WATCH AI detects archaeological site changes from space with 92.5% accuracy

A team of researchers from Microsoft, Planet, and other institutions have developed WATCH (Wide-Area Archaeological Site Tracking for Change Detection), a framework that leverages PlanetScope satellite mosaics from 2017 to 2024 at 4.7 meters per pixel resolution to detect month-level disturbances at archaeological sites. The framework employs three complementary scoring approaches: Temporal Embedding Distance (TED), a training-free method that measures month-to-month deviations; Self-Supervised Change Detection (SSCD), an ensemble of reconstruction and forecasting signals; and a Weakly Supervised (WS) temporal localization model. Using embeddings from six foundation models—CLIP, GeoRSCLIP, SatMAE, Prithvi-EO-2.0, DINOv3, and Satlas-Pretrain—alongside handcrafted spectral features, WATCH was benchmarked on 1,943 archaeological sites in Afghanistan and tested for cross-regional generalization in Syria, Turkey, Pakistan, and Egypt.

The results show that unsupervised approaches (TED and SSCD) consistently outperform the weakly supervised alternative. TED with SatMAE achieves the highest exact-month recall at 55%, while TED paired with GeoRSCLIP, CLIP, or Satlas-Pretrain reaches 92.5% recall within a three-month tolerance. A directional margin analysis reveals systematic temporal biases: SSCD with GeoRSCLIP or Prithvi-EO-2.0 exhibits the strongest early-warning profile, detecting anomalies before the recorded event, while TED favors confirmation-oriented detection after a change has occurred. These findings demonstrate that satellite imagery combined with foundation-model embeddings can enable scalable, decision-relevant heritage monitoring at a global scale.