin), followed by WOA (0.867 and 0.877) and Combine (0.790 and 0.825). Model reliability was further supported by independent field validation through the identification of four previously unknown archaeological findspots near Călugăreni, two of which currently provide clear Upper Palaeolithic evidence. The results demonstrate that predictive modelling provides an effective framework for identifying new archaeological sites and investigating the relationship between Upper Palaeolithic settlement patterns and landscape characteristics. The proposed methodology offers a reproducible approach that can be tested and adapted in other geomorphologically comparable regions of the Carpathians, supporting both future archaeological prospection and cultural heritage management.