Answers are grounded in this paper and cite their sources.
Buried archaeological remains can induce subtle and temporally variable spectral anomalies in overlying soils and vegetation, making their detection from multispectral satellite imagery challenging. This study evaluates whether unsupervised machine learning applied to multi-temporal multispectral data can identify archaeological anomalies without incorporating archaeological information during the clustering stage. A semi-automatic workflow based on K-means clustering was applied to Sentinel-2 and PlanetScope imagery acquired between January 2020 and March 2026 over two Roman archaeological sites in central Italy: Stalla Buoi (Ostia-Portus) and Podere Colle Agnano (Labro). Spectral indices were used to construct pixel-level feature matrices incorporating multi-temporal and seasonal statistics, while Jeffries-Matusita distance and phenological time-series analyses were used to identify the most discriminating spectral indices and optimal temporal windows. The resulting clusters were validated against archaeological evidence derived from magnetometry and very-high-resolution satellite imagery. Agreement with the reference data was moderate to substantial in three of the four site/sensor combinations (Cohen’s Kappa = 0.439–0.621), and fair for PlanetScope at Podere Colle Agnano (Kappa = 0.265). Spectral separability consistently peaked during winter (December-February), indicating that archaeological anomalies are primarily expressed through phenology-dependent vegetation responses. The results demonstrate the potential of long-term, unsupervised multispectral satellite analysis for label-free archaeological prospection and highlight phenology as a key factor in defining optimal detection windows.
No takes yet. Share an insight, caveat, or question.
Corbo et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: