WATCH: Wide-Area Archaeological Site Tracking for Change Detection

Girmaw Abebe Tadesse, Titien Bartette, Andrew Hassanali, Allen Kim, Jonathan Chemla, Andrew Zolli, Yves Ubelmann, Caleb Robinson, Inbal Becker-Reshef, and Juan Lavista Ferres · 2026

Read WATCH: Wide-Area Archaeological Site Tracking for Change Detection by Girmaw Abebe Tadesse, Titien Bartette, Andrew Hassanali, Allen Kim, Jonathan Chemla, Andrew Zolli, Yves Ubelmann, Caleb Robinson, Inbal Becker-Reshef, and Juan Lavista Ferres (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-08160.txt.

Introduction

Girmaw Abebe Tadesse and colleagues built WATCH. The system watches old sites from space and dates damage to the month. Guarding heritage at scale is hard. Damage cues in satellite images are subtle, and labeled cases are rare.

WATCH works with PlanetScope mosaics from 2017 to 2024. It covers 1,943 sites in Afghanistan. It offers three scoring routes. One is a training-free temporal embedding distance. One is a self-taught ensemble of signals. One is a weakly guided model trained on sparse event labels. The team tests embeddings from six foundation models beside handcrafted features. The result is a practical toolkit for heritage monitors.

Catalog record

CREATORGirmaw Abebe Tadesse, Titien Bartette, Andrew Hassanali, Allen Kim, Jonathan Chemla, Andrew Zolli, Yves Ubelmann, Caleb Robinson, Inbal Becker-Reshef, and Juan Lavista Ferres
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.08160

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

WATCH: Wide-Area Archaeological Site Tracking for Change Detection Girmaw Abebe Tadesse, Titien Bartette, Andrew Hassanali, Allen Kim, Jonathan Chemla, Andrew Zolli, Yves Ubelmann, Caleb Robinson, Inbal Becker-Reshef, and Juan Lavista Ferres Abstract Monitoring archaeological sites at scale is vital for protecting cultural heritage, yet pinpointing when disturbances occur remains difficult because visual cues are subtle and ground-truth data are sparse. We introduce WATCH, a framework for month-level change-event localization over PlanetScope satellite mosaics (2017–2024, 4.7 ​ m / px 4.7\,\mathrm{m/px} ) that supports three complementary scoring approaches: (i) Temporal Embedding Distance (TED), a training-free method that scores month-to-month deviations from a local temporal reference; (ii) Self-Supervised Change Detection (SSCD), an ensemble of reconstruction, forecasting, and latent-novelty signals; and (iii) a Weakly Supervised (WS) temporal localization model trained with sparse event-month labels. We benchmark WATCH on 1,943 archaeological sites in Afghanistan using embeddings from six foundation models (CLIP, GeoRSCLIP, SatMAE, Prithvi-EO-2.0, DINOv3, and Satlas-Pretrain) alongside a handcrafted spectral and texture baseline, and assess cross-regional generalization on sites in Syria, Turkey, Pakistan, and Egypt. The unsupervised approaches (TED, SSCD) consistently outperform the weakly supervised alternative. TED with SatMAE achieves the highest exact-month recall (55% at m = 0 m{=}0 ), while TED with GeoRSCLIP, CLIP, or Satlas-Pretrain reaches 92.5% within a three-month tolerance ( m = 3 m{=}3 ). Handcrafted features remain competitive for exact-month detection under weak supervision. Our directional margin analysis reveals systematic temporal biases: SSCD paired 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 materialized. These results show that satellite imagery combined with foundation-model embeddings enables scalable, decision-relevant heritage monitoring. Code: https://github.com/microsoft/WATCH . I Introduction Archaeological sites are irreplaceable records of human history and cultural identity, yet face growing threats from looting, unregulated construction, agricultural expansion, and conflict-driven destruction. Advances in earth observation (EO) now provide routine satellite imagery at monthly or sub-monthly intervals, establishing remote sensing as a key tool for heritage monitoring [ 14 , 6 , 13 , 11 ] . Because many sites are remote or situated in unstable environments, satellite-based monitoring has become indispensable for large-scale cultural heritage preservation [ 19 , 1 ] . For operational decisions, heritage practitioners require temporal localization of change events, not merely binary site-level classification, since accurate detection is critical for timely intervention. Satellite-based looting detection has a long history, but early approaches relied on visual inspection, sometimes augmented by simple image processing [ 12 , 19 ] . These manual workflows are labor-intensive, subjective, and difficult to scale. Looting leaves subtle, spatially dispersed traces, including disturbed soil and localized spectral anomalies, that are easily confounded with natural processes such as erosion or agriculture. Machine learning offers a scalable alternative [ 20 , 18 ] , yet automated systems still struggle to pinpoint when disturbances occur for three main reasons: ground-truth labels are sparse and often temporally imprecise; changes are visually subtle and heterogeneous, easily confounded by seasonal effects, illumination variation, and sensor noise; and large-scale monitoring must operate efficiently across thousands of sites with limited computational and annotation budgets. We address these challenges through site-centric monitoring, representing each site as a monthly time series of feature embeddings extracted from a masked (if available) spatial grid. Prior work on archaeological site monitoring has focused on binary classification, distinguishing looted from preserved sites using single images or full time series [ 20 , 18 ] . Recent geospatial foundation models, including SatMAE [ 5 ] , SatCLIP [ 10 ] , DINOv3 [ 16 ] , and Prithvi-EO [ 8 ] , have advanced performance on a range of downstream EO tasks, yet their effectiveness for temporal change detection at archaeological sites remains largely unexplored. We benchmark six foundation models in this study. Handcrafted features (e.g., spectral indices and texture descriptors) can capture looting-related patterns, but have not been rigorously compared to foundation-model embeddings in this setting. More broadly, change detection in remote sensing has been studied extensively, typically under fully supervised or bi-temporal protocols requiring dense annotations or paired imagery [ 7 , 4 ] . These methods are not directly applicable to our setting, which requires monthly temporal resolution under weak or absent supervision over long time series. Month-level change detection for archaeological sites, however, remains underexplored. No existing change detection framework for archaeological sites jointly (i) operates on monthly satellite time series, (ii) exploits modern foundation-model embeddings, (iii) functions under minimal or noisy supervision, and (iv) evaluates temporal tolerance in a manner relevant to real-world heritage monitoring. Fig. 1 : Overview of the WATCH framework. (a) PlanetScope monthly mosaics (RGB+NIR, 4.7 m/px, 2017–2024) are processed into 1 km 2 site patches with spatial masks for 1,943 archaeological sites in Afghanistan, and later validated across sites beyond Afghanistan, such as Syria, Turkey, Pakistan, and Egypt. (b) Foundation-model embeddings and handcrafted features are extracted and temporally normalized, then scored via three complementary approaches: TED (Temporal Embedding Distance), SSCD (Self-Supervised Change Detection), and WS (Weakly Supervised), each producing a monthly change score. (c) Scores are converted to change probabilities and evaluated using recall with symmetric and directional temporal margins, supporting heritage monitoring, looting detection, and temporal analysis applications. We present WATCH ( W ide-area A rchaeological site T racking for CH ange detection), a framework for month-level change detection in archaeological site monitoring using PlanetScope monthly mosaics (2017–2024, 4.7\,\mathrm{m/px} ) (see Fig. 1 ). WATCH operates on per-site time series of geospatial foundation-model embeddings and supports three complementary scoring approaches: (i) Temporal Embedding Distance (TED), a training-free deviation score; (ii) Self-Supervised Change Detection (SSCD), an ensemble of reconstruction, forecasting, and novelty signals; and (iii) a Weakly Supervised (WS) temporal localization model trained with sparse, truncated event-month labels. We benchmark embeddings from Satlas-Pretrain [ 3 ] , GeoRSCLIP [ 21 ] , SatMAE [ 5 ] , CLIP [ 15 ] , DINOv3 [ 16 ] , and Prithvi-EO-2.0 [ 17 ] against handcrafted features, and introduce a unified evaluation protocol based on Recall@ K (hereafter simply recall ) with symmetric and directional temporal margins with the top-K months based on their likelihood of change. We validate on sites in Afghanistan and assess cross-regional transfer on sites in Syria, Turkey, Pakistan, and Egypt (see Fig. 7 ). Our contributions are: 1. Formulating archaeological site monitoring as monthly change-event detection over site-centric satellite time series. 2. Proposing WATCH with three complementary scoring approaches (TED, SSCD, and WS) spanning unsupervised and weakly supervised regimes. 3. Systematically benchmarking six foundation-model embeddings against handcrafted features for this task. 4. Introducing a temporally tolerant evaluation protocol with directional margin analysis. To promote reproducibility and support continued research, the codebase has been made publicly accessible at https://github.com/microsoft/WATCH . II Dataset Our primary dataset comprises 1,943 archaeological sites in Afghanistan: 898 looted (46.2%) and 1,045 preserved (53.8%) as shown in Fig. 2(a) . Sites were compiled from the Archaeological Gazetteer of Afghanistan [ 2 ] and subsequently verified and expanded by expert archaeologists using high-resolution satellite imagery platforms (Google Earth, ESRI, Bing) to classify each site as “looted” or “preserved” based on evidence of deliberate damage before 2023. Of the looted sites, only 117 ( \sim 13%) have a reliably known

Read the complete text