Multi-sensor UAV predictive modelling for archaeological feature detection: from probability surfaces to excavation targets on Lesvos Island, Greece

Abstract

Archaeological predictive modelling has advanced substantially with machine learning, but four limitations persist: reliance on single sensor types, opaque model behavior, poor transferability between sites, and the absence of a defensible procedure for converting probability surfaces into discrete excavation targets. This study presents an integrated UAV-based workflow that addresses all four through a single pipeline applied at two sites on Lesvos, Greece. Thermal, Multispectral, and LiDAR-derived topographic predictors were extracted from 1.5 m grid cells at the documented training site of Thermi, reduced via grouped Principal Component Analysis with significance and collinearity filtering to 21 mutually independent components, and modelled through binary logistic regression. The Hybrid model achieved AUC = 0.964 and F1 = 0.726 at Thermi, with cross-domain Spearman correlations confirming largely independent information from the between thermal and topographic predictors and a 12% F1 improvement over the best single-sensor model. A Random Forest benchmark trained on the same 21 components produced only a 2.6 percentage-point F1 advantage at the cost of interpretability and transferability. The fitted model was transferred to the unexplored hilltop site of Arisvi via standardization–projection without refitting; thermal and topographic components transferred successfully while vegetation predictors failed due to seasonal mismatch. A two-stage Getis-Ord Gi*–Local Moran’s I cluster analysis converted the Hybrid output into 1,440 statistically prioritized target cells covering 3,240 m 2 , a 13-fold reduction, providing a defensible bridge from predictive modelling to fieldwork.

The authors' abstract, as published at the source. Journal of Archaeological Science Reports, 2026 · DOI ↗