Physically Based Predictive Modelling of Archaeological Proxies Using Cropmarks

ABSTRACT Cropmarks, as archaeological proxies, offer a valuable means of detecting buried sites through remote sensing. Yet, the scalability of such methods across varied archaeological contexts remains underexplored, and AI‐based modelling approaches are still in early stages. This gap stems from environmental variability, limited ground‐truth data due to intrusive validation and insufficient collaboration between archaeology and computer science. In this study, we assess the predictive performance of machine learning models trained on synthetic spectral signatures—representing cropmarks and healthy crops—generated via PROSAIL‐based radiative transfer inversion. These synthetic data serve as a necessary augmentation of the original observations for model training and statistical analysis. These models are tested on new observations from the same test field, collected in a different year. Our ensemble approach uses multiple classifiers trained on batches of synthetic samples with majority voting and explicitly examines the impact of input noise. Results show detection rates of up to 92% in specific cases, with peak accuracy achieved when the synthetic dataset is twice the size of the original observations. These findings underscore the potential of radiative transfer–based synthetic data to support scalable cropmark detection in both new and archival hyperspectral datasets.