WATCH: Wide-Area Archaeological Site Tracking for Change Detection

Researchers introduce WATCH, a satellite-based framework using foundation models to detect disturbances at archaeological sites across months and years. The system combines three approaches—temporal embedding distance, self-supervised change detection, and weakly supervised learning—achieving up to 92.5% accuracy within three-month tolerance windows when monitoring 1,943 Afghan sites and cross-validating in Syria, Turkey, Pakistan, and Egypt.

WATCH addresses a critical infrastructure challenge in cultural heritage preservation by automating change detection across wide-area satellite imagery. Traditional archaeological site monitoring relies on sparse ground surveys and manual inspection, making timely disturbance detection nearly impossible at regional or continental scales. This framework leverages freely available PlanetScope satellite mosaics (4.7m resolution) and foundation model embeddings to democratize heritage protection, enabling institutions with limited budgets to monitor vast areas continuously.

The research demonstrates that foundation models trained on diverse geospatial data outperform handcrafted spectral features for month-level change localization. Notably, unsupervised approaches (TED and SSCD) consistently exceed weakly supervised baselines despite minimal labeled training data—a significant finding suggesting these methods generalize across geographic regions without costly annotation. The directional margin analysis reveals interesting behavioral differences: GeoRSCLIP and Prithvi-EO-2.0 exhibit early-warning capabilities, flagging anomalies before official event records, while TED provides confirmation-oriented detection after disturbances materialize.

For the geospatial AI and earth observation sectors, this work validates that satellite imagery combined with modern foundation models creates viable commercial applications beyond traditional remote sensing. Organizations managing UNESCO sites, governments tracking looting or construction encroachment, and cultural institutions now have accessible tools to prevent heritage loss at scale. The cross-regional generalization results suggest deployment potential across diverse landscapes and political contexts. Future iterations might integrate real-time data streams, finer temporal resolution, or multimodal inputs combining satellite and radar imagery to enhance early-warning capabilities.

• →WATCH achieves 92.5% three-month detection accuracy using satellite imagery and foundation models without expensive labeled training data.
• →Unsupervised methods consistently outperform weakly supervised approaches, suggesting strong generalization potential across geographic regions.
• →GeoRSCLIP and Prithvi-EO-2.0 embeddings show early-warning profiles, detecting disturbances before official event records.
• →Foundation models prove more effective than handcrafted spectral features for archaeological site monitoring at scale.
• →The framework enables cost-effective heritage protection for resource-constrained institutions across Afghanistan, Syria, Turkey, Pakistan, and Egypt.