Blog

Using GO to Observe the Operational Status of Suspected Chinese “re-education” Camps

Challenge

Trying to understand trends in the construction rates and operational status of specific facilities in non-permissive environments can be time-consuming and difficult. other than having on the ground intelligence, satellite imagery may be used to explore these challenges. However, imagery still requires a human analyst to sift through imagery and identify additional construction manually before any contextual analysis can even take place.

Solution

Orbital Insight leverages computer vision algorithms to automatically classify land-use changes as well as detect objects. This automated analysis quickly helps gather more information, allowing analysts and users to focus on contextual analysis rather than just quantifiable trends.

By automatically classifying land use, a user can quantify land use classes as well as more easily detect changes between time ranges. Automatically detected changes over a multitude of camps allow an analyst to focus on the context surrounding those changes and pick which ones to evaluate more closely. Construction of these facilities provides information around the build-up, or break down, of the overall “Re-education” camp policies within China.

By analyzing historical imagery, and obtaining future imagery to come, a user has more time to create context surrounding a set of events or an issue. Using cloud computing and a robust imagery ingestion pipeline, 41 suspected Chinese re-education camps are analyzed in the do-it-yourself analytics platform, GO.

Additionally, automatically identifying cars at scale helps humanitarian and NGO organizations achieve another level of transparency in otherwise non-permissive environments. Car detections are used as a proxy for worker activity at foreign facilities of interest, and in this case, are observed at suspected camp parking lots and surrounding areas. Combined, car counts and land use classification datasets help users to more efficiently understand what is happening in specific areas of interest (AOIs).

Chinese re-education camp near Korla
Figure 1: Construction of a suspected Chinese “re-education” camp near Korla City, Xinjiang, May 15, 2018. (Imagery source: Digital Globe)

Background on Re-education Camps in Xinjiang

Up to an estimated one million people have been interned without trial, throughout Xinjiang Uighur Autonomous Region (XUAR), according to Amnesty International. XUAR has a large Muslim population, and following China’s regulations on “de-extremification” policies, camps were used to “re-educate” portions of the population. The reasoning, according to the Chinese government, was to combat religious extremism (Sudworth, 2019). Critics of Chinese policies argue human rights violations are taking place, while China maintains that Islamic extremists and separatists are to blame for unrest in the region, and therefore this is a way to make the region more peaceful (“China changes law to recognise ‘re-education camps’ in Xinjiang,” 2018). These events have also received global attention, with some countries praising China for taking action against extremism, and others condemning the camps (“Which Countries Are For or Against China’s Xinjiang Policies?, The Diplomat,” 2019). With scant evidence of what actually goes on inside the camps and opposing views of camps’ function, anecdotal evidence proves to be one of the only other sources of information.

Recently, China claimed it closed many of these suspected camps (Buckley & Myers, 2019). To attain a better understanding of if this is true, Orbital Insight can use GO to measure construction, employee parking lots, and potential destruction of suspected camp locations. This is a first step for analysts and non-governmental organizations (NGOs) to understand these locations’ operational status.

Land use Data

Land use aggregation and land use change detection projects were created in Orbital Insight’s GO platform and analyzed Planet imagery. The results were exported and visualized in QGIS 3.0. These results can also be accessed via an API and integrated with other analytic models. Metadata such as square meters per polygon and construction/destruction changes can be analyzed and tracked automatically over time.

Yuli re-education campyuli re-education camp construction
Figure 2: Left — Imagery ground truth showing construction changes year over year, Yuli Camp. No construction in 2016, partial construction in 2017, and finished construction in 2018. Right — Buildings detected in GO, in April 2018.
Yuli re-education camp
Figure 3: Example land-use change detection (LUCD) results from the LUCD algorithm comparing July 2017 changes to July 2018 changes. These polygons correspond to construction taking place between the center and rightmost image time frames in Figure 2. Polygons are created from well-known-text (WKT) strings and can be visualized in QGIS 3.0 for further analysis if needed.
Korla re-education campKorla re-education camp construction
Figure 4: Left — Imagery ground truth showing construction changes year over year, Korla Camp. No construction in 2017, mostly finished construction in 2018, and camp likely operational in 2018. Right — Buildings detected in GO, April 2018.
Korla re-education camp
Figure 5: Example land-use change detection (LUCD) results from the LUCD algorithm comparing July 2017 changes to July 2018 changes showing new buildings and roads, corresponding to the center image in Figure 4. Polygons are created from well-known-text (WKT) strings and can be visualized in QGIS 3.0 for further analysis if needed.
Korla re-education camp 2018 vs 2019
Figure 6: Example land-use change detection (LUCD) results from the LUCD algorithm comparing July 2017 changes to July 2018. Changes show new road segments and additional improved surfaces (detected as buildings in this case), corresponding to the rightmost image in Figure 4. Polygons are created from well-known-text (WKT) strings and can be visualized in QGIS 3.0 for further analysis if needed.

Object Detection: Car Detector

Based on the time series of these areas, raw car counts increase during construction and maintain a relatively high state once construction is complete. This would indicate that the facility is now in operation and employees of the facility are coming to work on a daily basis. These counts can be monitored in conjunction with LUCD to better understand when construction is complete and the facility is in operation. Conversely, if car counts decline, it could indicate a camp closure, which would verify the latest Chinese government claims of the facility shutting down.

Chinese re-education campsFigure 7: Blue dots indicating cars detected near Yuli camp, indicating worker presence following construction. Zoomed in of cars on right. (Imagery source: Digital Globe)
Figure 7: Blue dots indicating cars detected near Yuli camp, indicating worker presence following construction. Zoomed in of cars on right. (Imagery source: Digital Globe)
Car counts at Korla re-education camp
Figure 8: Korla cumulative mean counts over time, indicating an increase in cars as construction takes place and then leveling off once the facility is operational.
Car traffic at Chinese re-education camps
Figure 9: Cars automatically detected per camp between January 2017 and August 2019.

Conclusion

GO can be used to automatically track the rate of construction for foreign areas of interest in relatively non-permissive environments. The outputs allow for rapid interoperability for other analysis models concerning quantitative metrics, allowing users and analysts to focus on contextual and qualitative analysis.

Landuse change detection, in conjunction with car detection counts, can give insight into construction completion and/or facility operational status. Cars may indicate facility employees, indicating the facility’s status as well as how many workers it takes to operate a facility. By proxy, understanding employee numbers may also be an indication of how many non-employees are at the camp.

The automatically detected data generated by GO may be an indicator of a camp’s operational status. Therefore, with contextual analysis, the data may help prove or disprove Chinese claims that facilities are closing.


References:

Buckley, C., & Myers, S. L. (2019, August 9). China Said It Closed Muslim Detention Camps. There’s Reason to Doubt That. The New York Times. Retrieved from https://www.nytimes.com/2019/08/09/world/asia/china-xinjiang-muslim-detention.html

China changes law to recognise “re-education camps” in Xinjiang. (2018, October 10). Retrieved September 9, 2019, from South China Morning Post website: https://www.scmp.com/news/china/politics/article/2167893/china-legalises-use-re-education-camps-religious-extremists

Sudworth, J. (2019, June 21). Searching for truth in China’s “re-education” camps. BBC News. Retrieved from 

Which Countries Are For or Against China’s Xinjiang Policies? | The Diplomat. (n.d.). Retrieved September 9, 2019, from