Guide: Keeping pace with the quarantine facilities popping up across China

A viral video caught my attention last month. What struck most in the video was not the men with medical suits on an athletics track, but the prefabricated housing units, arranged in neat straight rows, spanning the full length of the small stadium.  

After watching several other videos of these alleged quarantine facilities under construction across China,  I decided to verify the claims made. As some people wondered how I did this, I wrote this guide to show and lead you through the steps that could help find and verify such facilities in China. 

Although following all four steps is hardly necessary in most cases, I’ve included extra detail to make it easier to find the more difficult ones. At the end, there’s an additional section to address the PRC coordinate ‘shift problem’.

1. Determining provenance and collecting claims

Like with many investigations, the first step is finding the source. This might provide valuable insight or even answer questions right away. 

In this case, it helps that many of the circulating videos contain visual clues from the source – such as logos from media outlets or social media from China. You might recognize one as TikTok, or more precisely it’s completely separate Chinese counterpart Douyin (抖音). The other logo is from the video-sharing app Kuaishou (快手). Fortunately for us, both of these services have a searchable web interface, allowing some searching without the need to log in. 

Image 1

Left: Still image taken from footage originating from Douyin. Right: Still image from footage originating from Kuaishou.

The numbers visible in some videos can be typed over, but if your mastery of the Chinese language is as poor as mine, Chinese characters are challenging. A better choice then is to extract the text. Various ways of doing that include extracting still images, or pasting several screenshots in one image, and uploading to a service like i2OCR. If you are running a recent version of an Apple build OS, you can use Live Text, which supports Chinese and a handful of other languages. 

Gif 1

Extracting text from imagery with Live Text

After extracting keywords, we can start searching. For the aforementioned reason, we cannot simply assume that Douyin works the same way as TikTok. Searching the username is not very successful, for example. However, Douyin videos also display a unique Douyin user ID, and searching this ID does return the associated user. This brings us one click away from the source video, unless the video is not visible on the profile page without logging in. If so, we can add an extracted keyword in the search field to search again and see if the video shows in the results.

Image 2

Using the user ID and keywords to search Douyin for the source video.

Downloading the video for further analysis is possible by searching “<video” in the source code. This element contains one or more hyperlinks that can be opened in a new tab – choose the last one if it shows several – and subsequently be downloaded via right-click. 

Image 3

Video link in source code

With Kuaishou, we can follow the same steps as with Douyin for searching the source and downloading the video.

Translating the video description and comments, or analysis of other videos uploaded by the same user might be sufficient to continue with the next step, though we could search additional imagery if desired in due time. For example, querying Kuaishou and search engines Baidu and Sogou for “隔离” (quarantine), with one suspected city name, returns numerous new videos.

Image 4

Search results for "隔离" (quarantine) and "紹興市" (Shaoxing) on various platforms.

Now, what if the video lacks any captions or logo’s? As an alternative, we can extract still images, and reverse image searching those with or

Image 5

Baidu reverse image search results of a still image from one of the videos.

2. Analysis of the video for clues of the location

Not all videos reveal as much of the surrounding area, though even small details might help to narrow down the search area. Does the environment look remote or urban? Any mountains visible in the background? Particular objects or buildings that might be visible on satellite imagery? 

Image 6

The camp on the left seems to be located close to some kind of tower or temple building. On the right, we are looking for facilities close to an urban area with high-rise buildings and a green strip in between.

If we discover distinctive elements, we can try searching for these elements at or nearby the city of interest, to further reduce the search area in the next step. River or mountain ranges can be matched with the area on satellite imagery. In practice, some going back and forth between this and next step may be necessary in some cases.  

3. Let’s look at satellite imagery—but don’t zoom in yet

Now that we have a search area, we can start looking at satellite images. As many facilities still seem under construction, we need imagery with recent captures. A good free tool for this purpose is Sentinel Hub’s EO Browser. This service requires an account for some features and has more learning curve than its sibling Sentinel Hub Playground, but will prove its worth.

The hypothesis is this: most of these facilities have a distinctive profile – thanks to white or blue roofing – and scale, to such an extent that identification is possible at relative low zoom levels. If this is correct, we can search large-scale areas at once. 

We will compare the most recent image that has clear views of the area, with an image captured about one year ago. This is not only long enough to see clear changes, it also avoids getting fooled by seasonal changes. To achieve this, first select the last couple of weeks under time range in the Discover tab. Thereafter, select the most recent image that isn’t too cloudy. The selected image can also be used to reduce the search area with the findings from previous steps – or this can alternatively be done with tools like Google Earth.

Image 15

Selecting a time range and non-cloudy image in Sentinel Hub.

Next, add this image to the compare tab as shown below, and repeat these steps to add an image with a date of about a year ago. Finally, uncheck the labels in the layers in the top-right corner for a clean view.

Image 7

The Compare button and tab.

Now we are ready to start the comparison. Open the compare tab, and use the slider to slowly slide the recent capture over the older capture. As we are looking for the facilities pictured in step 2, the one with the green strip, we must especially pay attention to areas close to the city. But don’t zoom in yet, to avoid repetition. The image below shows this juxtaposing, revealing the appearance of a white block that requires further scrutiny.

Gif 2

Comparing satellite images with Sentinel Hub.

A closer look shows a place that bears resemblance with the image from step 2; a location close to an urban area with a green strip in between and an indentation filled with trees. It’s time to verify our findings.

Gif 3

The same comparison zoomed-in.

4. Standard practice: verification

What remains is standard practice. To verify that the identified site is the same location as on footage – instead of some random construction site, we can move on to cross-reference geographic and urban details seen on video with reference imagery. While satellite imagery with better spatial resolution can be of value here, there’s no need to go over much detail this time, as juxtaposing shows clear similarities. Matching a few distinctive buildings demonstrates a clear match.

Image 8

The identified construction site next to reference satellite imagery. | Maxar

Going backwards in Sentinel Hub through time, show whether the site is a recent construction as claimed. To demonstrate the period of construction, slick time-lapse videos can be made with the EO-Browser. How to make such time-lapse is explained in this video. The time-lapse for this location shows it was built in December 2021, in only a few weeks time.

Image 9

Time-lapse showing the construction. | Courtesy of Sentinel Hub

Dealing with the PRC coordinate system offset 

Anyone who has ever used Google Maps in China, is probably familiar with the offset or shift in coordinate systems. Google Maps shows this clearly, as it uses the PCR mandated GCJ-02 system for its maps of China and the common WGS-84 for the satellite layer. As the direction and distance of the offset differs per location, side by side use of different services can get frustrating. To worsen it, some mapping applications like Baidu maps use yet another coordinate system with an additional offset.

Image 10

The two coordinate systems visible in Google Maps at Shanghai.

Another side effect is the inability of Chinese mapping applications to properly handle coordinates. There’s no easy way for either retrieving as interpreting coordinates from a specific location.

Few solutions are available to deal with some of these issues. Scripts are available for the transformation of coordinates, and maps offer API’s to handle coordinates. But there’s also another, maybe easier way, to switch from one service to another one that uses a different system. Let’s have a look and walk through an example of switching from Baidu maps to Google maps.

When browsing Baidu maps, the numbers in the address bar look suspiciously similar to coordinates. But they aren’t. Still, these numbers can be used to get the coordinates of the location. As these numbers represent the coordinates at the centre of your screen, zoom in on the location of interest before copying the numbers from the URL.

Image 11

The numbers in the URL represent the location of centre screen.

For the next step, we need to use Baidu’s Get Point service to convert these numbers to coordinates. We’ll tick the box next to the search bar to activate the reverse lookup, and search for the numbers we copied from the URL, which will show our location of interest on the map.

Image 12

Reverse lookup with Baidu's Get Point service.

Clicking the pin on the map, makes the corresponding coordinates appear in the top right box, which we can easily copy to the clipboard. But bear in mind that these coordinates are BD-09 coordinates. Using these in other mapping services will still show an offset location.

Image 13

Retrieving the coordinates from Baidu's Get Point Service.

To make the step of transforming the coordinates more accessible, I’ve created a front-end, available here, which uses the eviltransform script to switch between the different systems. It works as follows: (1) Choose the coordinate system of origin, (2) paste the coordinates in the coordinates field, or in the latitude and longitude fields underneath if copied separately and click Split or Transform as required, resulting in (3) transformed coordinates that can be copied or opened in a corresponding mapping application right away. It’s also possible to go the other way around, or convert to GCJ-02.

Image 14

Steps to transform coordinates to another system.

In this guide, we went through the steps of finding videos on Chinese social media, using Sentinel Hub to spot changes in infrastructure, and switching between mapping services with different coordinate systems. As investigation work in China comes with challenges, I hope that this guide will be of help for anyone dealing with these and other issues addressed.

Further reading:
Environmental monitoring of conflicts using Sentinel-2 - Wim Zwijnenburg
A short guide to Chinese coordinate system