Satellite imagery: what it can and cannot tell us

Satellite data is particularly damning because it adds a more tangible layer of confirmation to that which has already been demonstrated through analysis of procurement contracts: namely, that the Chinese government has been building a massive network of internment camps/detention centers in Xinjiang. In fact, whereas procurement analysis was limited to just those bids that happened to be made public on a certain platform (and not all facilities would need to be constructed via competitive bids), satellite imagery can reveal a wider range of potential camps.

This post covers three major sources of satellite imagery by the Australian Strategic Policy Institute (ASPI), independent researcher Shawn Zhang, and Buzzfeed. These are not the only ones who have done this work (see also the BBC and Reuters, for example), but they have been incredibly prolific and meticulous, and taken together demonstrate the stunning breadth of evidence for mass incarceration in prison-like facilities, despite denial by the Chinese government. They are also ones for which I have seen more substantive objections on Twitter, some in good faith (and some not).

ASPI: The Xinjiang Data Project

Note that the Australian Strategic Policy Institute derives much of its funding from the Australian government and the defense industry, which creates a possible financial/political motive to exaggerate claims it makes about China, and leading to questions of bias. These questions are totally fair, and indeed ones that must be posed, but they mean nothing if they are not paired with substantive criticisms of ASPI's findings; no matter how wretched you think they might be, they might still be right. It's important to interrogate the motives and biases of sources, but anyone who has taken above a freshman history seminar knows that cannot be the sole locus of your critique, because that'd just be an ad hominem. So let's look at the data!

The Xinjiang Data Project maps what ASPI states it has identified as detention centers, cultural sites, and mosques in Xinjiang. It is impressive in its breadth and depth, and attracted a lot of attention on Twitter after its launch in September 2020, including by some for perceived inconsistencies—none of which I have ultimately seen pass basic muster, which I'll get to later.

The project dates back at least two years, to November 2018, when ASPI published an initial analysis of 30 camps into a single mini-database that you can view on Google Spreadsheets yourself. That report showed that the camps were expanding at a rapid pace.

For all entries in its database, ASPI provides the exact geo-coordinates that you can punch into Google Earth (Baidu Maps grays out imagery below a fairly high altitude in Xinjiang). It usually also contains a satellite image and occasionally a screenshot of a government document that mentions the facility. Some, like Peyziwat Facility #4, are also annotated.

Blank entries?!

ASPI relied on commercial satellite imagery in addition to free imagery. As a result, a handful of the sites appear barren in Google Maps/Earth, like Onsu Facility #3. If we put those coordinates into Google Maps today, we see an empty site with the tile on the bottom left showing more recent imagery (notice how the buildings on the bottom are split by the different images).

image

I reached out to ASPI analyst Nathan Ruser about some of these "blank" cases. For Onsu Facility #3, Ruser told me ASPI also contacted a journalist to confirm for them that the facility existed. Long story short: for some blanks, you have to trust ASPI. If you're skeptical, by all means, discount those blank sites, because fortunately, there aren't that many. I spent ten minutes or so clicking through 100-ish entries in Aksu, Hotan, and Kashgar prefectures and found four blanks: Atush Facility #8, Karakash Facility #5, Karakash Facility #12, and Kelpin Facility #2.

Sometimes, DigitalGlobe will have better coverage, but not always; in the case of Onsu Facility #3, the only available DigitalGlobe imagery is completely blocked by heavy clouds. DigitalGlobe data from December 2020 does show something has been built at the coordinates of Kelpin Facility #2, but the resolution is too low to make out.

image

385 identified camps

Satellite imagery is tricky. You can zoom in on your house in Google Earth, after all, but that does not (usually) show what's going onn inside. Internment camps are a bit more conspicuous than a random house, though, so there is still much to be gleaned from satellite imagery examination. The documentation report written by lead analyst Nathan Ruser explains some of the tricks used to track down camps. Sometimes it is as simple as looking at government procurement bids (see paper of Zenz explained here). Other times, ASPI states it contacted local journalists to check sites for them, if possible, as mentioned above. Finally, there's straight up geospatial intelligence analysis, including some rather clever techniques:

💬

Because the vast majority of the camps that we located were built on previously unused land in remote or peri‑urban areas, it was possible to compare illuminated areas in the first few months of 2017 ‑ before most of those camps had been constructed ‑ with presently illuminated areas. The new areas of night‑time light emissions were cross‑referenced against high‑resolution daytime satellite imagery that showed much greater detail. We discovered that many of the newly illuminated areas in these parts of Xinjiang were either newly constructed detention facilities or significant new highway checkpoints used to monitor the movement of people across Xinjiang.

Not all camps were built the same; there are higher and lower security facilities designed for different types of detainees. Higher security camps are more easily identifiable and "share identical distinguishing architectural features", Ruser writes in the report. Lower security camps are trickier because—as we know from procurement documents—they are sometimes converted civilian buildings, like schools. When examining such a facility, Ruser notes several features that can distinguish it as a detention center. These include:

  • Walls, barbed wire fencing, and guard towers; while many schools in Xinjiang and China more generally have some form of wall (and even barbed wire), the presence of barbed wire and towers is clearly reason for suspicion.
  • Relatedly, these camps generally have "an extensive network of barbed wire fencing that cages individual buildings, restricting the access detainees have to outside areas and channelling people through wire ‘tunnels’ between buildings."
  • If you've ever seen your neighborhood or school in Google Earth or the like, you've probably noticed that the imagery includes people, well, doing things. Going to classes, playing sports, walking around. Prison camps are not like that. Thus, "another key visual clue that can distinguish detention facilities is the lack of cars inside the facility and the absence of people in the satellite imagery acting normally" therein.
  • Camps are also sometimes colocated with factory complexes or prisons, unusual for schools.

Ultimately, ASPI divides camps into four tiers, from least to greatest security: Tier 1, lower security re‑education camps; Tier 2, dedicated re‑education camps; Tier 3, detention centers; and Tier 4, maximum security prisons. Again, these details are laid out in this report.

Objections to ASPI's findings generally seem to fall prey to an ignorance of geospatial data or of the nature of the camp system as documented in government procurement bids. Recall that such bids (e.g. this one I translated) explicitly show that reeducation facilities can be been set up in converted schools. See the final section for answers to various answers to detractors of ASPI's Xinjiang work.

Mosque and sacred site destruction

The other set of data in XJDP covers mosques and Uyghur holy sites. The Chinese government frequently touts the large number of mosques in Xinjiang as evidence against a widespread crackdown. ASPI produces strong evidence that many of these mosques have been shut down, and moreover that culturally and religiously significant sites have been destroyed as well. Writing in its September 2020 report:

💬

The Chinese Government’s 2004 Economic Census identified more than 72,000 officially registered religious sites across China, including more than 24,000 mosques in Xinjiang. Given the lack of access to Xinjiang and the sheer number of sites, we used satellite imagery to build a new dataset of pre-2017 mosques and sacred sites. We found the precise coordinates of more than 900 sites before the 2017 crackdown, including 533 mosques and 382 shrines and other sacred sites. Each of those sites was then cross-referenced against recent (2019–2020) satellite imagery and categorised as destroyed, significantly damaged, slightly damaged or undamaged. In most cases, significant damage relates to part of the site being destroyed or to Islamic-style architecture (such as domes and minarets) being removed. We then used a sample-based methodology to make statistically robust estimates of the region-wide rates of destruction by cross-referencing it to data from the 2004 Economic Census, by prefecture. [...] Extrapolating those figures on a prefectural level from official statistics allowed us to estimate the full number of destroyed and damaged mosques in Xinjiang. We found that across the XUAR approximately 16,000 mosques have been damaged or destroyed and 8,450 have been entirely demolished. The 95% confidence range of our regional findings is ±4% for the estimates of demolished, destroyed and undamaged mosque numbers.

As noted previously, satellite data, of course, cannot tell us everything. Some limitations to keep in mind:

  • Satellite imagery generally cannot tell if a mosque is still active or has been closed down/abandoned, nor can it reveal anything about conditions inside a mosque; it is " only able to determine demolition or other visible structural changes to the sites." (This is arguably less salient for mazars/shrines, which are inherently significant given their history.) ASPI notes third-party reports and unnamed first-person testimony to ASPI strongly suggests many mosques, particularly ones with less visibly Islamic features, have been converted into public/commercial spaces, or have been completely closed down, but does not factor this into its numbers. Thus, ASPI may be systemically undercounting the loss of mosques in the region.
  • Satellite imagery alone cannot tell why a mosque or other structure may have been demolished or destroyed or renovated. For this reason, the report cites extensive literature that plausibly ties the mass changes to deliberate government policy. Even so, there may be exceptions—a mosque may have been old and decrepit, for instance. (It should be noted, however, that this would also be tied to questions of government policy: the Xinjiang government has significantly decreased the ability to communities to raise money for repairs and renovations of mosques, and has also greatly slowed approval of such repairs. See "Analysis of the changes in number and management policy of Xinjiang mosques" (新疆清真寺的数量变化及管理政策分析), p. 44.)
  • This is a statistical argument made because of the infeasibility of locating each of 24,000-odd mosques given imprecise official data. As testimony elsewhere has demonstrated, systematic, on-the-ground attempts by outside journalists to verify the status of mosques and cultural sites in person would likely see them halted by the police.
From ASPI's report
From ASPI's report

Supplementary evidence

The following are examples of evidence that corroborate the central thesis of the ASPI report.

  • Similar investigation several months prior to ASPI's report: In addition to the above-mentioned BBC reports, the work by Buzzfeed explained below, an additional investigation in May 2019 by The Guardian and Bellingcat based in part on sites initially identified by Shawn Zhang identified "more than two dozen Islamic religious sites partly or completely demolished since 2016". The piece includes plenty of comparative imagery showing such destruction:
  • image
  • Published academic work from Party School researchers suggest the idea is far from fringe: Yang Weiwei, a researcher at the Altay Prefectural Party Committee Party School, published "Operational Research on Restraining the Infiltration of Religious Extremist Thought,” in 2015. Yang argued Xinjiang had far too many mosques, and that their number ought to be brought down, with remaining ones strictly regulated for architectural style (presumably, removing Arab-Islamic architectural features, as is documented practice elsewhere), "and that their opening hours should be limited to a single day every week and holidays." She also advocates strict prohibition of veils, Islamic dress, and facial hair. (Chinese copy of Yang's paper available via the Xinjiang Documentation Project here. The abstract was archived by ASPI; the article is also listed in the table of contents by CNKI as of September 26, 2020.)
  • Threads like the following, written by Professor Timothy Grose, a specialist on Xinjiang, show that even ignoring the number of destroyed or damaged mosques, there is strong evidence that they have suffered spiritual damage, so to speak.

Shawn Zhang

Shawn Zhang is a PhD student in Canada whose early work personally combing through Chinese government bid contracts and satellite imagery of Xinjiang helped inspire later analysis by other organizations. He started in May 2018—before the Chinese government admitted to the existence of the facilities—and by June 2019, he had identified 92 camps. Note that these 92 include 9 camps for which he found no evidence besides satellite imagery but which he assessed as "very likely" to be camps. Most, however, he tracked down based on public procurement documents explicitly calling for the construction of detention centers.

Sean is a thorough amateur very willing to make corrections when he is wrong. His posts are worth reading because they frequently include links to primary sources he refers to (although some are deleted or otherwise disappear due to link rot). These often include some very revealing findings, such as in the post which links to a full procurement bid (saved here) for Shache (Yarkant) County Technical School "security equipment" that includes things like teargas launchers and riot batons. A quick translation of pages 47-48 of said bid (I am unsure of the alfalfa thing, that's definitely what it said in Chinese—perhaps preparing for using laborers to grow crops? The School may or may not be a formal reeducation camp.):

image

It's rare to see full bids, usually one just sees the notices. I am unable to completely verify the document's authenticity, but I did establish that a bid with the same title, code and basic details was posted on a procurement platform and subsequently indexed by Tianyancha, a corporate registry website. Note that Tianyancha does not allow foreign IP addresses; you'll need a VPN with servers in China to access the link. Here's a screenshot:

image

Buzzfeed

In late 2020, Buzzfeed News reporter Megha Rajagopalan—the previous Buzzfeed Beijing bureau chief, before her visa was allowed to expire after her reporting in Xinjiangin 2018—and British architect Allison Killing published a major investigation into camps in Xinjiang based on a novel method of identification. In short, they used censorship of Baidu Maps to narrow down the range of possible camp locations, which helped reveal several previously unknown camps.

Geospatial information is subject to many restrictions in China (Wikipedia has a decent overview here), and this includes satellite imagery. Baidu Maps, the dominant mapping software in China, blanked out sensitive areas—usually military bases, etc.—at a specific height with unique grey tiles, which, of course, sometimes just made them stand out even more. The feature was not intentional; these tiles only appeared at zoom level 18. Buzzfeed used these grey tiled areas in Xinjiang as a starting point for their investigation, reasoning that China would likely seek to cover up camps. You can read the article detailing their methodology here, or take a look at this short thread by Killing with pictures:

In short, after discovering these grey tiles corresponded with already-existing camps (along with many other sensitive sites), they looked at every single such grey tile in Xinjiang and then started to narrow down which ones might be internment camps or other detention-related facilities as opposed to other sensitive sites. As they explain:

💬

Prisons and internment camps need to be near infrastructure — you need to get large amounts of building materials and heavy machinery there to build them, for starters. Chinese authorities would have also needed good roads and railways to bring newly detained people there by the thousand, as they did in the early months of the mass internment campaign. Analyzing locations near major infrastructure was therefore a good way to focus our initial search. This left us with around 50,000 locations to look at. We began to sort through the mask tile locations systematically using a custom web tool that we built to support our investigation and help manage the data. ... After looking at 10,000 mask tile locations [in Kashgar prefecture] and identifying a number of facilities bearing the hallmarks of detention centers , prisons, and camps, we had a good idea of the range of designs of these facilities and also the sorts of locations in which they were likely to be found.

Objections of the Twitter hordes

ASPI is wrong about __________ building!

For objections to ASPI's findings, just look at the lead analyst's thread answering them as they popped up. He's very thorough, and there's no way I'd do a better job explaining it than him.

Most of the objections are fairly asinine and come largely from random trolls. An exception is a thread by Chengxin Pan, a professor in Australia, whose failure to read the ASPI report and documenting evidence is very clear. Ruser's response thread detailed and worth specifically highlighting:

This camp is actually a ___________!

A common issue is that people seeking to verify will punch in coordinates onto Google Maps and instantly tweet about how the supposed coordinates don't align with what is on the map; therefore, what is actually an alleged internment camp is clearly a chicken farm. So insinuated the less funny Bad China Takes at one point about Buzzfeed's investigation (see my angry quote tweet response here).

Specifically, in response to this objection, architect Allison Killing—lead author of the Buzzfeed reports—gives a concise explanation: "while Google uses the WGS84 projection system, China uses its own different one, GCJ-02, also known as ‘Mars coordinates’, which makes it difficult to align GIS data." What that means in functional terms can be seen in this image:

image

Buzzfeed's results aren't reproducible, the tiles don't exist!

Let's deal with the most common objection by denialists: the grey tiles aren't on Baidu now, so the entire premise is unreproducible. The Global Times, a nationalist outlet under the auspices of the Party publication People's Daily, quoting a random anonymous Twitter user ("Chi", i.e. @_tchiek, quite the abrasive one), put it this way:

💬

When reached by the Global Times, Chi said that Killing claimed to have started the research using Baidu in 2018, and Baidu has removed the light gray tiles that were used to find the "camps" - "Baidu has changed." "They said Baidu has changed, which cannot be confirmed… anyway, they are not objective [on the topic]," Chi said.

This misunderstands the significance of the grey tiles—the proof is not the existence of the tiles themselves, but the underlying satellite imagery revealed on non-Chinese maps. Xinjiang is huge, the tiles were just the way to find the sites to further investigate. The evidence itself is the imagery and corroboration that Buzzfeed (says it) did as a journalistic organization. But ignoring that for a second, let's examine why those tiles are't there. Two hypotheses:

1) Buzzfeed is lying as a tool of the imperial media apparatus; the tiles were never there 2) Buzzfeed is telling the truth; the tiles were there and taken down later

Now, if we assume hypothesis 1 is true, we first have to concede that Buzzfeed is at worst a contract hire for the Man, not a full-time employee. They have regularly published confidential information leaked or pried in court from the US government, so unless you are a complete tinfoil hat and think all of that is a front, then we can only say it sometimes goes along with US propaganda. This, in itself, is not necessarily conspiratorial thinking; there are cases of prestigious media outlets being duped by sham intelligence and bad informants, like the New York Times in its coverage of Iraq. The Times, like countless others, failed spectacularly post-9/11 to balance the claims of Bush and Blair pushing for the catastrophic war in Iraq.

Yet the circumstances surrounding Iraq and Xinjiang are different in several ways. First, there is no serious calls by Washington for an invasion of Xinjiang at present, unlike there was for Iraq in 2002/3, when mainstream media paraded its own army of military and retired military professionals on TV to advocate for war. Second, whereas evidence supporting the invasion of Iraq came from a handful of sources—Colin Powell's speech at the UN, the Times' faulty informant-activist—the investigations conducted by the above media outlets are their own ventures. Reporters have gone to Xinjiang, interviewed witnesses, and received leaked documents themselves; the sources of information are far more diverse.

Evidence for hypothesis 1 I have seen on Twitter, besides repeated assertions that "this is just like Iraq", is either: a) Buzzfeed is saying things China doesn't like, so it must be conspiring with the State Department; or b) the satellite investigation was funded partially by the Open Tech Fund, a nonprofit backed by the US government (as Buzzfeed discloses in the opening lines of the relevant articles).

Now let's consider hypothesis 2! Evidence in support is first, anecdotal acknowledgement that these tiles did exist (see e.g. this thread for the input of myself, an open-source Chinese missile analyst, and another respectable member of the Twitterati—hardly conclusive, but still). More importantly, though, are references to the phenomena by others who deal with maps and whatnot in China. Isaac Stone Fish, for instance, wrote in 2013:

💬

Chinese tech companies provide their own free mapping apps; leading search engine Baidu, for example, has one, and Chinese map firm AutoNavi has an app that provides high-end navigation services. On the one hand, both take pains to expand China’s contested borders to include places like the disputed Diaoyu Islands (the Senkakus to the Japanese). But some Chinese military sites are simply blank spaces, and even Baidu’s detailed Beijing map has an odd black hole: Zhongnanhai, the seat of the Chinese government, for which the map lists only the names of the two lakes in the compound.

Map World, a Google Maps rival launched by the Chinese government in 2010, also did this, as noted by the Financial Times and Maclean's in 2010. (That bit about Zhongnanhai, it seems, is still true; the right side of the road in the image below is part of the Zhongnanhai complex, and the left is regular old Beijing. Satellite imagery of the right side shows it is filled with buildings, though none are labelled. A grey tile right smack in the middle of Beijing, I guess, would have been two conspicuous. But I digress.)

image

Buzzfeed writes in its articles that it contacted Baidu Maps and others for comment multiple times, as is standard journalistic practice, before publication. Hypothesis 2 then can be more fully expressed as "Buzzfeed is telling the truth; the tiles were there and taken down after Buzzfeed contacted Baidu, alerting it of the ongoing investigation." It's your choice as to which hypothesis you think has the strongest support.

So it is at least plausible that such censorship of sensitive sites would exist. My guess is the blank spaces were eliminated at some point after 2013 but for whatever reason accidentally kept at z = 18 until notified by Buzzfeed of their existence.

Conclusion

I am personally fairly confident in the analyses by ASPI, Zhang, and Buzzfeed, but I would not be surprised if there are errors. Such is the natural result when independent verification on the ground—outside of tightly choreographed media tours—is not possible. That does not excuse any potential sloppiness, of course, but I have yet to see that. Beware of Twitter hordes picking one or two cases of less-than-clear evidence and generalizing it to imply all satellite imagery is fake/bad/CIA propaganda; refusing to address the core argument presented by the above investigations is a form of cherry picking.