Identifying spectral anomalies such as cropmarks is a valuable approach to detecting concealed archaeological sites. In this study, we explore a physically based modeling strategy, interpreting cropmarks as stress-induced changes in vegetation, governed by radiative transfer dynamics. Measurements were obtained from a controlled test field during two campaigns conducted 13 years apart. Using PROSAIL in both forward and inverse modes, combined with statistical modeling, we construct synthetic spectral datasets to augment limited observations, thus facilitating robust training of classification models. An ensemble of machine learning algorithms is applied, with classifiers trained on synthetic data and refined via majority voting. The models achieve detection rates exceeding 90% on earlier observations, demonstrating effective retrospective application. Results highlight the influence of plant growth phase, with peak greenness data yielding stronger performance. Interestingly, injected noise improves robustness, albeit modestly. This study establishes a reproducible pipeline for archaeological prospection, merging physical simulation with machine learning. The integration of synthetic data offers a promising solution to data scarcity, and opens pathways for applying predictive models to archive aerial and satellite imagery, thereby potentially transforming remote sensing strategies in heritage research.
Gravanis et al. (Thu,) studied this question.