Technical Summary: WATCH – Wide-Area Archaeological Site Tracking for Change Detection

Problem Statement

Monitoring archaeological sites at a global scale is critical for protecting cultural heritage from looting, unregulated construction, and conflict. However, pinpointing the exact timing of disturbances remains a significant challenge. Visual cues of looting (e.g., disturbed soil, spectral anomalies) are often subtle and easily confounded by natural processes like erosion or seasonal agricultural cycles. Furthermore, ground-truth data regarding the specific timing of looting events is extremely sparse and often temporally imprecise.

Existing automated systems struggle to localize when disturbances occur because:

1. Label Scarcity: High-quality, temporally precise labels are rare.
2. Visual Subtlety: Changes are heterogeneous and masked by seasonal effects and sensor noise.
3. Scalability: Monitoring thousands of sites requires efficient operation with limited computational and annotation budgets.

Current remote sensing change detection methods typically rely on fully supervised, bi-temporal protocols requiring dense annotations or paired imagery, which are not directly applicable to the task of monthly temporal localization over long time series with weak supervision.

Methodology: The WATCH Framework

The authors introduce WATCH (Wide-area Archaeological site Tracking for CHange detection), a framework designed for month-level change-event localization using PlanetScope satellite mosaics (4.7 m/px, 2017–2024). The system processes per-site time series of geospatial feature embeddings and employs three complementary scoring approaches:

1. Feature Extraction

The framework extracts feature vectors from masked 1 km² monthly mosaics for 1,943 archaeological sites in Afghanistan (and validates on sites in Syria, Turkey, Pakistan, and Egypt). It utilizes embeddings from six geospatial foundation models alongside a handcrafted baseline:

• Foundation Models: CLIP, GeoRSCLIP, SatMAE, Prithvi-EO-2.0, DINOv3, and Satlas-Pretrain.
• Handcrafted Baseline: 60 statistical, spectral-index, and textural features (e.g., NDVI, NDWI, GLCM, LBP) computed directly from imagery without learned models.
• Normalization: A two-stage normalization process removes seasonal effects (standardizing by calendar month statistics) and global scale mismatches.

2. Scoring Approaches

WATCH supports three distinct strategies for generating monthly change probabilities:

• Temporal Embedding Distance (TED): A training-free, unsupervised method. It calculates the distance (L2 or cosine) between a target month's embedding and a robust reference vector (coordinate-wise median) of the preceding $R=3$ months. It is designed to detect deviations from local temporal history.
• Self-Supervised Change Detection (SSCD): An unsupervised ensemble method that combines three signals:
  1. Reconstruction error (masked autoencoder).
  2. Forecast error (next-month predictor).
  3. Latent novelty (inverted k-NN density in latent space).
     These signals are weighted and aggregated to produce a change score.
• Weakly Supervised (WS) Localization: A supervised temporal classifier (LSTM-based) trained on sparse, truncated event-month labels. It predicts the probability of a change event at each month, utilizing Gaussian-smoothed targets around known event months to handle label uncertainty.

3. Evaluation Protocol

The authors introduce a temporally tolerant evaluation metric based on Recall@K (top-K months with highest change probability). They define:

• Symmetric Margin ($m$): A hit occurs if the top-K prediction is within $m$ months of the ground truth.
• Directional Margins:
  • Positive Margin ($h^+$): Detection occurs after the event (confirmation-oriented).
  • Negative Margin ($h^-$): Detection occurs before the event (early-warning oriented).

Key Contributions

1. Formulation: Reframed archaeological site monitoring as a monthly change-event detection problem over site-centric satellite time series.
2. Framework: Proposed WATCH, integrating unsupervised (TED, SSCD) and weakly supervised (WS) scoring regimes.
3. Benchmarking: Systematically evaluated six foundation-model embeddings against handcrafted features, demonstrating that foundation models generally provide stronger signals for month-level localization.
4. Evaluation Protocol: Introduced a temporally tolerant evaluation with directional margin analysis to distinguish between early-warning capabilities and post-event confirmation.

Results

The framework was benchmarked on 1,943 sites in Afghanistan (898 looted, 1,045 preserved), with only 117 looted sites having a reliably known month of looting.

• Performance: Unsupervised approaches (TED and SSCD) consistently outperformed the weakly supervised alternative.
  • TED with SatMAE achieved the highest exact-month recall (55.0% at $m=0$).
  • TED with GeoRSCLIP, CLIP, or Satlas-Pretrain reached 92.5% recall within a three-month tolerance ($m=3$).
  • SSCD with Satlas-Pretrain also reached 92.5% at $m=3$ and 100% at $m=6$.
• Foundation Models vs. Handcrafted: Foundation models consistently outperformed handcrafted features at strict margins (e.g., GeoRSCLIP exceeded handcrafted by 10 percentage points at $m=0$ under TED). However, handcrafted features remained competitive, particularly for exact-month detection under weak supervision (45.0% vs. 35.0% for Satlas-Pretrain at $m=0$ in WS).
• Directional Bias:
  • SSCD paired with GeoRSCLIP or Prithvi-EO-2.0 exhibited the strongest early-warning profile (negative directional margin), detecting anomalies before the recorded event.
  • TED favored confirmation-oriented detection (positive directional margin), reacting after a change had materialized.
• Generalization: When applied to sites in Syria, Turkey, Pakistan, and Egypt without retraining, unsupervised approaches (TED, SSCD) transferred effectively, producing temporally structured score trajectories. The WS model failed to generalize, remaining near-zero outside the Afghan training domain.

Significance and Claims

The paper claims that combining satellite imagery with foundation-model embeddings enables scalable, decision-relevant heritage monitoring under minimal supervision.

• Operational Guidance: The directional margin analysis provides actionable insights for practitioners. Those prioritizing early intervention should pair SSCD with geospatial foundation models (GeoRSCLIP, Prithvi-EO-2.0) to leverage pre-event anomaly detection. Those requiring post-event forensic confirmation can rely on the simpler, training-free TED approach.
• Robustness: The unsupervised nature of TED and SSCD allows them to generalize to out-of-distribution sites (different regions) without retraining, a critical advantage for global heritage monitoring where labeled data is scarce.
• Future Work: The authors identify the expansion of ground-truth coverage (currently limited to 117 known-month sites) as the most pressing avenue for future research, suggesting that active learning could strengthen both supervised training and evaluation reliability.

The codebase is publicly available to promote reproducibility and further research in this domain.