U-2 / KH Declassified Reconnaissance Imagery Project

Pilot Study · 2026/2027 · Computational Humanities

AI-Assisted Georeferencing and Segmentation of Declassified Reconnaissance Imagery

A U.S. Air Force C-119J 'Flying Boxcar' snags a parachute-borne CORONA film-return capsule in mid-air, 1960.
A C-119J recovers a CORONA film capsule in mid-air, 1960. U.S. Air Force, public domain.

And its potential for historical research

Principal Investigator Thorben Pelzer
Host School of Humanities and Social Science, HKUST
Duration 12 months (August 2026 – July 2027)

Overview

From the Second World War well into the Cold War, American intelligence services observed China from afar. At altitudes above 70,000 feet the Lockheed U-2 captured strikingly detailed imagery, while the CORONA Keyhole satellites exposed kilometres of Kodak film that was parachuted back to Earth. Shot during a period of rapid industrialisation and transformation, these images constitute a trove of historical insight into China.

Situated in the computational humanities, digital history, and historical GIS, this pilot study develops an AI-assisted pipeline to automatically georeference and segment the declassified reconnaissance imagery, with a focus on China.

Research directions the data may open

  1. Land use & settlement. Historical change in agricultural, industrial, and urban built-up areas and population (Stratoulias & Grekousis 2021).
  2. Transportation networks. Tracing historical transportation networks, including street layouts (McCarthy et al. 2026).
  3. Rivers & environment. Reconstructing former river flows and environmental change (Sinha 2021).
  4. Heritage. Reconstructing lost heritage sites and buildings of significance (Ryavec et al. 2025).
  5. Archaeology. Discovering locations of archaeological relevance (Fowler 2004; Goossens et al. 2006).

Background

Ever since the gradual declassification of these images, and parallel advances in GIS, researchers have pointed to the potential of U-2 and CORONA imagery. To date, discoveries using CORONA imagery have relied on the manual georeferencing of a small subset of images from a pre-defined area of scholarly interest, while the automated georeferencing of the full corpus has so far largely been theorised, with promising approaches beginning to emerge (Li et al. 2023).

The archaeologist Jason A. Ur pioneered the humanities use of U-2 and CORONA imagery, demonstrating its value for the study of ancient Mesopotamia (Ur 2010). This project opens complementary ground — a focus on China and the use of AI-assisted georeferencing and segmentation — and builds directly on the PI’s prior historical-GIS work on geographic mobility (Pelzer 2025) and the simulation of historical transportation from georeferenced archival maps (Pelzer 2026), extending the available spatial data both in coverage and backward in time.

Methodology

Generating the dataset poses two computational problems at the intersection of remote sensing, machine learning, and GIS.

01

Georeferencing

Locating each image on a geographic coordinate raster and warping it onto a modern map projection. The project stays open to a range of computational approaches, from feature-based image stitching (SIFT/SURF, homography, RANSAC) to topographic and natural-feature mapping — for example, detecting street intersections that can serve as ground-control points (Shensky et al. 2024; Cléri et al. 2014).

02

Segmentation

Vectorising each image through a segmentation model that reliably identifies features such as crop fields and roads. This is expected to be the less difficult half, advanced through collaboration with HKUST remote-sensing groups, manual tagging and training, and a recently released preliminary model (Hao et al. 2025).

Sources

Two main pillars supply the project: U-2 aerial photography held on analogue reels at the National Archives and Records Administration (NARA) in Washington, D.C., and CORONA Keyhole satellite imagery being digitised by the U.S. Geological Survey (USGS) on a rolling, on-demand basis. Less-established wartime aerial reconnaissance — partly captured from Japanese inventories — forms a third strand.

Overview of available declassified reconnaissance imagery
Type Time period Archives / license vendors
Original & captured wartime aerial photography 1938–1945 NARA; National Collection of Aerial Photography
U-2 aerial photography 1952–1978 National Archives and Records Administration
CORONA Keyhole satellite images 1959–1972 United States Geological Survey
A wide CORONA Keyhole satellite strip showing the mountainous terrain and river valleys around Lanzhou, China, in the early 1960s.
A CORONA Keyhole satellite strip showing the region of Lanzhou, China, in the 1960s — the kind of imagery the project will georeference and segment. United States Geological Survey (DS1114-1103DA011).

Schedule

The pilot study runs for twelve months across five overlapping phases.

Targeted schedule for the pilot study (months 1–12)
Phase 123 456 789 101112
Set-up phase
Archival research
Georeferencing work
Segmentation training
Pilot report

Team & Collaboration

The project fosters interdisciplinary integration between the Division of Humanities (HUMA) and other HKUST units, bringing remote-sensing domain knowledge — including expertise in using satellite imagery for archaeological discovery — to a humanities research agenda.

Thorben Pelzer

Principal Investigator · HUMA, HKUST

Digital history, historical GIS, history of engineering, history of modern China.

Piyush Yadav

Collaborator · EMIA, HKUST

PhD researcher contributing remote-sensing domain expertise to the project.

Óscar N. Bernal

Collaborator · CIVL CliMet Lab, HKUST

Prospective PhD researcher supporting satellite-imagery methods and segmentation.

Selected References

  1. Cléri, I., Pierrot-Deseilligny, M., & Vallet, B. (2014). Automatic Georeferencing of a Heritage of Old Analog Aerial Photographs. ISPRS Annals, 2(3), 33–40. doi:10.5194/isprsannals-II-3-33-2014.
  2. Duan, W., Chiang, Y.-Y., Leyk, S., Uhl, J. H., & Knoblock, C. A. (2020). Automatic Alignment of Contemporary Vector Data and Georeferenced Historical Maps Using Reinforcement Learning. IJGIS, 34(4), 824–849. doi:10.1080/13658816.2019.1698742.
  3. Dyke, K. (2026). Historical GeoAI: The Promise of Unlocking Analog Spatial Data. Journal of Map & Geography Libraries, 22(1). doi:10.1080/15420353.2026.2628590.
  4. Fowler, M. J. F. (2004). Archaeology through the Keyhole. Interdisciplinary Science Reviews, 29(2), 118–134. doi:10.1179/030801804225012635.
  5. Goossens, R., de Wulf, A., Bourgeois, J., Gheyle, W., & Willems, T. (2006). Satellite Imagery and Archaeology: The Example of CORONA in the Altai Mountains. Journal of Archaeological Science, 33(6), 745–755. doi:10.1016/j.jas.2005.10.010.
  6. Hao, T., Zhang, L., Zhang, Y., Chen, M., Zhang, J., Dong, R., & Fu, H. (2025). WakeupUrban: Unsupervised Semantic Segmentation of Mid-20th-Century Urban Landscapes with Satellite Imagery.
  7. Kwong, C. M. (2024). Japanese Occupation of Hong Kong, 1941–1945: A Spatial History Project. HKBU Library.
  8. Li, Z., Zhang, M., Zhang, Y., & Li, Y. (2023). Automatic Detection of Beacon Towers in Historical Aerial Images Using Improved FCOS. IJRS, 44, 1089–1113. doi:10.1080/01431161.2023.2174388.
  9. McCarthy, C., Phillips, S., Sternberg, T., et al. (2026). Mapping the Great Mongolian Road. Journal of Historical Geography, 91, 89–107. doi:10.1016/j.jhg.2025.11.015.
  10. Pelzer, T. (2025). Engineers on the Move: Elite Geographic Mobility in Republican China. Twentieth-Century China, 50(1), 25–55. doi:10.1353/tcc.2025.a950426.
  11. Pelzer, T. (2026). Towards a Historical Simulation of Spaces as Networks. In C. Armand, C. Henriot, & L. Lien (Eds.), Performing Power. Berlin: De Gruyter.
  12. Ryavec, K., Nyandak, T., & Bhum, Y. (2025). Monastic Architectural Reconstruction From a 1962 U-2 Aerial Photograph of Nartang in Central Tibet. Studies in Digital Heritage, 9(1), 1–36. doi:10.14434/sdh.v9i1.39637.
  13. Sato, R., Kobayashi, S., & Jia, R. (2016). Aerial Photographs of Mainland China Acquired by U-2 Spy Planes. Teledetekcja Środowiska, 54(1), 61–74.
  14. Shensky, M. G., Strickland, K. M., Marden, A. W., & Dubbe, H. (2024). An Automated Georeferencing Workflow for Historical Sanborn Maps. Journal of Map & Geography Libraries, 20(3), 137–163. doi:10.1080/15420353.2025.2462737.
  15. Sinha, R. (2021). Reconstructing the Ganga of the Past from Corona Archival Imagery. New Delhi.
  16. Stratoulias, D., & Grekousis, G. (2021). Information Extraction and Population Estimates of Settlements from Historic Corona Satellite Imagery in the 1960s. Sensors, 21(7), 2423. doi:10.3390/s21072423.
  17. Ur, J. A. (2010). Urbanism and Cultural Landscapes in Northeastern Syria: The Tell Hamoukar Survey, 1999–2001. Oriental Institute of the University of Chicago (OIP 137).