Data Journalism on Visualizing.JP

Drawing Shelling Locations in Two Ways by Certainty: TEXTY's "Under attack"

Wed, 04 Nov 2026 00:00:00 +0900

TEXTY drew the locations of Russian shelling, gathered from news reports, in two ways. Locations backed by NASA fire detections or by confirmed eyewitness reports are small, sharp dots. Locations where the report named only a settlement, so the settlement’s coordinates stand in, are large, blurred dots. In the drawing code, the former have a radius of 4 and a blur of 0; the latter have a radius of 8 and a blur of 1.5.

The work is “Under attack. What and when Russia shelled in Ukraine” by TEXTY.org.ua, an independent media outlet in Kyiv. The Ukrainian version came out on August 18, 2022, and the English version on September 1. It was shortlisted for the Sigma Awards 2023. What follows draws on the work’s text and its METHODOLOGY section, the drawing code written into the page, and the author’s conference talk.

February 24, 2022: It Starts with a Small Spreadsheet

The work was made by three people: Petro Bodnar, Yevhenia Drozdova, and Nadja Kelm. In 2023, Kelm presented this map at “Visualizing Knowledge,” a conference hosted by Aalto University. TEXTY published a transcript of the talk on June 27, 2023. According to the transcript, the theme of that year’s conference was “Visualizing Uncertainty.”

In the talk, Kelm describes the beginning like this.

“On February 24, 2022, we ran a quick search for news about the first missile strike. That information was transformed into a map, and that was the final goal. However, the strikes continued on the next day and the day after that. The small spreadsheet started to grow.”

On the first day of the invasion, they searched for news of the first missile strike and put it on a map. That was supposed to be the end of it. But the attacks continued the next day and the day after, and the small spreadsheet began to grow.

As the spreadsheet grew, so did its columns: one for the type of attack, and one for the certainty of the coordinates.

“I added extra columns to record various types of strikes (missiles, artillery, air, and, beginning autumn 2022, drones). Besides I have to identify the precision of coordinates since reports often mention only the name of a town or even of a region.”

She added columns for the types of strikes (missiles, artillery, air strikes, and, from autumn 2022, drones). She also had to identify how precise the coordinates were, because reports often give only the name of a town or even of a region.

A spreadsheet shown on a talk slide. Columns for place name, latitude, longitude, date, precision, source link, region, and type are lined up, and the precision column continues with “неточні”

Source: Nadja Kelm / TEXTY.org.ua, “Data visualisation during wartime.” TEXTY.org.ua, CC BY 4.0 (as stated on the site)

Kelm says she collected this data by hand every day for half a year. In the summer of 2022, Planet.com let the team use its database of high-resolution satellite imagery. Combining the spreadsheet with images of destroyed towns into an interactive map produced this work.

Checking Place Names in Reports Against a Fire Map

The basis for splitting the dots in two is written in the work’s METHODOLOGY section. In the first few weeks of collection, the team watched several news sites. There were many duplicates, so afterward they watched only the news feed of Ukrainska Pravda (www.pravda.com.ua/news).

Locations picked up from reports are checked in the following steps.

Write out the names of all the settlements that were shelled.
Compare them with the fire map for that date.
If a fire is found, record the fire’s coordinates in the spreadsheet and mark the entry “accurate.”
If there is no fire and the report gives no detailed description of the location, use the settlement’s coordinates from Google Maps and mark the entry “inaccurate.”

“If we saw a fire that day, we indicated its coordinates in the table and marked the entry as accurate. In most instances, fires recorded on NASA satellite maps corresponded with the locations of shelling.”

In short: if a fire was visible that day, they recorded its coordinates and marked the entry “accurate,” and in most cases the fires on NASA’s satellite maps matched the locations of shelling.

Information from confirmed eyewitnesses (friends and relatives) was also treated as “accurate” in most cases, even when it had no source link. In the talk, Kelm reveals that the fire map is NASA’s FIRMS (Fire Information for Resource Management System). She also says that long-burning fires, such as those at large factories, have to be excluded.

A NASA FIRMS screen. Red fire-detection dots are scattered over a map of eastern Ukraine, and a table with the latitude, longitude, and observation time of one location is open

Source: Nadja Kelm / TEXTY.org.ua, “Data visualisation during wartime” (the screen is NASA FIRMS)

The work’s text is also clear about its limits. Shelling that did not appear in the news is not included. The daily list of settlements published by the General Staff of Ukraine has not been included since the first day of the invasion. The work states that the map and charts do not reflect reality 100%.

Small Sharp Dots, Large Blurred Dots

This “accurate / inaccurate” distinction becomes the way the dots are drawn on the map. The work is built with Mapbox GL JS, and the drawing code is written directly into the page. The values below are read from that code. I did not operate the map in a browser to check the rendering.

Data / variable	Appearance (visual variable)	Specific value (from the code)
Position of one shelling	Position of a dot on the map	Mapbox circle layer. Shown only below zoom 11
Type of attack (4 types)	Color of the dot	Shelling `#E45C04` (orange), air strike `#C14DB2` (magenta), rocket `#901B1B` (dark red), drone `#003FFC` (blue)
Certainty of coordinates	Blur and size of the dot	Accurate: radius 4, blur 0. Approximate: radius 8, blur 1.5
Date	Time slider below the map	Start fixed at February 24, 2022. Shows the cumulative total up to the selected date
Occupied territory	Color of the area	Before 2022 `#A69B99`, from 2022 `#DCB5B6`. The latter changes with the slider’s date
Places with satellite imagery	Light-blue rectangle, then polygon, then satellite image	Rectangles at zoom 4.5 to 6.9, polygons of the imaged area at 7 to 11, Planet imagery at 11 and above
Details of one incident	Popup opened by clicking	Place name, date, exact / approximate, source link

What determines certainty is the spreadsheet column “координати.” If the value is “неточні” (inexact), the dot is drawn with a radius of 8 and a blur of 1.5. Otherwise, it is drawn with a radius of 4 and a blur of 0. Color is set by type and is independent of certainty. Orange shelling dots, for example, appear both as sharp dots and as blurred dots.

The legend explains this distinction only in terms of blur.

“Clear outline = exact coordinates / Blurred outline = approximate coordinates”

A clear outline means exact coordinates, and a blurred outline means approximate coordinates. The legend’s sample is a gray circle with a CSS blur applied. The legend does not say that the dot’s radius doubles.

A slide from the author’s talk. The top shows a map of all of Ukraine with the legend and shelling dots; the bottom shows a zoomed-in screen that has switched to satellite imagery

Source: Nadja Kelm / TEXTY.org.ua, “Data visualisation during wartime.” Satellite imagery from Planet.com (via TEXTY; whether it falls under CC BY is unknown)

The presentation also changes with the zoom level. While viewing the whole country, places with satellite imagery are shown as light-blue rectangles. As you move closer, they become polygons of the imaged area, and past zoom 11 the shelling dots disappear and the Planet imagery itself is overlaid. In the talk, Kelm describes this design as letting you zoom “from the level of abstract points to the level of specific consequences.”

“On the map you can choose the type of shelling, move along the timeline and zoom from the level of abstract points to the level of specific consequences.”

The work’s cover image. A screen combining satellite imagery of a town with a map scattered with orange and magenta dots

Source: Petro Bodnar, Yevhenia Drozdova, Nadja Kelm / TEXTY.org.ua, “Under attack.” Satellite imagery from Planet.com (via TEXTY; whether it falls under CC BY is unknown)

Blurred Dots Make Up Most of the Map

So how many blurred dots are there? I tallied the public Google spreadsheet that the map loads. The counts here are from the public version retrieved on November 4, 2026. They differ from the data at the time of the Sigma Awards judging (early 2023).

Point in time	Rows with a place name	Approximate coordinates	Share
Through August 31, 2022 (around the English version’s release)	9,588	7,183	about 75%
Through December 31, 2022 (around the Sigma entry)	16,474	13,911	about 84%
Whole dataset retrieved November 4, 2026	48,623	44,908	about 92%

In the public version, over 90% of the dots have “approximate coordinates.” The code suggests that when you open the map, most dots are drawn large and blurred. The share of approximate coordinates rises over time. Neither the work nor the talk gives a reason for this.

Let us set this beside research on visualizing uncertainty. An experiment by MacEachren, Roth, and colleagues, published in IEEE TVCG in 2012, compared ways of expressing uncertainty with map symbols.

“Fuzziness and location work particularly well; value and arrangement are also rated highly and both size and transparency are potentially usable.”

The conclusion: fuzziness and location work particularly well, value and arrangement are also rated highly, and size and transparency are potentially usable. The same paper reports which directions read intuitively. For fuzziness, the blurrier, the more uncertain; for size, the smaller, the more uncertain.

TEXTY’s design agrees with this finding in the direction of blur. The direction of size is reversed: uncertain dots are drawn larger. I read TEXTY as using size to mean the breadth of the possible area. This is the same idea as a GPS error circle, showing “somewhere in this settlement” as a spread of the dot. Read that way, it makes sense, but it departs from the experimental finding. And because the legend does not mention size, I suspect readers may take a large dot to mean “heavy damage.”

How Did Third Parties See This Map?

The work was shortlisted for the Sigma Awards 2023. The entry form describes the map’s limits in the same words as the work’s text.

“This project gives a general overview and does not aim to show absolute accuracy since some instances of shelling were not mentioned in the news we processed.”

It gives an overall picture and does not aim at absolute accuracy, because some shelling did not appear in the news they processed.

On the maps blog Google Maps Mania, Keir Clarke covered the work on August 19, 2022, the day after the Ukrainian version was published. After describing the color-coded dots, the links to news reports, and the zoom into Planet satellite imagery, he writes about the time slider like this.

“You can adjust the end-date on this timeline to view the cumulative amount of shelling from the beginning of the Russian invasion until the selected date.”

Moving the end date shows the cumulative shelling from the start of the invasion up to the selected date. The work’s text says “choose a specific time range,” but in the code the start date is also fixed. Clarke’s wording, it can be read, is the one that matches the implementation.

On the personal blog ix5.org, “Useful Ukraine databases and maps” lists “Attack map by Texty.org.ua” alongside the GeoConfirmed and DefMon3 maps. There is no evaluative sentence, only the link. This suggests the work is treated not only as journalism but as a practical tool for understanding the situation.

Uses

What we can take away from this work is that, at the data-collection stage, certainty was recorded as a column. In the talk, Kelm describes the purpose of data visualization as “bringing at least a little order to chaos.”

“We aren’t trying to restore a sense of security, but rather try to restore the ability to navigate the situation.”

The words mean: the aim is not to restore a sense of security, but to restore the ability to find one’s way through the situation.

To show uncertainty, you first need a procedure for measuring it. TEXTY recorded in a spreadsheet column whether a location was backed up (a matching fire, a confirmed eyewitness report). Because that column has only two values, accurate and inaccurate, the map’s distinction is also binary. That is how I read it.

On the other hand, if you layer two variables to tell the dots apart, it is safer to describe both in the legend. TEXTY changes blur and size at the same time but explains only blur in the legend. Together with the fact that the direction of size is opposite to the experimental finding, I suspect that either distinguishing the dots by blur alone, or matching the legend’s samples in size as well, would reduce misreading.

References

Primary sources

Petro Bodnar, Yevhenia Drozdova, Nadja Kelm, “Under attack. What and when Russia shelled in Ukraine,” TEXTY.org.ua, 2022-09-01 (English version; main text, METHODOLOGY section, legend, drawing code). https://texty.org.ua/projects/107577/under-attack-what-and-when-russia-shelled-ukraine/
“Під ударом. Що й коли обстрілювала Росія,” TEXTY.org.ua, 2022-08-18 (Ukrainian version, the original). https://texty.org.ua/projects/107440/pid-udarom-sho-i-koly-obstrilyuvala-rosiya/
Public data loaded by the map (Google Sheets CSV, retrieved 2026-11-04). https://docs.google.com/spreadsheets/d/e/2PACX-1vSyFuA7nxANnn7BwXn7az5D5L-V7yKnETgTybfKSIGmoYz2qVkc6FWSH7f0l-1Gt_dML1VpywPUzXwp/pub?gid=1376631421&single=true&output=csv
Location data for satellite imagery. https://texty.org.ua/documents/313/tif_shapes_rects.geojson / https://texty.org.ua/documents/315/tif_shapes_new_AW4Mn67.geojson
List of GeoJSON files for occupied territory. https://thallium.texty.org.ua/maps/shelling_map/json_list.csv
“Data visualisation during wartime. Nadja Kelm`s speech at the data-journalisn conference in Finland,” TEXTY.org.ua, 2023-06-27. https://texty.org.ua/articles/110009/data-visualisation-during-wartime-nadja-kelms-speech-data-journalisn-conference-finland/
Sigma Awards 2023 entry page, “What and when Russia shelled in Ukraine” (shortlisted; data courtesy of The Sigma Awards / GIJN, CC BY-NC-SA 4.0). https://sigma-awards.data-visualization.jp/entries/2023-s-0017/

Secondary sources

Keir Clarke, “Ukraine Under Attack,” Google Maps Mania, 2022-08-19. https://googlemapsmania.blogspot.com/2022/08/ukraine-under-attack.html
“Ukraine War – Useful Ukraine databases and maps,” ix5.org. https://ix5.org/thoughts/2022/ukraine-war-useful-tools-guides/
GIJN, “Data Journalism Top 10” (said to be the issue that covered TEXTY’s 2022 work; the text could not be read because of a bot check, so this is unconfirmed). https://gijn.org/stories/data-journalism-top-10-a-year-of-war-in-ukraine-russian-casualties-trans-identities-unexpected-nba-performances/
(Related research) Alan M. MacEachren, Robert E. Roth, James O’Brien, Bonan Li, Derek Swingley, Mark Gahegan, “Visual Semiotics & Uncertainty Visualization: An Empirical Study,” IEEE Transactions on Visualization and Computer Graphics, 18(12), 2012. https://geography.wisc.edu/cartography/projects/publications/MacEachrenEtAl_2012_TVCG.pdf

How Many Mbps Were You Offered at the Same Price? The Markup's "Still Loading"

Wed, 04 Nov 2026 00:00:00 +0900

The Markup collected, address by address, the internet speeds that four ISPs (AT&T, Verizon, EarthLink and CenturyLink) offered at the same price. It then compared “the share of addresses offered under 25 Mbps at the same price” across groups defined by income, share of white residents, and 1930s redlining grades. The data covers 848,618 offers in 38 US cities.

The main article, “Dollars to Megabits, You May Be Paying 400 Times As Much As Your Neighbor for Internet Service,” was published on October 19, 2022, in partnership with AP. The same day, the methodology (“Show Your Work”) and a guide for reporters (“Story Recipe”) came out too. The package won a Sigma Awards 2023 Winner prize. Below, I read the main article, the methodology, and the code published on GitHub, following the order of the graphics.

When the price is flat, speed is what you compare

The starting point is how the pricing works. According to the methodology (Show Your Work), NDIA pointed out in a 2018 report, drawing on the ISPs’ promotional materials, “tier flattening”: charging the same base price for different speeds. The Markup checked this with offers at individual addresses. The methodology puts it this way.

Our analysis differs from previous work by examining pricing in conjunction with speed and the demographics of areas that are offered the same price for worse service (below 25 Mbps)—we call these the ‘worst deals.’

In short, the difference from earlier work is that it examines price together with speed and the demographic makeup of areas, and it calls the cases where the same price buys under 25 Mbps “worst deals.”

The data source is the address lookup tool on each ISP’s website. According to the methodology, the team followed a 2020 Princeton University study and queried the underlying APIs automatically. The period was April 15 to October 1, 2022, with 1,606,964 queries. Addresses were drawn at random, 10% per block group. The speeds are those the ISPs offered, not measured speeds.

Speeds were split into four bands, from under 25 Mbps (Slow, which falls short of the FCC’s definition of broadband) to 200 Mbps and above (Blazing). Income and share of white residents were each divided into quartiles of block groups within each city. For historical grades, according to the methodology, addresses were overlaid on the HOLC (Home Owners’ Loan Corporation) maps published by Mapping Inequality. Only the 22 cities for which maps exist are covered.

The “400 times” in the headline comes from converting prices to price per Mbps. According to the methodology, CenturyLink works out to $100/Mbps at 0.5 Mbps and $0.25/Mbps at 200 Mbps, a 400-fold difference. The main article also has a graphic that connects each ISP’s cheapest and most expensive per-Mbps price with a line segment. However, the graphic’s data and the methodology’s text partly disagree. AT&T’s cheapest price is $0.13 in the graphic and $0.18 in the text. CenturyLink’s values match. The reason for the discrepancy cannot be confirmed.

The opening pair of houses: you can’t compare with your neighbor yourself

The main article opens with two houses in Kansas City. Two homes 0.4 miles apart were offered 300 Mbps and 5 Mbps by AT&T for the same $55 a month.

An illustration in which a bundle of thin dashed lines flows into a blue-gray house, while only a single red dashed line reaches a red house

Source: Gabriel Hongsdusit and Joel Eastwood / The Markup, “Dollars to Megabits”

This is the article’s key visual. The opening graphic is built with HTML and SVG and is read by scrolling down. To the fast house, 41 thin blue-gray wavy lines flow on a 12-second cycle. To the slow house, only one thick red wavy line flows slowly on a 50-second cycle. Speed is drawn through the number, thickness, and flow rate of the lines.

Next, the numbers appear with the two houses kept side by side. They come in the order of speed, distance, median household income and share of white residents, and monthly price, and finally a map of the 38 US cities.

Item	Fast house	Slow house
Offered download speed	300 Mbps	5 Mbps
Median household income	$63,214	$23,836
White residents	70%	1%
Monthly price	$55	$55

The two houses were not found by hand. The notebook on GitHub uses scikit-learn’s BallTree to find, for each address, the nearest address that was offered 200 Mbps or more. Julia Angwin, the editor-in-chief, wrote in her newsletter:

Very few people are likely to enter other addresses into the site to compare speeds that their neighbors are getting

That is, hardly anyone will type in other addresses to compare the speeds their neighbors get. Opening with two houses side by side can be read as a response to this problem: you can’t compare with your neighbor yourself.

Most of the graphics after the opening use a single metric: the share of addresses offered under 25 Mbps at the same price. What changes is how the groups are divided and how many cities are shown.

Graphic	Data	Visual encoding
Opening graphic	Speed, income, share of white residents, and monthly price for two addresses	Speed as number, thickness, color, and flow rate of lines. Houses in blue and dark red
Price per 1 Mbps	Cheapest and most expensive per-Mbps price for each ISP	Line segments. Circle at the cheapest end in blue-gray, at the most expensive end in red
Map by city	Gap between the shares in low-income and high-income areas	One same-sized square per city. Gap shown in four color steps
Las Vegas bars	Composition of the four speed bands within each income quartile	100% stacked bars. Only Slow in red; the faster, the darker the blue-gray
Scatter plot by city	Share in the areas with the lowest share of white residents (x) and the highest (y)	Position of points. Dotted lines at ±5 points on both sides of the diagonal
Kansas City maps	Offers by address and HOLC grades	Dots for addresses over a background colored by grade. Two maps: slow addresses and the fastest addresses

The same red (#FF335F) is used for slow speeds in every graphic. Even in stacked bars with many bands, red is the only warm color. This can be read as telling the reader that finding the red finds the worst deals.

The graphic split by income is the map by city. From the gap boundaries at +5, +10, and +15 points, the color deepens from beige to red.

A US map with a square for each city; most squares are deep red, and only three cities are beige

Source: Joel Eastwood and Paroma Soni / The Markup, “Dollars to Megabits”

A note points to Oklahoma City, saying that the share in low-income areas (53.3%) is more than seven times that in high-income areas (7.1%). The graphic split by share of white residents is a scatter plot.

A scatter plot of points, one per city. Most points lie below and to the right of the diagonal; New Orleans sits at 77% on the horizontal axis and 45% on the vertical axis

Source: The Markup / The Markup, “Dollars to Megabits”

The dotted lines on both sides of the diagonal carry the note “Within this band, the difference was 5 percentage points or less.” The methodology’s criterion is as follows.

We consider a disparity to exist when the difference of the proportion of slow speeds between any of the comparisons is five percentage points or more.

That is, a disparity is deemed to exist when the share of slow speeds differs by five points or more between groups. No p-value test is used. The dotted band on the scatter plot can be read as this criterion drawn directly onto the graphic. The further a point lies to the lower right outside the band, the more the city disadvantages areas with fewer white residents. According to the methodology, worst deals were concentrated in low-income areas in 92% of cities, and in areas with fewer white residents in 66% of cities.

Splitting the Kansas City map in two

The third way of grouping is the 1930s redlining grades. HOLC graded neighborhoods in four tiers: Best, Desirable, Declining, and Hazardous. According to the methodology, in all 22 cities, worst deals were concentrated in redlined areas.

The main article places AT&T’s offers in Kansas City on a map, address by address. The left map is “Addresses offered slow speed,” and the right is “Addresses offered blazing speed.”

A Kansas City map colored by HOLC grade, with addresses offered slow speeds shown as downward-pointing triangles, densely clustered in the Hazardous areas to the north

Source: Joel Eastwood (Source: Mapping Inequality) / The Markup, “Dollars to Megabits”

The same map extent, with addresses offered 200 Mbps or more shown as upward-pointing triangles, clustered in the Best and Desirable areas to the southwest

Source: Joel Eastwood (Source: Mapping Inequality) / The Markup, “Dollars to Megabits”

In the background, Best is green and Hazardous is a muted red. Slow addresses are drawn as downward triangles and the fastest addresses as upward triangles. On the slow-address map the Hazardous areas are shaded more darkly, and on the fastest-address map the Best and Desirable areas are. Because the two kinds of dots are split across two maps rather than overlaid on one, the places where the dots cluster appear to swap between left and right.

The note reads “68% of addresses in redlined areas were offered slow speeds.” “redlined areas” is underlined in the same color as Hazardous, linking the legend to the text. In the methodology, Slow accounts for 68.5% in areas graded D and 4.2% in areas graded A. Reporter Aaron Sankin wrote on X:

The 1939 notes explaining the decision to redline a Kansas City, Mo., neighborhood cited ‘Negro encroachment from the north.’ More than 80 years later, we didn’t see AT&T offer a single connection there that the FCC would consider ‘broadband.’

That is, the 1939 notes explaining the decision to redline the neighborhood cited “Negro encroachment from the north,” and more than 80 years later, AT&T offered not a single connection there that would count as broadband.

How third parties saw the investigation

The Sigma Awards 2023 judges singled out that the investigation put a double disadvantage into numbers.

It showed the injustice didn’t stop there: slow internet connections cost the same as fast ones, meaning communities are disadvantaged twice.

Because slow connections cost the same as fast ones, communities are disadvantaged twice, the judges say. The price per Mbps can be read as folding this doubleness into a single unit price.

The investigation took first place in the 2023 Philip Meyer Journalism Award (IRE/NICAR). The judges’ points concern method and tools.

The reporters adapted methods from an academic study to identify internet offers by address and then used Census data and historical maps to tell a powerful story about a critical social injustice. The judges applaud the team for their resourcefulness, robust validation process and, along with their partner Big Local News, commitment to sharing their bespoke mapping tool with the public.

In short, the reporters adapted methods from an academic study to identify offers by address and told the story with Census data and historical maps; the judges praise the resourcefulness, the validation process, and the public release of the bespoke mapping tool.

From the academic side, a limit in scope was pointed out. The ACM SIGCOMM 2023 paper by Udit Paul et al. treats this investigation as the closest prior work and writes:

The Markup study does not include cable-based ISPs, which serve most of the US population.

That is, the study omits the cable-based ISPs that serve most of the US population. The paper compares plans in Mbps per dollar, the reverse of The Markup’s unit. This can be read as The Markup choosing the price paid per unit as its axis, and the paper choosing the value received.

What to take from it

First, let the system decide the unit of comparison. When the price is flat, price differences make no graphic. The Markup held the price fixed and counted speeds, then converted to a unit price. What you hold fixed and what you count determines the graphic’s metric.

Second, keep the metric to one. Whether scatter plot, symbol map, stacked bar, or address map, the form changes but what is counted stays “the share of addresses offered under 25 Mbps.” Red also always means slow speed. The reader only has to follow the difference in grouping.

Third, draw the criterion for judgment into the graphic. The ±5-point band on the scatter plot has the same width as the methodology’s criterion. Because it is drawn as lines, readers can check for themselves which cities exceed the criterion.

Fourth, show first the comparison readers can’t make themselves. The opening two houses were chosen by an algorithm that finds nearby addresses. Before the story of the whole, it places a single pair: 300 Mbps and 5 Mbps for the same $55.

Finally, publish the means of reproduction together. The Story Recipe released per-city CSVs and per-ISP Kepler.gl maps, and it led to nine local-outlet stories. According to the application materials, a Detroit city official said the city would use this data in deciding the order in which to lay its municipal fiber network.

References

Primary sources

Leon Yin, Aaron Sankin, “Dollars to Megabits, You May Be Paying 400 Times As Much As Your Neighbor for Internet Service,” The Markup, 2022-10-19: https://themarkup.org/still-loading/2022/10/19/dollars-to-megabits-you-may-be-paying-400-times-as-much-as-your-neighbor-for-internet-service
Methodology, Show Your Work, “How We Uncovered Disparities in Internet Deals,” 2022-10-19: https://themarkup.org/show-your-work/2022/10/19/how-we-uncovered-disparities-in-internet-deals
National Digital Inclusion Alliance (NDIA), report on tier flattening (PDF, July 2018; referenced from the methodology): https://www.digitalinclusion.org/wp-content/uploads/2018/07/NDIA-Tier-Flattening-July-2018.pdf
David Major, Ross Teixeira, Jonathan Mayer, “No WAN’s Land: Mapping U.S. Broadband Coverage with Millions of Address Queries to ISPs,” ACM IMC 2020 (the study the methodology followed): doi:10.1145/3419394.3423652 https://doi.org/10.1145/3419394.3423652
Story Recipe, 2022-10-19 (updated 2023-03-16): https://themarkup.org/story-recipes/2022/10/19/journalists-investigate-which-neighborhoods-in-your-city-are-offered-the-worst-internet-deals
GitHub repository the-markup/investigation-isp: https://github.com/the-markup/investigation-isp (README: https://raw.githubusercontent.com/the-markup/investigation-isp/main/README.md , Notebook 5: https://github.com/the-markup/investigation-isp/blob/main/notebooks/5-find-closest-fiber-address.ipynb )
Kepler.gl map (AT&T): https://markup-public-data.s3.amazonaws.com/isp/at%26t-kepler.gl.html
Datawrapper (main article graphics): https://datawrapper.dwcdn.net/LK8HK/11/ / https://datawrapper.dwcdn.net/g6Lw9/7/ / https://datawrapper.dwcdn.net/yv3Vw/14/ / https://datawrapper.dwcdn.net/YXy1q/8/ / https://datawrapper.dwcdn.net/eFcrH/18/
Datawrapper (methodology graphics): https://datawrapper.dwcdn.net/3AUkE/6/ / https://datawrapper.dwcdn.net/mKzuZ/3/ / https://datawrapper.dwcdn.net/4WZtw/6/ / https://datawrapper.dwcdn.net/m0AU9/6/ / https://datawrapper.dwcdn.net/DX6tS/11/ / https://datawrapper.dwcdn.net/mr1oj/9/ / https://datawrapper.dwcdn.net/n7R75/8/ / https://datawrapper.dwcdn.net/vZUIW/6/
Julia Angwin, “Broadband Bias,” Hello World (The Markup’s newsletter), 2022-10-22: https://themarkup.org/newsletter/hello-world/broadband-bias
Aaron Sankin’s X thread (reposted by Thread Reader): https://threadreaderapp.com/thread/1582739500903456768.html
Follow-up map “See the Neighborhoods Internet Providers Excluded from Fast Internet,” 2023-05-11: https://themarkup.org/still-loading/2023/05/11/see-the-neighborhoods-internet-providers-excluded-from-fast-internet
Reports of awards (Scripps Howard, NABJ): https://themarkup.org/inside-the-markup/2023/10/24/the-markup-wins-scripps-howard-innovation-award / https://themarkup.org/inside-the-markup/2023/08/15/the-markup-wins-nabj-salute-to-excellence-award
Sigma Awards entry (work page on the listing site; the Japanese is machine-translated): https://sigma-awards.data-visualization.jp/entries/2023-s-0006/
Sigma Awards official 2023 Winners: https://www.sigmaawards.org/tag/2023-winners/
AP’s co-published version (body text not checked): https://apnews.com/article/01a99247a08b355e89cc54595aecdafa

Secondary sources

Sigma Awards 2023 judges’ comments (original text on the entry page above): https://sigma-awards.data-visualization.jp/entries/2023-s-0006/
IRE, “2023 Philip Meyer Journalism Award” winner announcement and judges’ comments: https://www.ire.org/2023-meyer-winners/
Udit Paul et al., “Decoding the Divide: Analyzing Disparities in Broadband Plans Offered by Major US ISPs,” ACM SIGCOMM 2023: https://arxiv.org/abs/2302.14216 (PDF: https://arxiv.org/pdf/2302.14216 , doi:10.1145/3603269.3604831 https://doi.org/10.1145/3603269.3604831 )

On a Day with a High of 36°C, the Heat Index Reached 69°C: The Hindu's 'Chennai's heat index'

Wed, 04 Nov 2026 00:00:00 +0900

The Hindu calculated a heat index from the readings of sensors carried by four workers. The piece gives the maximum temperature in Chennai, in southern India, on June 9, 2025 as 36°C. On the same day, the heat index for a street-food cook and a fisherman who went out to sea reached as high as 69°C. The sensors were Arduino Nano boards fitted with a temperature and humidity sensor, and the heat index was calculated with the formula of the U.S. National Weather Service (NWS).

The piece is a standalone interactive page published on July 12, 2025 by the Data team of The Hindu, an English-language daily in India. The page’s headline is “TAKING ON THE HEAT,” and its entry name in the Sigma Awards is “Chennai’s heat index: How heat stressed are you?” Below, I read it in the order a visitor sees the page, following the text and code of the piece and its Sigma Awards entry.

On June 9, the high was 36°C

When you open the page, the large red-orange letters “HEAT” appear over a black-and-white photograph of a person drinking water. Below them is a single lead sentence.

The opening of the piece: the headline “TAKING ON THE HEAT” and a lead sentence saying the maximum on June 9 was 36°C, laid over a black-and-white photograph

Source: The Hindu Data team / The Hindu, “Heat index: how heat-stressed are you?”

“The maximum temperature on June 9 in Chennai was 36°C. But, high humidity and the kind of work they did compounded the heat stress for many.”

In short, the high was 36°C, but humidity and the kind of work made the heat stress heavier for many people. With this first line, the article can be read as fixing the maximum temperature as the thing everything else will be compared against.

Next come cards for four people, each with a portrait. A card gives the person’s name, occupation, the highest heat index of the day with its time, and a quote from the person.

Person	Occupation / place	Maximum written on the card
Sathya	Tiffin cook (a street-food cook) at Parry’s Corner in the city center	69°C at 11:30 a.m.
Anand (age 33)	Fisherman who goes out to the Bay of Bengal from Marina Beach	69°C at sea at 12:56 p.m.
Prakash	Factory worker polishing stainless-steel utensils in Korukkupet	About 66°C at 3:10 p.m.
Leon (a pseudonym)	Auto-rickshaw driver	56.8°C at 2:20 p.m.

All four numbers on the cards are more than 20°C above 36°C.

Handmade sensors and the NWS formula

The method is described in the body text after the cards.

“The Hindu Data team used Arduino Nano devices with temperature and humidity sensors to continuously record time-stamped data on an on-device SD card. We handed these devices out to four people from different parts of Chennai.”

The team recorded time-stamped data continuously on the device’s SD card and handed the devices to four people working in different parts of Chennai. The piece says the recording day was June 9, 2025, that there was light rain at 4 p.m., and that this was a few weeks after the city’s heat peak.

How the heat index was calculated is explained in a note at the end of the page. The team used a formula from the U.S. National Weather Service (the article writes “U.S. National Weather Service (NOAA)”) that takes air temperature and relative humidity. The India Meteorological Department (IMD) provisionally divides the value into four levels: below 35°C, 36 to 45°C, 46 to 55°C, and above 55°C. The body text gives an example: 35°C at 85% relative humidity gives a heat index of 60°C. Recalculating it with the NWS formula gives 60.0°C, which matches.

The piece does not hide gaps and outliers in its data. The body text says that there are periods when the devices were switched off by mistake and the data is missing, and that the team filled those periods in through interviews. Sathya’s chart carries a note that unusually high values appeared because the device was placed near the stove. Anand switched his device off at night, and the piece says the nighttime heat index could not be calculated.

Three numbers that change as you scroll

Each of the four people gets one scene that follows their day. A square illustration and three numbers are pinned to the left third of the screen. The numbers are “TIME,” “TEMP,” and “HEAT INDEX.” Paragraphs of text flow through the right two thirds, and when a paragraph reaches the middle of the screen, the illustration and numbers switch.

On a black background, an illustration of a woman wiping her face with the end of her sari, with three numbers, 11:30 A.M., 48.3°C and 69°C, and a paragraph of text on the right

Source: Soumyadip Sinha (Illustrations in the piece’s credits) / The Hindu, “Heat index: how heat-stressed are you?”

The figure above is the scene where Sathya sets down her ladle: 11:30 A.M., air temperature 48.3°C, heat index 69°C. Which values appear with which paragraph is written in the HTML, as the attributes data-time, data-temp, and data-felt on each paragraph.

The table below shows how the data is mapped to visual forms.

Component	Data	Visual form / material	Specifics
Lead sentence	Maximum temperature given by the piece	Text	36°C
Four cards	Each person’s maximum heat index and its time	Portrait + name + one sentence + the person’s words	69 / 69 / about 66 / 56.8°C
Pinned panel: time	The time of the paragraph	Numbers animated like a clock	Counts in steps of 1 to 10 minutes depending on the time difference
Pinned panel: temperature	The temperature of the paragraph	Number + text color (5 levels)	Color changes at 30 / 33 / 36 / 39°C
Pinned panel: heat index	The heat index of the paragraph	Number + text color (5 levels)	Thresholds at 35 / 45 / 55 / 65°C, from #ffe5e5 to #f24405
Pinned panel: image	The scene of the paragraph	Swapping black-and-white line illustrations	Sathya1 to 4, Fisherman1 to 5, and so on
Time-series line	Heat index calculated from the sensor	A single red-orange (#f24405) line on black	Horizontal axis = time, vertical axis = heat index
Line annotations	Peaks, outliers, missing data	White circle and curved arrow + one sentence	“peaked to 69°C at around 1 p.m.” and others

The only place where values are turned into color is the text color of the panel’s temperature and heat index. The color thresholds for the heat index are the IMD’s four levels (35 / 45 / 55°C) with one more boundary, 65°C, added, making five levels. As a result, the deepest red-orange appears only in the scenes of the two people at 69°C and Prakash at 66°C. Leon (56.8°C), who falls in the same “above 55°C” class under the IMD’s division, stays one shade lighter. The extra step can be read as added to show differences among the four.

At the end of each scene, a line chart of that person’s heat index is placed. The line itself is a PNG image, while the axis numbers and annotations are HTML text. HTML comments show that the charts were finished in Adobe Illustrator and ai2html (v0.121.1).

The four charts differ from one another in both time range and vertical-axis range.

Person	Horizontal axis	Vertical axis	Annotations
Sathya	9:00 a.m. to 9:00 a.m. the next day (about 24 hours)	40 to 80 (steps of 10)	Outlier near the stove / “peaked to nearly 69°C at around 11.30 a.m.”
Anand	9:00 a.m. to 6:00 p.m.	30 to 70 (steps of 5)	“peaked to 69°C at around 1 p.m.” (desktop version)
Prakash	10:00 a.m. to 12:00 a.m.	40 to 65 (steps of 5)	“peaked to 66°C in the afternoon”
Leon	9:00 a.m. to 11:00 p.m.	40 to 55 (steps of 2.5)	“peaked to 56.8°C during one of his afternoon rides” / “Missing data”

The peak that reaches near the top of the frame is in the 56°C range in Leon’s chart and above 80°C in Sathya’s. The four are not built to be compared side by side. What lets readers compare one person with another is the numbers in the annotations and the four cards at the start. The charts can be read as cut to show the shape of each person’s day.

A red-orange line on black. A peak above 80 around noon, with a note that the device was placed near the stove and a note for 69°C, each marked with a circle and an arrow

Source: The Hindu Data team / The Hindu, “Heat index: how heat-stressed are you?”

In Sathya’s chart, the line rises past the 80 mark around noon. That peak is circled in white and annotated as the device having been placed near the stove. The “69°C” note points to the lower peak before it. The 69°C on the card can be read as a value chosen with the outlier peak excluded.

A red-orange line on black. A peak around 1 p.m. is circled, with a note that it reached 69°C

Source: The Hindu Data team / The Hindu, “Heat index: how heat-stressed are you?” (desktop version)

Anand’s chart contains an inconsistency. The page switches between two versions depending on screen width. In the desktop version, at widths of 1144px or more, the annotation reads “Anand’s heat index peaked to 69°C at around 1 p.m.” In the version for narrower screens it reads “peaked to 48°C at around 1 p.m.” Both versions have a vertical axis of 30 to 70, and the shape of the line is the same. The top of the line sits just under the 70 mark, which fits the desktop version’s 69°C. The 48°C on the small screen can be read as a missed rewrite. A person reading on a smartphone sees 69°C on the card and 48°C in the chart’s annotation.

A red-orange line on black. The afternoon peak carries a note of 56.8°C, and for several evening hours the line breaks off with the words Missing data

Source: The Hindu Data team / The Hindu, “Heat index: how heat-stressed are you?”

Leon’s chart does not connect the missing hours with a line; it leaves a blank and writes “Missing data.” The line color is the same single red-orange in all four charts, and it does not change with the value. Bands showing the IMD levels are not drawn in the background either.

How do third parties see this method?

I could not find any critique that discusses this piece. Instead, I list sources that assess the method, by point at issue.

The first is the meaning of measuring personal exposure. Kuras et al. write in a 2017 paper:

“Personal exposure research can help reveal the extent of exposure misclassification that occurs when individual exposure to heat is estimated using ambient temperature measured at fixed sites”

In short, estimating an individual’s exposure from temperature measured at fixed sites causes misclassification, and personal exposure research can reveal how large it is. The piece’s framing, the citywide high of 36°C against what people actually experienced, can be read as the journalistic version of this point.

The second is the limits of the heat index formula. The NWS Weather Prediction Center writes on its page explaining the formula:

“The Rothfusz regression is not valid for extreme temperature and relative humidity conditions beyond the range of data considered by Steadman.”

That is, the regression formula is not valid for extreme conditions beyond the range of Steadman’s data. In a piece for Down To Earth (June 18, 2024), Ashok Gadgil and Elif Kilic of UC Berkeley caution that the heat index assumes a healthy young adult moving lightly in the shade. The piece’s 69°C comes from conditions of 47 to 48°C air temperature. A fisherman at sea and a cook at a pot can be read as standing outside the formula’s assumptions. The piece reveals its outliers in annotations, but it does not mention the formula’s range of applicability.

The third is research using the same approach. According to an article from Princeton University Engineering (December 20, 2024), a team including Anu Ramaswami gave residents of Delhi and Chennai pendants with heat sensors. They are devices built on the same idea as the piece, computing a heat index from temperature and humidity. The difference can be read this way: the research gathers many participants, while the piece narrows to four people and one day and shows it as a story.

The piece made the shortlist of the Sigma Awards 2026. There were 543 entries from 84 countries, and 31 finalists. The Hindu’s name does not appear in the announcement of the ten winners.

At the end of the piece’s page, The Hindu’s making-of video, “Behind the scenes: How the data and science teams measured heat stress | In Focus podcast,” is embedded (YouTube). I have not watched the contents of the video.

Takeaways

What this piece lets you take away is to fix one thing to compare against and put it first. Because the lead sentence has 36°C, the 69°C on the cards reads as “high.” Show the reference number first and put the number people actually experienced beside it. This order works for reporting that re-measures something.

Next is following a small number of people with numbers and words. Four people over one day is not statistics. Instead, three numbers are tied to each person’s actions, paragraph by paragraph.

The third is to state gaps and anomalies inside the chart. The blank of “Missing data” and the note about the stove tell readers how far the numbers can be trusted. On the other hand, annotations that differ between versions and yardsticks that do not match leave readers who want to compare confused. When you build charts separately for different screen widths, check even the numbers in the annotations against both versions. If you change the vertical axis for each person, provide a separate place for comparison. This piece is an example of both.

References

Primary sources

The Hindu Data team, “Heat index: how heat-stressed are you?” (headline “TAKING ON THE HEAT”), The Hindu, 2025-07-12 (updated 2025-07-30). Reporting and writing: Nitika Francis / Production and data: Srinivasan Ramani / Hardware Assembly: Sanjeevi Maran / Interactive and Code: Areena Arora / Illustrations: Soumyadip Sinha, and others. The piece
The piece’s scroll handling and color levels (gif.v2.js). JavaScript
The piece’s CSS (color variables, two-column pinned layout). gif.v2.css / project_styles.v8.css
Sigma Awards 2026 entry “Chennai’s heat index: How heat stressed are you?” (2026 · Shortlist). Sigma Awards entries
The Hindu, “Behind the scenes: How the data and science teams measured heat stress | In Focus podcast”. YouTube
The Sigma Awards, “Remarkable diversity, confronting some of the most urgent issues of our time: announcing the 2026 Sigma Awards shortlist”. Announcement
The Sigma Awards, “From toxic supply chains in Africa to Baltic sabotage to stolen children in Syria: 10 data projects win 2026 Sigma Awards”. Announcement

Secondary sources

E. R. Kuras et al., “Opportunities and Challenges for Personal Heat Exposure Research,” Environmental Health Perspectives, 125(8):085001, 2017. PubMed / doi:10.1289/EHP556
NOAA/NWS Weather Prediction Center, “The Heat Index Equation”. Explanation
Ashok Gadgil, Elif Kilic, “Get the Heat Index right: an important first step to protect people from heatwaves,” Down To Earth, 2024-06-18. Op-ed
Molly Sharlach, “Measuring the human impacts of extreme heat to guide cities’ climate action plans,” Princeton Engineering, 2024-12-20. Article

Showing a Ledger of About 80,000 Parcels at Different Scales: High Country News's "Land-Grab Universities"

Wed, 04 Nov 2026 00:00:00 +0900

High Country News split a ledger of about 80,000 parcels that became university endowments under the Morrill Act into three views at different scales: a national dot map, maps by cession and by university, and parcel cards. For each parcel, the project sets the amount paid to Indigenous nations beside an estimate of what the university raised. The university figure is not a sales record. It is an estimate that spreads each state’s average per-acre figure across the parcel’s acreage.

The work is “Land-grab universities,” published on March 30, 2020. The text was written by historian Robert Lee and HCN’s Tristan Ahtone. The maps come from two lines of work: the maps in the text and the print edition are by Margaret Pearce, while the interactive web map, landgrabu.org, was designed by Geoff McGhee and developed by Cody Leff. It won a Sigma Awards 2021 prize (winner). Below, I read it from the largest scale to the smallest, following the text, the methodology article, landgrabu.org, and the published data.

Turning Land Called a “Donation” into a Ledger

On July 2, 1862, Abraham Lincoln signed the Morrill Act, a law that gave states public land and used the proceeds from its sale to support colleges. The text rejects the “donation of land” story right at the start.

Behind that myth lies a massive wealth transfer masquerading as a donation.

The sentence says that behind the myth of a gift lies a large transfer of wealth disguised as a donation. The text then explains why this went unseen.

But with a footprint broken up into almost 80,000 parcels of land, scattered mostly across 24 Western states, its place in the violent history of North America’s colonization has remained comfortably inaccessible.

In short, because the grant was scattered across almost 80,000 parcels in 24 Western states, its place in the history of colonization stayed out of view. The scattering itself can be read as the cause of the invisibility.

Lee’s methodology article describes how the ledger was built in seven steps. The team extracted parcels from General Land Office patent records and collected more than 99% of each state’s granted acreage. The center point of each parcel was overlaid on Charles Royce’s cession maps, tying each parcel to the land cession through which the land was taken.

The data is published on GitHub under the Open Database License and CC BY-SA 3.0. When I tallied Parcels.csv myself, it had 79,461 rows and the payments totaled $397,247. That matches the text’s “less than $400,000.” Parcels with a $0 payment make up 29.8%, which also matches “more than a quarter.”

The National Dot Map: Placing 80,000 Parcels on One Sheet

The announcement for Pearce’s talk (Northrop, University of Minnesota, 2022) describes the design of the maps this way.

Cartographer Margaret Pearce will share how the maps and graphs are designed to amplify the project at multiple scales …

In short, the maps and graphs were designed to work at multiple scales. Because this is an event announcement, it is impossible to tell whether the words are Pearce’s own or the organizer’s summary.

The largest scale is the national map. The first Pearce map in the text opens with a headline saying that 79,461 parcels, about 10.7 million acres, were given to 52 universities.

Map of the contiguous United States. Small purple parcels are scattered across the West and Midwest, and red dots mark the locations of 52 universities

Source: Map by Margaret Pearce, from High Country News “Land-grab universities”

Parcels are small purple fills and universities are red dots. The beige background is Indigenous land that the United States acquired by treaty, cession, or seizure. The purple parcels cluster in the West and Midwest, while most of the red dots are in the East.

The web map, published on April 6, 2020, also shows the whole first as dots. In the Mapbox style embedded in the app, the background is a dark blue-gray. The parcel center points are cyan (#8ef7fb) with a small radius of 0.5 to 1 pixel, and they disappear at zoom 9. Instead, from zoom 8 the parcels’ shapes appear as white dotted outlines.

The dot map can be read as the stage that keeps all 80,000 parcels intact and shows only their distribution. Which parcel belongs to which university is left to the next scale.

Maps by Cession and by University: Connecting Distant Land with Lines

According to the text, Western states chose parcels from within their own borders. States in the East and South, and some in the Midwest, were given scrip (land vouchers), which they used to select land in the West.

In fact, the grants were as big or bigger than major cities, and were often located hundreds or even thousands of miles away from their beneficiaries.

In short, the grants were as large as big cities and were often hundreds or thousands of miles from the schools that benefited. In her maps by university, Pearce showed this distance.

Two maps of the United States. On the left, short lines fan out from the University of California to parcels within the state; on the right, long lines stretch from Cornell University to parcels in the West and Midwest

Source: Map by Margaret Pearce, from High Country News “Land-grab universities”

The legend reads “lines connecting a university to the individual parcels that funded it.” The University of California received 2,395 parcels from land inside the state, and its lines stay within the state. Cornell University obtained 6,716 parcels through scrip, and its land comes from 63 cessions and other sources across 15 states. The endowment principal around 1914 was $740,146 for UC and $5,739,657 for Cornell.

There is also a map that takes a single cession as its unit: the 1851 Dakota cession.

Map around Minnesota. Inside the boundary of the 1851 Dakota cession, dots color-coded by university are scattered, with a legend of 35 universities below

Source: Map by Margaret Pearce, from High Country News “Land-grab universities”

Parcels inside the boundary are color-coded by the 35 universities that benefited, and parcels from other cessions are gray. According to the text, about 830,000 acres from this cession funded 35 universities. The University of Minnesota is said to have received $250 for every $1 paid to the Dakota.

The web map extends these lines to all universities. It draws arcs with deck.gl’s ArcLayer from each parcel to the university in the beneficiary state. The arcs start in a cyan tone and end in red-orange, and the more arcs a university has, the thinner the lines. The site’s own note reads “This map connects Indigenous land (in blue) redistributed to universities (in red).”

Satellite-image map of the United States. Bundles of white lines stretch from the West and Midwest to red dots in the East, and cession boundaries are drawn as gold dashed lines

Source: High Country News (Web App Design: Geoff McGhee, Development: Cody Leff), from High Country News “Land-grab universities”

Cession boundaries are gold dashed lines. The lines are neither roads nor rivers; they represent the connection between a parcel and a university. At this scale, the distance between land and university becomes visible for the first time, as I read it.

Parcel Cards and the Caveat of Estimation

The smallest scale is a single parcel. The text has 14 cards (covering 12 parcels), each listing fields in a fixed order beneath a photograph by Kalen Goodluck. The photographs show what is on the site today: railways, cemeteries, scrapyards.

The first card is for the Niles Canyon Railway parcel in Sunol, California.

PARCEL ID: CA210040S0010W0SN020AE½SWAL
OWNERSHIP TRANSFER METHOD: Seized by unratified treaty, May 28, 1851
GRANTED TO: State of Alabama
FOR THE BENEFIT OF: Auburn University
AMOUNT PAID FOR INDIGENOUS TITLE: $0
AMOUNT RAISED FOR UNIVERSITY: $72.01

Land in California ends up funding Auburn University by way of the State of Alabama. The scrip mechanism shows in the two lines “GRANTED TO” and “FOR THE BENEFIT OF.” The two amounts are set side by side as plain text.

In the Arizona section, there is a figure that links three images: a map, an aerial photograph, and a photograph.

Three panels. A map of Tucson marks the parcel, an aerial photograph outlines the 640-acre parcel in red, and a photograph shows the motel standing on the site

Source: Map by Margaret Pearce, from High Country News “Land-grab universities”

Within one figure, the scale steps down from a city map to an aerial photograph with one parcel outlined in red to a photograph from the ground. The parcel is 640 acres seized from the Apache in 1886, with a payment of $0 and a university endowment principal of $2,004.39.

Here I check the nature of the university-side numbers. This is Step 7 of the methodology.

Using the principal raised and value of the unsold land for each state’s grant, we calculated the average contribution to endowment per acre, then multiplied that average by the number of acres in a parcel to estimate each parcel’s average contribution to an endowment.

In short, for each state the team took the principal raised and the value of unsold land, worked out a per-acre average, and multiplied it by the parcel’s acreage. The “amount raised for university” is therefore an estimate, not a sales record. The payment to Indigenous nations is likewise a single cession’s total, allocated by area.

Using the CSV, I compared parcels in the same section in Sunol. The Auburn parcel is 80 acres at $72.01, and the neighboring UC parcel is 39.84 acres at $195.90. That works out to about $0.90 and about $4.92 per acre. The differences in how each state sold its land can be read as showing up directly in the estimates.

The web map changes its fill by the accuracy column. According to the methodology, 99% of the acreage is accurate to the acre and 1% to the square mile. In the style, different fill patterns are applied to Aliquot and Section in the Accuracy column. These two can be read as corresponding to the methodology’s one-acre and one-square-mile precision.

How the Views Map to the Data

Scale and medium (creator)	Data	Visual and material correspondence
National map (text and print, Pearce)	Parcels, cessions, universities	Parcels as purple fills, universities as red dots with names, cessions as beige areas
National map (web, McGhee and Leff)	Parcel center points	Cyan dots. At zoom 8 and above they give way to white dotted outlines
Map by university (text, Pearce)	Parcels matched to beneficiary universities	Purple lines from a university’s red dot to its parcels, with parcel count, acreage, and principal added
Map of a cession (text, Pearce)	Parcels in the cession and beneficiary universities	Parcel dots color-coded by university inside the cession boundary
Arcs between parcels and universities (web, McGhee and Leff)	Parcel to university in beneficiary state	Arcs. Cyan-tone start, red-orange end; more arcs make thinner lines
Cession boundaries (web, McGhee and Leff)	Royce’s cession areas	Gold dashed lines, with italic serif labels
Location accuracy (web, McGhee and Leff)	Accuracy column (Aliquot/Section)	A different fill pattern for each accuracy level
Parcel cards (text, photographs by Goodluck)	One row of the CSV	Beneath a photograph of the site: ID, transfer method, granted state, beneficiary university, and two amounts, all as text

How Third Parties Saw the Work

Map of the Week, a cartographers’ blog, covered the work on April 8, 2020.

Here is a detailed view since the map above is hard to read at this size.

In short, the national map is hard to read at this size, so the blog adds a detailed view. It also names the land lying far from the universities as a highlight. This is a review from the reader’s side that shows the national figure alone is not enough and a zoomed-in scale is needed.

Tiara Roxanne wrote in February 2021 from the standpoint of Indigenous data sovereignty (IDS).

By organizing this data utilizing differing digital markers (graph, table, map), the project contributors create a space for IDS.

In short, organizing the data in different forms (graph, table, map) created a space for data sovereignty. She also adds a note of caution that colonialism is embedded in digital infrastructure.

Historian Jon Parmenter introduced a forum in the Spring 2021 issue of Native American and Indigenous Studies. One of its themes, he notes, is the power of a mapping method that links specific dispossessions to the benefit of specific universities.

The work also won a George Polk Award (education reporting) and an IRE Award.

Takeaways

What this work offers to visualization practice is a way of splitting one ledger into views at different scales.

Show the overall distribution as dots, and switch to shapes as the viewer moves closer. The aim is to keep 80,000 parcels without lumping them together. Draw the relationship between two distant parties as a line. Unlike a table, it conveys the distance as well. Turn a single record into a card and lay out the numbers as text. In a bar chart, $0 has no length, but as text “$0” reads as it is.

The other point is to treat an estimate as an estimate. The methodology states that the university-side figures are a state-average allocation, and the data is published. Making it possible for readers to check how reliable the figures are is what supports a strong claim.

References

Primary sources

Robert Lee & Tristan Ahtone, “Land-grab universities,” High Country News, March 30, 2020. https://www.hcn.org/issues/52.4/indigenous-affairs-education-land-grab-universities
Robert Lee, “How we investigated the land-grant university system,” High Country News, March 30, 2020. https://www.hcn.org/articles/indigenous-affairs-education-how-we-investigated-the-land-grant-university-system/
Land-Grab Universities (interactive site; credits and data license). https://www.landgrabu.org/
JS of the landgrabu.org app (Mapbox style and deck.gl arc-layer). https://www.landgrabu.org/app-0797f4bb73a5a4989184.js
HCNData/landgrabu-data, README. https://raw.githubusercontent.com/HCNData/landgrabu-data/master/README.md
HCNData/landgrabu-data, Parcels.csv. https://raw.githubusercontent.com/HCNData/landgrabu-data/master/Morrill_Act_of_1862_Indigenous_Land_Parcels_Database/CSVs/Parcels.csv
Kalen Goodluck, Tristan Ahtone & Robert Lee, “The land-grant universities still profiting off Indigenous homelands,” High Country News, August 18, 2020. https://www.hcn.org/articles/indigenous-affairs-the-land-grant-universities-still-profiting-off-indigenous-homelands
Sigma Awards 2021 entry page (original application and Japanese translation). https://sigma-awards.data-visualization.jp/entries/2021-s-0032/
High Country News, “‘Land-Grab Universities’ wins big.” https://www.hcn.org/?p=121248
University of Cambridge, “Historian wins major journalism award for Indigenous land project.” https://www.cam.ac.uk/research/news/historian-wins-major-journalism-award-for-indigenous-land-project
University of Cambridge, “The great university land grab.” https://www.cam.ac.uk/stories/great-university-land-grab
Northrop, University of Minnesota, “Keeping Accountability at the Table: Mapping Land-Grab and What’s Next,” October 20, 2022. https://www.northrop.umn.edu/events/keeping-accountability-table-mapping-land-grab-and-next-2022

Secondary sources

Map of the Week, “Land Grab Universities,” April 8, 2020. https://mapoftheweek.blogspot.com/2020/04/land-grab-universities.html
Tiara Roxanne, “Indigenous Data Sovereignty and the Land Grab Universities Project,” Heinrich-Böll-Stiftung, Gunda-Werner-Institut, February 11, 2021. https://www.gwi-boell.de/en/2021/02/11/indigenous-data-sovereignty-and-the-land-grab-universities-project
Jon Parmenter, “Assessing Cornell’s Response to Revelations it was Funded by Taking Indigenous Land,” History News Network, August 11, 2022. https://historynewsnetwork.org/article/183730

Showing How 227,000 Was Counted, Split into Maps of Place, Time and Numbers: Reuters' Hong Kong Protest Count

Wed, 04 Nov 2026 00:00:00 +0900

For the July 1, 2019 protest in Hong Kong, Reuters estimated that 227,000 people marched. It shows where it counted with a map and photos, when it counted with a timetable of dots, and the result of the count with a bar chart of 15-minute estimates, each as a separate graphic.

The piece is Reuters Graphics’ “Hong Kong protests: How many protesters took to the streets on July 1?”, published on July 4, 2019. It is signed by three people: Simon Scarr, Manas Sharma and Marco Hernandez. It was also an entry for the Sigma Awards 2020. Below, I read it graphic by graphic, following the piece’s text and figures, the Backstory in which Reuters described how it was made, and the Sigma entry form.

July 1, 2019: four numbers side by side

July 1 is the anniversary of Hong Kong’s handover, and the day on which pro-democracy groups march every year. The march started at Victoria Park. The planned end point was the Government Headquarters, but clashes at the morning ceremony moved it to Chater Road, about 400 m away.

At the top, the piece lines up four estimates in prose.

Who estimated	Number	How the piece describes it
Organizers	550,000	People who took to the streets
Police	190,000	People in the march “at its peak”
Hong Kong Public Opinion Research Institute (HKPORI)	260,000	People who took part in the protest
Reuters	227,000	People who marched past in the Admiralty district

A note at the end says the HKPORI figure was wrongly given as 265,000 when the piece was first published and was later corrected. The piece has no graphic that puts the four numbers side by side.

The piece says that in Hong Kong, protest turnout is treated as a gauge of public opinion and is also a political fight between the organizers and the government. The police told Reuters that their figure was “only a rough figure solely for the purpose of effective manpower deployment.” The organizers did not respond.

In its Sigma entry form, Reuters gives its motive as follows.

For Reuters, this was an opportunity to calculate our own estimate and provide an unbiased statistic.

In a June 20 wire story, Scarr and Tong had relayed scholars’ view that for a march it is better to count the people passing a section of the route in a short time and multiply by the length of the march. The July piece uses this method.

Showing where it counted, with a map and photos

Scarr chose the counting location on a scouting trip. According to the Backstory, Scarr studied the roughly 3 km route on Google Earth and looked for a place the crowd had to pass. From photos posted on TripAdvisor, he also confirmed that a room in a certain hotel overlooked that spot. The hotel was the main observation point, and there were two more on a pedestrian bridge.

The first graphic the piece shows is a map of the march route.

A map of Hong Kong Island with the march route drawn as a gold line, the counted location as a black rectangle on the route, and the direction of the crowd’s flow as black arrows

Source: Scarr, Sharma, Hernandez (Reuters) / Reuters Graphics, “How many protesters took to the streets on July 1?”

The graphic was exported with ai2html, and the place names are laid over the image as HTML text. That is why the image quoted here has no lettering. In the piece, “START Victoria Park” appears at the right edge, “END Chater Road” at the left edge, “Counting location” below the black rectangle, and “Flow of people” next to the arrows.

According to the piece, part of the march headed for the Government Headquarters along the way. The counting location is before that fork, and there is no shortcut nearby. So the people heading for the Government Headquarters were counted too, the piece explains.

A black-and-white aerial photograph annotated with the government buildings filled in gold, white arrows branching off, and the counted area as a yellow rectangle

Source: Scarr, Sharma, Hernandez (Reuters) / Reuters Graphics, “How many protesters took to the streets on July 1?”

The second graphic is an enlargement of the area around the fork, with the branching arrows and a “Counting area” rectangle drawn over the photo. The video that follows looks down on Hennessy Road from the observation point, with a note that the full width of the road was in frame.

A black-and-white photograph looking down from a height at a road filled with a crowd, with a single yellow line drawn across the road

Source: Scarr, Sharma, Hernandez (Reuters) / Reuters Graphics, “How many protesters took to the streets on July 1?”

The image above is the piece’s share image. It can be read as showing the line that the text refers to when it says people “passing a line in the road” were counted.

After showing the place, the piece states the limits of the number.

That means Reuters’ calculations are not an estimate of the total number taking part in the protest - only the march.

People who left before the counting location, and people who went straight to Tamar Park without marching, are not counted. So 227,000 is the number of marchers, not the total number of protest participants, the piece says.

Showing when it counted, with dots

Next, the piece makes a graphic of when it counted.

The team recorded 30 seconds of video at least every 15 minutes and counted the number of people passing a line in the road during that time.

That is, the team filmed 30 seconds of video at least every 15 minutes and counted the people who crossed a line in the road in that time.

A graphic with time on the horizontal axis and one yellow dot for each time a 30-second measurement was taken. The dots are packed close in some stretches and sparse in others

Source: Scarr, Sharma, Hernandez (Reuters) / Reuters Graphics, “How many protesters took to the streets on July 1?”

The horizontal axis runs from 3 p.m. to 9 p.m. The legend reads “Each dot shows when Reuters analysed crowd flow,” so one dot is one measurement. Arrows are labeled “Front of march arrives” and “Tail passes and main march is over,” and the wide-screen version adds “Closer bursts of counting as crowd density fluctuated.” Counting the dots in the image myself, I got about 30.

The method for turning a 30-second count into a total is as follows.

Reuters extrapolated the number of people measured in 30 seconds over the period until the next reading - for instance, if there was a five-minute gap, the 30-second measurement would be multiplied by 10.

The count for 30 seconds is stretched over the time until the next measurement; if the gap is five minutes, it is multiplied by 10. As a formula, the total is the sum, over all measurements, of the count times (seconds until the next measurement ÷ 30).

In other words, the spacing of the dots is itself the extrapolation multiplier. It can be read as meaning that the sparser the dots, the more a single 30-second sample is stretched. Professor Keith Still, a crowd scientist, says in the piece that if you count the people passing through an area of known width over a short time, you should be able to get an accurate figure for the number who passed that spot.

The counting was done by hand. According to the Backstory, filming the march took more than four hours, and the team watched and counted 18 hours of video.

Adding up the 15-minute bars gives 227,000

Below the dot graphic, the piece places a bar chart of the estimated number of people for each 15 minutes.

A bar chart with time on the horizontal axis and number of people on the vertical axis, showing the estimated size of the march every 15 minutes as yellow vertical bars. Tall bars run through the middle

Source: Scarr, Sharma, Hernandez (Reuters) / Reuters Graphics, “How many protesters took to the streets on July 1?”

The vertical axis is labeled “Number of people,” with ticks at 5,000, 10,000 and 15,000. The subtitle is “15 minute estimates,” and the horizontal axis is the same as in the dot graphic. The piece concludes that adding these estimates gives the total number who passed the counting location, and that this is 227,000. When I read the heights of the 20 bars off the ticks and added them, I got about 227,000.

However, there is no graphic showing the result of the addition. Right after this paragraph there is an empty element called cumulative, but the script has no code to draw it, so nothing appears in the current version. I could not confirm whether a graphic existed when the piece was first published.

The correspondence between graphics and data is summarized below.

Graphic	Data shown	Correspondence to appearance
Route map	Route, start and end points, Government Headquarters, counted location	Position on the map. The gold line is the route, the black rectangle is the location, the arrows are the direction of flow
Counting-area graphic	The fork and the counted area	Arrows and an area rectangle laid over an aerial photo
Observation-point video	Traffic on Hennessy Road	The road itself as seen from the observation point
Measurement-time graphic	Times of the 30-second measurements	Horizontal position is time, one dot is one measurement, spacing is the density of measurements
15-minute bar chart	Extrapolated numbers of people, totaled for each 15 minutes	Bar length is the number of people, horizontal position is the time slot

The dots and bars line up vertically on the same time axis. Comparing them, the reader can check which bars are values stretched from few measurements, it can be read. On the other hand, none of the graphics shows the extrapolation multipliers, the raw counts or the margin of error. The source is given as “Source: Reuters reporting,” and as far as the piece and its links show, the data has not been published.

How third parties saw this piece

On the Sigma Awards listing site, this piece is classified as a 2020 “entry” and is not on the shortlist. The results from other awards are as follows.

Online Journalism Awards 2020: the series “Visualising the Hong Kong protests,” which includes this piece, won Excellence in Visual Digital Storytelling, Large Newsroom.
Same awards: an entry that paired it with the June piece “Measuring the masses” was a finalist for Explanatory Reporting, Large Newsroom.
Kantar Information is Beautiful Awards 2019: longlisted in the Politics & Global category.

The description text on the award pages is Reuters’ own entry text, so only the results can be used as an evaluation.

Researchers who counted the same day by a different method discuss it more concretely. In a 2023 paper, Chow, Yip and Wong combined capture-recapture with automated counting by a CNN and estimated 276,970 people (95% confidence interval 263,663 to 290,276). They mention Reuters as follows.

However, their approach used a single check point late in the rally course and excluded those who left before it. Considering the estimated early departure of 72,981 based on the field sample, the Reuters count is also similar to the overall count reported in this study.

In other words, Reuters counted at a single point late in the route and did not include people who left before it. Taking into account the estimated early departure of 72,981, the Reuters count is also close to the authors’ own estimate. The paper’s references list the URL of this piece.

However, the paper and the piece disagree on the number and the location. The paper gives Reuters’ number as “about 220,000” and the location as “near Arsenal Street.” The piece gives 227,000 and Hennessy Road (Admiralty district). I could not confirm the reason for the discrepancy.

The paper further points out, without naming anyone, that methods which multiply time samples assume a constant flow, which is difficult for a crowd whose walking speed changes minute by minute. The piece stated in words that its number is “not the total.” The paper can be read as putting a number, about 73,000 people, on that difference.

What to take from it

What we can take away is a structure for showing how a single number was made by splitting the process into separate graphics.

The graphic of place fixes the scope of what was counted. That the counting location was chosen before the fork and with no shortcut can be checked with the map and photos. For the people outside that scope, the piece says so in words. The piece makes this division of roles clear.

The graphic of time shows how the sample was taken. Just by lining up dots, we can see when and how densely the counting was done. If the result bars are aligned with the time axis, we can compare which values were stretched from a thin sample, it can be read.

On the other hand, the extrapolation multipliers, the raw counts and the margin of error are not in a graphic. As in the 2023 paper, there is another path: attaching the range of the estimate as a number. A graphic that discloses the method and a graphic that shows the range are different things. This piece drew the former, split into place, time and number of people.

References

Primary sources

Simon Scarr, Manas Sharma, Marco Hernandez, “Hong Kong protests: How many protesters took to the streets on July 1?,” Reuters Graphics, 2019-07-04. The piece (original URL: graphics.reuters.com)
Text and graphic labels of the same piece (script). main.js
Sigma Awards 2020, “How many protesters took to the streets on July 1?” (original entry form and Japanese translation). Sigma Awards entries site
Reuters, “Backstory: How Reuters counted a quarter million people at Hong Kong’s protests,” 2019-07-18 (Yahoo reprint). Article
Reuters, “Mass movement: counting marchers in Hong Kong,” 2019-07-04 (The Star reprint, first version). Article
Reuters, “CORRECTED-Mass movement: counting marchers in Hong Kong” (Devdiscourse reprint, corrected version). Article
Reuters, “In Hong Kong, science of crowd counting comes into focus,” 2019-06-20 (signed by Scarr and Tong, Yahoo reprint). Article
Online Journalism Awards entry pages (the description text is Reuters’ entry text). Visualising the Hong Kong protests / Hong Kong protests

Secondary sources

T. E. Chow, P. S. F. Yip, K.-P. Wong, “An integrated framework of mobile crowd estimation for the 2019, July 1st rally in Hong Kong,” Multimedia Tools and Applications, 2023. doi:10.1007/s11042-023-15417-7 / PMC full text
Online Journalism Awards 2020, Excellence in Visual Digital Storytelling, Large Newsroom (winner). Result
Online Journalism Awards 2020, Explanatory Reporting, Large Newsroom (finalist). Result
Kantar Information is Beautiful Awards 2019, Politics & Global (longlist). Showcase

Teal before, yellow after: the 2,312 mosques in the FT's "How China is tearing down Islam"

Wed, 04 Nov 2026 00:00:00 +0900

The Financial Times (FT) examined 2,312 mosques with Islamic architecture in Google Earth satellite imagery and counted 1,714 that had been modified. It fixed teal (#35e0cb) for before and yellow (#ffec1a) for after, and uses these two colors for the dots on the national map and for the before-and-after switches between satellite photos.

The piece is “How China is tearing down Islam.” It was published on 27 November 2023 as an FT Visual Investigations project, and it won a Sigma Awards 2024 prize (according to its entry page on the Sigma Awards entries site). Below, I read it in the order a reader scrolls, based on the body text and source code (HTML, CSS, JavaScript) of the project page, the Flourish data for the regional chart, and the Sigma Awards application.

It starts with one mosque in Beijing

When you open the piece, a photograph of a single mosque fills the screen. It is the Doudian Mosque on the outskirts of Beijing. The text describes it as one of the most magnificent mosques in northern China.

A photograph of the white Doudian Mosque, with a dome and two minarets, taken from directly above the front

Source: Financial Times (© The Financial Times Limited 2023) / How China is tearing down Islam

In the next step, a second image, whose alt text reads “Mosque, outlined,” fades in over three seconds. Then, as you scroll, drone footage plays. It was shot by Marine Zambrano. The text goes on:

But earlier this year, its minarets were removed, its domes replaced with pagoda-style cones and its Arab-style arches squared off.

Summary: earlier that year the minarets were removed, the domes were replaced with pagoda-style cones, and the arches were squared off.

From here, the screen moves from a single building to a map of all China. At first only the area of Xinjiang is shaded dark blue-green, and yellow dots appear in it. In the next step, dots spread across the country. The map data contains 1,714 dots. In descending order of dot count: Xinjiang 387, Gansu 355, Ningxia 268, Henan 205, Yunnan 146, Qinghai 118.

After the map come satellite photos. First four large tiles appear, switching one by one at one-second intervals from before to after. Then a full-screen grid of 70 before-and-after satellite pairs scrolls automatically. The images date from January 2017 to March 2023.

The sequence runs from one photo of one building, to a national map, to 70 pairs of satellite photos. It can be read as a structure that first shows what happened to one building, then shows how many places the same thing happened.

From 4,450 locations to 1,714 mosques

The numbers are given in the methodology paragraph at the end of the article.

We built a dataset of 4,450 reported locations of mosques across China by combining search results from Baidu Maps, Google Maps, and OpenStreetMap, as well as mosques in Xinjiang from the Australian Strategic Policy Institute. We inspected each location using Google Earth satellite imagery, identifying mosques and comparing changes with their architecture over time.

Summary: the team combined search results from Baidu Maps, Google Maps and OpenStreetMap with the Xinjiang mosques from the Australian Strategic Policy Institute (ASPI) into a list of 4,450 locations. It then checked each location in Google Earth satellite imagery, confirmed whether it was a mosque, and compared how the building changed over time.

The result is in the next paragraph.

In total, we located 2,312 mosques with Islamic architecture. Of these, 1,714 (74.3 per cent) had Arabic-style features removed between 2018 and 2023.

Summary: the team located 2,312 mosques with Islamic architecture, and in 1,714 of them (74.3%) Arabic-style features had been removed between 2018 and 2023.

Lined up by stage, the numbers look like this.

Stage	Count	Wording in the text
Candidate locations	4,450	reported locations of mosques
Confirmed as Islamic architecture	2,312	mosques with Islamic architecture
Modified or removed, 2018–2023	1,714 (74.3%)	had Arabic-style features removed

What was judged is minarets, domes, and other exterior Arabic-style features. The methodology also states its limits. The mosques examined are only a portion of those in China as a whole. In some regions the most recent satellite imagery dates from 2020, so buildings modified after that were not counted. It therefore presents 1,714 as probably an underestimate.

The article is not built from satellite imagery alone. The text also brings in photos taken on the ground and interviews with residents, a local government’s renovation drawings in Sichuan, and an image of a planning document from Xining. In its Sigma application, the team lists Drone, Scraping, QGIS, Google Sheets, OpenStreetMap, Node.js and others among the technologies used.

A renovation elevation of the prayer hall of the Maoxian Chengguan Mosque (茂县城关清真寺). Dashed lines mark the position of the dome, and a multi-tiered Chinese-style roof is drawn above it

Source: Renovation drawing for Maoxian (a local government document), Financial Times (© The Financial Times Limited 2023) / How China is tearing down Islam

Teal for before, yellow for after

The page’s stylesheet repeats two colors. Teal (#35e0cb) is before the modification, and yellow (#ffec1a) is after. This can be read as the one rule that ties the data to its appearance.

Near the top of the text, the two words “before” and “after” appear in colored bold. “Before” is teal and “after” is yellow. On the satellite tiles, the border and the date label at upper right are teal for before and yellow for after. The dots on the map of China are also small yellow squares.

Where the text discusses Doudian in detail, two satellite images sit side by side. The project page labels them 2017 and 2023 (the images themselves carry no year). The 2017 image is tinted teal and the 2023 image yellow. Around the building’s perimeter runs a line made of a solid black line overlaid with a white dashed line.

Two satellite images of the Doudian Mosque. On the page, the left one, labeled 2017, is tinted teal and the right one, labeled 2023, is tinted yellow; a solid black line overlaid with a white dashed line traces the building’s perimeter

Source: Financial Times (© The Financial Times Limited 2023) / How China is tearing down Islam (the two images placed side by side by the author)

The table below summarizes how the colors are used on the page.

Presentation (location)	Data	Mapping to appearance
Opening photo and drone footage	One mosque, Doudian, before and after	After the before photo, an image with the alt text “Mosque, outlined” fades in over 3 seconds. Drone footage advances as you scroll
Map of China	Locations of the 1,714 modified mosques	Small yellow squares (thin black outline). Xinjiang first, then the whole country. Each dot appears with a delay of 0 to 750 ms
Colored words in the text	Before / after	“Before” in teal bold; “after” and “demolitions and modifications” in yellow bold
Four large satellite tiles	Four before-and-after satellite pairs	Switch from before to after every second. The before image blurs and disappears. Border and date label: teal for before, yellow for after
Satellite photo grid	70 before-and-after satellite pairs	The grid scrolls automatically, and each tile that enters the screen switches from before to after
Before-and-after per mosque	Doudian, Najiaying and Weizhou, before and after	Satellite images and ground photos placed side by side before and after, each with its year
Maoxian renovation drawing	Government drawings and documents	In the text, “dome” is highlighted in teal and “pagoda-style structure” with a yellow background
Regional scatter plot (Flourish)	31 provinces and regions	x = Muslim population share (log scale), y = modification rate, circle area = number of mosques with Islamic architecture, fill yellow

The map dots are yellow, which can be read as meaning the dots are buildings “after.” The 1,714 modified mosques line up on the map as yellow dots. On the satellite tiles, the same yellow goes on the frame of the after image. Whether you read the text, look at the map, or compare the photos, yellow appears to be built to mean after.

The two-color switch also helps with comparing photos. In the Doudian satellite images, the roof shape seen from above the dome differs between 2017 and 2023. Because the colors differ, you can tell which image is the before without reading the date.

The regional scatter plot, and totals that do not match

Later in the text, a scatter plot made with Flourish is embedded. Its title is “Mosque modifications were more systematic in areas with larger Muslim populations.” Each circle is one province or region. The horizontal axis is the Muslim population share (log scale), and the vertical axis is the share of mosques with Islamic architecture that were modified. Circle size represents the number of mosques with Islamic architecture examined.

In the embedded data, the figures for some regions are as follows.

Province / region	Muslim population share	Mosques with Islamic architecture	Modification rate
Xinjiang	56.51%	539	71.8%
Ningxia	35.08%	294	91.16%
Qinghai	18.43%	146	78.77%
Gansu	8.15%	434	81.8%
Henan	0.96%	212	96.7%
Liaoning	0.53%	43	9.3%

The text also says more than 90% in Ningxia and more than 80% in Gansu. The sources are given as “2020 Chinese Census” and “FT Research.”

Adding up the 31 regions in this Flourish data gives 4,440 locations examined and 2,310 with Islamic architecture. These differ from the text’s 4,450 and 2,312 by 10 and 2 respectively. The piece does not say why they differ.

This is a point to watch when following the numbers. The national figures (4,450 / 2,312 / 1,714) are the methodology figures in the text, while the regional figures are from the scatter plot data. When citing both in the same article, it is safer to make clear which set a number comes from.

How third parties saw the piece

The Sigma Awards judges cite both how the data was built and how the page was put together.

The team behind “How China is tearing down Islam” built a database of 4,450 buildings and manually reviewed each one of them for accuracy. … This entry successfully combined photos, videos, data visualization and coherence page design to tell a powerful story.

Summary: the team built a database of 4,450 buildings and manually checked each one. It combined photos, video, data visualization and a coherent page design to tell the story. The judges also note that many of the reporters in China worked at personal risk.

The Guardian cited the FT’s analysis in its article of 25 May 2024. It wrote that last year’s FT analysis showed that three quarters of more than 2,300 mosques had been modified or removed since 2018. It also reports on what later happened to the mosques in Najiaying and Shadian.

The anthropologist Ruslan Yusupov made a different point in an essay for the Political Theology Network in March 2026. On the Chinese internet, he notes, even those who celebrate Sinicization use “before-and-after” paired images.

when the narratives about Sinicization emerged online, they did so through a powerful medium of visual juxtaposition called “before-and-after” images.

Summary: when narratives about Sinicization emerged online, they did so through a powerful medium of visual juxtaposition, “before-and-after” images. Yusupov argues that paired images keep the “before” as something lacking. Yusupov is also one of the people interviewed in the FT’s piece.

The three assessments make different points. The judges look at how the piece was made, The Guardian at the numbers, and Yusupov at the before-and-after form itself. Yusupov’s point can be read as showing that the before-and-after form the FT used is just as available to the opposing side.

Takeaways

What a visualization practitioner can take from this piece is to narrow the color convention down to one and keep to it to the end.

The FT’s color mapping can be read as narrowed to a single rule: before = teal, after = yellow. It carries this through the words in the text, the tinting of the satellite photos, the borders and date labels of the tiles, the dots on the map, and the highlights on the drawing. It can be read as built so that it works without a legend.

This method is useful when you set different kinds of material side by side in one article. Photos, maps and drawings are each made with different tools. If you decide on one color convention first, then even when each kind of material looks different, they read as parts of the same article.

The way of counting is also worth taking. The FT put the narrowing in numbers in the text: from 4,450 candidates to 2,312, then to 1,714. It also wrote the limits in the same paragraph. In an article that presents counts, stating where the counting began and what was not counted, next to the numbers, lets readers check them.

Finally, there is Yusupov’s point. The before-and-after form is powerful, but either side can use it. Placing before and after images side by side does not by itself convey what was lost. In the FT’s piece, the residents’ words and the government documents can be read as carrying that part.

References

Primary sources

Financial Times, “How China is tearing down Islam,” 2023-11-27. Project page
Stylesheets of the piece: styles.a686d242.css, ChinaMap.58928fe1.css
Scripts of the piece: index.page.client.a5b40edf.js, chunk-de8e8e48.js (map of China), chunk-74e2fd96.js (satellite tile grid)
Flourish, “China Mosques - Scatter” (visualisation 15487273). Embed
Sigma Awards 2024, “How China is tearing down Islam” (original application text and judges’ comments). Sigma Awards entries
Sam Joiner, post on X, 2023-11-27. Thread Reader
Human Rights Watch, “China: Mosques Shuttered, Razed, Altered in Muslim Areas,” 2023-11-22 (an earlier report cited in the text of the piece). HRW

Secondary sources

Sigma Awards 2024 judges’ comments. Sigma Awards entries
Amy Hawkins, Elena Morresi, “Last major Arabic-style mosque in China loses its domes,” The Guardian, 2024-05-25. The Guardian (checked via a reprint on Yahoo News UK)
Ruslan Yusupov, “The Montage of Privation: Islam and the Architecture of Sinicization in China,” Political Theology Network, 2026-03-10. Political Theology Network

Total Area, Per Capita, Share of Land: Country Rankings Reshuffle in Green to Grey

Wed, 04 Nov 2026 00:00:00 +0900

Green to Grey shows the natural land and cropland that turned into buildings and roads across Europe between 2018 and 2023, country by country, in three ways: total area, per capita, and share of national territory. Norway ranks 11th by total area, 1st per capita, and 26th by share of land. Depending on what you divide by, the ranking of countries reshuffles. The three metrics are three identical grey bar charts that you switch between with buttons; they are not placed side by side in a single figure.

The work is a cross-border investigation published on October 1, 2025. It was led by the Dutch nonprofit Arena for Journalism in Europe and the Norwegian public broadcaster NRK, with scientists from the Norwegian Institute for Nature Research (NINA) and news organizations from several countries. It is a winner of the Sigma Awards 2026. Below, I read it in the order a reader sees it, drawing on the main story, the methodology page, the data behind the country charts, an interview with the creators, and a later academic paper.

Repeating “green to grey” in color

The main story opens with three sets of satellite photos: a coastal forest in Greece that became a five-star hotel, farmland in the UK that became an Amazon logistics hub, and a pine forest in Germany that became a Tesla factory. A 2018 Pléiades image and a 2023 Pléiades Neo image are stacked, and as you scroll the 2023 image fades in. It is a scroll-linked crossfade, not a before-and-after slider.

The text is split into two lines. The earlier state (“In Greece, a forest on the coast…”) sits on a green background, and the later state (“became a five-star hotel.”) sits on a grey one. The site’s colors, in charts, maps and buttons alike, come from just two families: green #39b54a and grey #5d676b. The title’s “green to grey” becomes the color scheme of the whole page.

The investigation started from a domestic Norwegian investigation that NRK published in January 2024. In 2024, Arena invited NRK, NINA and news organizations from other countries to a meeting in Bologna. The team began work in January 2025 and published in October 2025. The main story says it involved “41 journalists and scientists from 11 countries.”

The encoding of each component is as follows.

Component	Data	Mapping to the visual
Three opening satellite photo pairs	The same places in 2018 and 2023	Scroll amount = opacity of the 2023 image (0 to 1)
Dot map	Random sample of constructions larger than 10,000 m²	Position = place, color = grey (construction), land = pale green. Dot size is unrelated to the data
Counter of area lost	1,500 km² per year converted to seconds	Seconds elapsed since loading × 47.5 m²
Three country bar charts	Estimates for 30 countries	Length = value, all the same grey. Buttons switch to another chart
Wetlands of İzmir	Satellite images from 2018 and 2023	Drag-to-compare before-and-after slider

The dot map shows spread, not location

The first figure in the main story is a map of Europe on which grey dots are placed. Its caption reads “Random sample of constructions larger than 10,000m² identified by our methodology.” The text says each dot is one place where nature or cropland was lost.

About 16,000 dots are drawn at first, and once the map enters the screen, 100 more are added every 100 milliseconds, for about 29,000 in total. Each dot’s radius is picked at random from three values: 0.5, 1.0 and 1.5. The size of a dot does not represent the area of the construction.

In an interview with the Royal Geographical Society (RGS), the creators Zeynep Şentek and Léopold Salzenstein explain the design of the map as follows.

We specifically did not include a zoom feature on the map. This was to emphasise the overall of scale of nature loss within Europe rather than focusing on detailed data for local areas.

In short, they deliberately left out a zoom feature because they wanted to show the scale of loss across Europe rather than detailed local data. They also say that the dots are added as the page loads in order to show that the loss is still going on. At the end of the map sits a counter of the area lost since you started reading. The note beside it reads “(47.5 m² per second, 4.1 km² each day).”

The academic paper that came later draws the same land take in a different form. It is a map that aggregates the data into a 50 km grid and colors natural land and cropland areas with a bivariate palette.

A map dividing Europe into 50 km grid cells and coloring the land take of natural land and cropland with a bivariate palette, together with aerial photos of two places where cropland and forest turned into buildings

The paper’s map lets you read where and how much was lost through color intensity. The work’s dot map gives up on letting you read quantity and hands that role to the motion of dots accumulating. Even with the same data, the journalistic figure seems to have chosen to convey spread and continuity before quantity.

Three denominators, three bar charts

The country comparison is in the “No country spared” section of the story. Above a Datawrapper horizontal bar chart sit three buttons: “Total area (km²),” “By population (m²/person/year)” and “By country size (%).” Pressing one switches to a different Datawrapper chart. Only the selected button is filled in green.

Three identical grey horizontal bar charts. From the left, Turkey is at the top by total area, Norway per capita, and the Netherlands by share of land

Source: Arena for Journalism in Europe, NRK et al., Green to Grey (three Datawrapper charts placed side by side by the author)

The three charts share the same bar color and ordering; only the denominator differs. Here are the rankings of four countries, taken from the values in the published charts.

Country	Total area (km², 2018-2023)	Per capita (m²/person/year)	Share of land (%)
Norway	207 (11th)	6.26 (1st)	0.06 (26th)
Turkey	1,860 (1st)	3.55 (10th)	0.24 (6th)
Netherlands	192 (12th)	1.77 (23rd)	0.46 (1st)
UK	604 (5th)	1.47 (26th)	0.25 (4th)

Countries with equal values are counted in the order they appear in the chart. Norway’s 0.06% is tied with Latvia and Ukraine, and Turkey’s 0.24% with Ireland.

The text also puts all three views into words. By total area, Turkey leads with 1,860 km², followed by Poland. Adjusted for the size of the country, the Netherlands and Belgium come first. Adjusted for population, the Nordic countries fare worse.

But if we adjust the figures based on population size, Scandinavian countries fare worse. Norway’s construction translates to roughly six square metres of nature and cropland every year for each resident. In Switzerland, per capita losses are ten times less.

In short, divided by population the Nordic countries look worse: in Norway it is roughly 6 m² per resident each year, and in Switzerland a tenth of that.

The units of the three charts are not aligned. Total area is a six-year sum, while per capita is “per year.” Share of land corresponds to total area divided by the country’s area. For the Netherlands, 192 km² divided by about 41,500 km² gives 0.46%, a cumulative figure over six years (my own check). The definition of share of land is not written in the chart.

From their headlines, The Black Sea (“Turkey Leads In European Nature Loss”) seems to stand on the total-area metric, and NRK (“Nordmenn bygger ned mest natur”) on the per-capita metric. The setup can be read as letting every news organization pick the denominator that makes its own country stand out.

Method and uncertainty placed beneath the numbers

The bars in the country charts carry only point estimates. The uncertainty is gathered on a separate methodology page, signed by Salzenstein and NINA’s Zander Venter.

The input is Dynamic World, from Google and the World Resources Institute, which converts 10 m resolution Sentinel-2 imagery into per-pixel land cover probabilities. For each pixel, a regression line is fitted to the 2018-2023 trend of the “probability of being built up,” and a pixel counts as land take when the slope exceeds 0.05. About 84 billion pixels and nearly 185,000 images were processed.

The area is not obtained by counting pixels. People verified the extracted points, and the estimate corrected for the resulting errors is given with a 95% confidence interval. Across EEA-39 it comes to about 9,000 km², or 1,500 km² per year on average.

Forty journalists and fact-checkers manually verified more than 10,000 locations across Europe, a process that took more than 500 hours in total.

In short, 40 journalists and fact-checkers checked over 10,000 locations by hand, taking more than 500 hours in total. The AI’s judgments were hidden from them, and points were assigned regardless of country. The methodology page also has a figure comparing pixel counts with corrected values and a figure of accuracy by country.

The paper also draws per-capita values by country, with 95% confidence intervals.

Three panels showing per-capita land take by country as point estimates with horizontal 95% confidence intervals, in the order total, natural land, cropland

Source: Venter et al., Nature Communications (2026), Fig. 2 (CC BY 4.0)

The main story shows easy-to-read point estimates, while the methodology page and the paper show estimates with ranges. This can be read as reconciling readability of the numbers with ease of verification by separating them into layers.

At the same time, the sources disagree on some numbers. For each figure I use, here is how the sources compare.

Item	Figures by source
Participating countries	Main story: “11 countries.” Counting the home countries of the two lead organizations and the 10 participating outlets gives 12
Countries analyzed	The methodology page lists 30 countries but also mentions a stratum of “28 countries.” The paper has 29 countries, Datawrapper 30 (including Ukraine)
Number of stories	About: “51 stories in 11 languages”; application: “51 stories published in 10 languages”; Arena’s award announcement and The Spark: “47 publications in 10 languages”
Verified points	Main story and methodology: “more than 10,000”; paper: “9,893”
How clear-sky months were chosen	Methodology: “four-month period”; paper: “three consecutive months”
Ratio to earlier estimates	Main story: “up to one and a half times faster”; paper’s abstract: “nearly twice as large”
Norway’s share of land	The CSV distributed for the chart says 0.03% (which would rank 29th). The share-of-land chart rewrote this value to 0.06% in an edit on October 1, 2025

The last row is something I confirmed from the edit history that remains on the Datawrapper chart.

How third parties saw the work

On the Sigma Awards listing site, the work is labeled “2026 · Winner · Large.” The judges’ comment cites the clarity of the presentation.

What’s more, the presentation is intuitive and married with strong storytelling and on-the-ground reporting. Sometimes it can be hard to see environmental impacts over time. This project made it easy.

In short, the presentation is intuitive and tied to storytelling and on-the-ground reporting; environmental impacts over time are hard to see, and this project made it easy.

On April 13, 2026, a paper was published in Nature Communications. Its authors include people from NINA, NRK and Arena, so the part that counts as third-party is the peer review. On per-capita values, the paper says Norway and Estonia are the largest and Switzerland and Spain the most efficient. It also introduces a fourth denominator, land consumption per change in population. Among countries whose population fell, Hungary is highest at “250 m2 land take per citizen lost.”

The British commentary magazine Country Squire Magazine (March 11, 2026), citing the UK edition of the Guardian, wrote that the UK is “fifth-worst in Europe.” It then explained this as a ranking relative to the size of the country. In the shared data, however, the UK is 5th by total area, 4th by share of land, and 26th per capita. This looks like a mix-up of denominators. I have not been able to check the Guardian’s text. It can be read as an example of a ranking traveling apart from its denominator.

In a feature by Greece’s iMEdD Lab (March 10, 2026), Şentek says that in many cases where they re-checked, the AI was right and the human was wrong. In the range I have searched, I have not found any article that criticizes the method itself.

How to use it

What you can take from this work is not to settle on a single denominator. Whether you choose total area, per capita or share of land, each is a claim. Green to Grey makes all three into charts of the same form, and the text also puts all three into words.

However, the three charts are not lined up in a single figure. Unless the reader presses the buttons, the reshuffling does not come into view. If you want to show the reshuffling of ranks itself, you could also choose a figure that links the three ranks with lines. This work seems to have left the reshuffling to the words of the text.

The dot map removed meaning from dot size. If you do not use size, it is safer to make sizes uniform or to note that size is not used. Here the random radii leave room for misreading quantity.

Uncertainty can be placed in a separate layer. The main story relies on point estimates, and the methodology page and the paper show the ranges. Finally, check that the table you distribute matches the chart on screen. Norway’s 0.03% versus 0.06% was something I could not have known until I saw the edit history. The more a work depends on choosing a denominator, the more you need to check the values of the denominator itself.

References

Primary sources

Zeynep Sentek, Jelena Prtoric, Hazel Sheffield, Leopold Salzenstein, “Green to Grey: How Europe is squandering the little nature it has left,” Arena for Journalism in Europe et al., 2025-10-01. https://greentogrey.eu/
Léopold Salzenstein, Zander Venter, “Data Methodology,” Green to Grey. https://greentogrey.eu/methodology/
Green to Grey, “About.” https://greentogrey.eu/about/
Green to Grey, “Publications.” https://greentogrey.eu/publications/
Country bar charts (Datawrapper): total area https://datawrapper.dwcdn.net/p3y79/5/ (data https://datawrapper.dwcdn.net/p3y79/5/dataset.csv), per capita https://datawrapper.dwcdn.net/D5ozx/7/, share of land https://datawrapper.dwcdn.net/2bLXf/5/
Dot map data: https://greentogrey.eu/data/spots-data-static.csv, https://greentogrey.eu/data/spots-data-animated.csv
Site JS / CSS: https://greentogrey.eu/_next/static/chunks/143-d8b84dbd4ceb054a.js, https://greentogrey.eu/_next/static/css/bb6b167e4a7bb8eb.css
Sigma Awards 2026 entry page (original application text and Japanese translation): https://sigma-awards.data-visualization.jp/entries/2026-s-0003/
Royal Geographical Society, “Green to Grey” (interview with the creators). https://www.rgs.org/about-us/what-is-geography/geovisualisation/green-to-grey/
Arena for Journalism in Europe, “‘First of its kind’: Behind the scenes of Green to Grey’s groundbreaking data investigation,” 2025-10-03. https://journalismarena.eu/first-of-its-kind-behind-the-scenes-of-green-to-greys-groundbreaking-data-investigation/
Arena for Journalism in Europe, “Arena’s groundbreaking data investigation, Green to Grey, wins Sigma Award.” https://journalismarena.eu/arenas-groundbreaking-data-investigation-green-to-grey-wins-sigma-award/
The Spark (TBIJ), “‘Green to Grey’ captured Europe’s changing landscape.” https://the-spark-tbij.beehiiv.com/p/green-to-grey-captured-europe-s-changing-landscape
NRK, “Europa fra grønt til grått: Nordmenn bygger ned mest natur, viser europeisk kartlegging.” https://www.nrk.no/dokumentar/xl/europa-fra-gront-til-gratt_-nordmenn-bygger-ned-mest-natur_-viser-europeisk-kartlegging-1.17577504
NRK, “NRK avslører: 44.000 inngrep i norsk natur på fem år,” 2024-01-05. https://www.nrk.no/dokumentar/xl/nrk-avslorer_-44.000-inngrep-i-norsk-natur-pa-fem-ar-1.16573560
Venter, Czúcz, Kumano-Ensby, Salzenstein, Simensen, Solvang, Støstad, “Europe’s land take and the loss of nature and cropland to artificial surfaces,” Nature Communications, 2026-04-13. https://www.nature.com/articles/s41467-026-71931-w (doi:10.1038/s41467-026-71931-w, CC BY 4.0. Its authors overlap with the work’s creators)

Secondary sources

The Sigma Awards, “From Toxic Supply Chains in Africa to Baltic Sabotage to Stolen Children in Syria: 10 Data Projects Win 2026 Sigma Awards” (judges’ comment). https://www.sigmaawards.org/from-toxic-supply-chains-in-africa-to-baltic-sabotage-to-stolen-children-in-syria-10-data-projects-win-2026-sigma-awards/
Nature Communications (as a peer-reviewed paper). https://www.nature.com/articles/s41467-026-71931-w
“Britain’s Disappearing Countryside,” Country Squire Magazine, 2026-03-11. https://countrysquire.co.uk/2026/03/11/britains-disappearing-countryside
Athina Thanasi, “Green to Gray: The data, the mountains, and the politics of land,” iMEdD Lab, 2026-03-10. https://lab.imedd.org/en/?p=49962

Tracing Track Maps by Hand to Turn Images Back into Maps: The Reporter (報導者), 〈解放軍如何進逼第一島鏈〉 (China's War Drums to the First Island Chain)

Wed, 04 Nov 2026 00:00:00 +0900

The Reporter (報導者) traced by hand the track maps that Taiwan publishes as images and Japan as PDFs, and turned them back into geodata (992 tracks for the Taiwan side). It then layered them onto a map, period by period, inside scroll-driven videos. The track maps show the movements of Chinese People’s Liberation Army aircraft and ships, drawn by Taiwan’s Ministry of National Defense and Japan’s Ministry of Defense.

The piece is 〈【圖解】解放軍如何進逼第一島鏈：台海衝突下牽動的美日軍事布局〉 (China’s War Drums to the First Island Chain), published on October 17, 2023 (Taiwan time). It won in the small newsroom category at the Sigma Awards 2024. Below, I read it in the order in which the data was made, following the article text and its appendix (the methodology notes), the dataset published on GitHub, and The Reporter’s article on Medium.

Chinese Aircraft Routes, Published Every Morning at 9

The article opens with a daily routine of the Ministry of National Defense.

每天早上9點，國防部新聞官準時在媒體群組公布過去24小時的共機騷擾路徑…2022年8月至2023年8月一年間，國防部便監測到近6,000架次共機起飛，其中2,231架次繞行台灣周遭

In short: every morning at 9, a Ministry of National Defense spokesperson posts the routes of Chinese military aircraft from the past 24 hours to a media group. Over the year from August 2022 to August 2023, about 6,000 sorties of Chinese aircraft took off, and 2,231 of them circled Taiwan.

The track maps showing the routes are published as JPG images. The appendix states the premise of the work as follows.

台、日官方提供資料皆為JPG圖檔或PDF文件檔，團隊須先將圖檔轉換為地理圖資檔案，以便後續分析。

That is, the official materials from both Taiwan and Japan are JPG or PDF files, so before any analysis the team first had to convert the images into geospatial files.

The Ministry of National Defense track map for August 4, 2022. Lines with arrows are drawn over the Taiwan Strait, and a table at upper left lists aircraft types and counts

Source: 中華民國國防部 (Ministry of National Defense, ROC) / data-reporter “PLA_Path_Database” (GitHub)

According to the README, maps from before August 5, 2022 were originally PDFs, which The Reporter converted to JPG.

The materials collected are as follows. The Ministry of National Defense issues a track map when military aircraft are “逾越海峽中線及西南空域”, that is, when they cross the median line or enter the airspace to Taiwan’s southwest. The Ministry does not routinely publish ship tracks, so all ship data comes from Japan’s Ministry of Defense materials.

Source	Format	Period	Volume (as described in the appendix)
Taiwan Ministry of National Defense, “即時軍事動態” (real-time military developments)	JPG images	2022-08-01 to 2023-09-15	「共317份活動圖、992條軌跡資料（1份報告可能有數條軌跡）」
Japan Ministry of Defense, Joint Staff Office press materials	PDF	2022-08-01 to 2023-08-31	「共計22份共機資料、99份共艦資料」

Russian military tracks that appear in the Japanese materials are excluded from the analysis, while joint China-Russia exercises are included.

Cutting the White Margins, Aligning the Four Corners, Tracing One Line at a Time

The appendix divides the Taiwan-side process into five steps.

Using image recognition, cut the white margins off every track map and unify the vertical and horizontal dimensions. Because the map extent is the same every time, once aligned the maps can be overlaid.
In the GIS software QGIS, georeference a map and obtain the coordinates of its four corners.
Load all the maps into Figma and draw each track by hand, one line at a time, on separate layers.
Attach the date, aircraft type, and number of aircraft, and export everything together as SVG.
In Python, using the four-corner coordinates as the reference, restore coordinates in EPSG:3857 (Web Mercator) and save them as latitude and longitude in EPSG:4326 (WGS84).

The third step is explicitly described as following another outlet’s method.

參考《天下雜誌》，將所有圖檔匯入繪圖兼介面設計軟體Figma中，在各圖層中逐一人工描繪各段軌跡。

In short: following CommonWealth Magazine (《天下雜誌》), all the images were imported into the drawing and interface-design software Figma, and each segment of track was traced by hand, one by one, on its own layer.

The same day’s track map with the white margins cut and the date “2022-08-04” added at upper right

Source: 報導者 Data Team (The Reporter) / PLA_Path_Database (GitHub, MIT License)

According to the README, the processed maps have their margins cut, the date added at upper right, and their dimensions unified.

For the Japan side, Python searches each PDF for the phrase “行動概要” (action summary) and turns the page containing the map into an image. Maps with the same extent are grouped together and converted to GeoJSON in the same way as the Taiwan maps. The remaining maps, whose extents differ, are aligned one by one using intersections of latitude and longitude lines or terrain points as references.

In both cases, the final step is to draw the GeoJSON in mapping software and compare it with the original map by eye to check for positional errors.

Looking at how the process is divided, it can be read that the machine took on aligning the images and computing coordinates, while picking out the lines was left to people. The appendix also says that this manual work affects accuracy.

由於軌跡是從官方圖檔復刻而來，各重繪環節都會使資料精準度打折，地圖相對適合從「宏觀」的小比例尺觀之，並不適合過度拉近為大比例尺進行解讀（位置可能會存在誤差）。

In short: because the tracks are reproduced from official images, every redrawing step reduces the accuracy of the data; the map is better suited to viewing at a “macro”, small scale, and is not suited to zooming in too far and reading at a large scale (positions may contain errors). This sentence can be read as telling readers up front to what scale they can trust the map.

Three Scroll-Driven Videos

The retraced tracks are shown inside “捲動式影片” (scroll-driven video). Zeng Wei-ting (曾偉婷), The Reporter’s product manager, explains it as follows in the Medium article.

「捲動式影片」將瀏覽影片從點擊播放的行為，轉為使用者更為熟悉的網頁向下滑動的閱讀操作模式，而且影片「播放」速度由閱讀者自己掌控

In short: it turns watching a video from clicking to play into the more familiar reading action of scrolling down a web page, and the reader controls the “playback” speed.

Three videos are embedded in the article.

Video	Main content	Last subtitle timestamp
Video 1	The First Island Chain, U.S. overseas bases as of 2021, Self-Defense Force deployments in Okinawa, Ishigaki Island	28.8 s
Video 2	Tracks of military aircraft around Taiwan	41.5 s
Video 3	Tracks of military aircraft and ships around Japan, missile ranges	28.7 s

The embed code sets secondsPer100vh: 1.2, so scrolling by one screen height advances the video by 1.2 seconds. Subtitles appear tied to the video’s time. By this setting, it can be read that getting through Video 2 to the end takes roughly 35 screen heights of scrolling.

Video 2 does not show the whole year at once. After the first subtitle, it splits the year as follows.

以3次解放軍發起大型軍演，軍機數量達最高峰的關鍵時間點做切分

In short: it splits at the key points in time when the PLA launched three large-scale exercises and the number of military aircraft peaked.

Period	Event that marks the break
August to end of November 2022	The first encirclement exercise on August 7, 2022, after U.S. House Speaker Nancy Pelosi’s visit to Taiwan
December 2022 to end of March 2023	Exercises after President Biden signed the National Defense Authorization Act on December 23, 2022
April to August 2023	Exercises after President Tsai Ing-wen returned from her U.S. visit, April 10, 2023

For each period, the subtitles describe how the routes widened from the southwest airspace to the Bashi Channel, and then to the east of Taiwan. By aircraft type, fighters account for 69% (1,541 sorties) and reconnaissance aircraft for 16% (357 sorties), with drones about half of reconnaissance (177 sorties). The video then pulls out, in turn, the tracks of reconnaissance aircraft, Y-8 anti-submarine patrol aircraft, and Ka-28 anti-submarine helicopters. The subtitles say the Ka-28 tracks cluster along the edge of the Taiwan Strait where the depth drops sharply from 100 m to 2,000 m.

A video frame with unmanned reconnaissance aircraft tracks overlapping around Taiwan. Photos and descriptions of the Y-8 and BZK-005 are at left, and subtitles at right

Source: 報導者 The Reporter / Medium 〈捲動式影片2.0：最受歡迎的數位敘事模式再進化〉 (Scrollytelling Video 2.0: The Most Popular Digital Storytelling Format Evolves) (one frame from a GIF)

According to the description attached to this GIF, the subtitles on the right are a standardized component on the page side, while the illustrations of aircraft types on the left are drawn into the video itself.

A video frame in which military aircraft tracks for September 1 to 15, 2023 overlap from the Taiwan Strait to the southwest of Taiwan, with subtitles at right

The table below summarizes the correspondence between data and visuals, as far as it can be verified. This article has not verified the correspondence between line colors and the legend.

Data	Visual / material	Basis
One track (with date, aircraft type, number of aircraft)	A line on the map. One line may represent multiple aircraft or ships	Appendix, “三、資料呈現” (3. Data presentation)
Ship positions and activity areas	Circles	Note in the Video 3 subtitles
Period	Tracks are added for each of the three periods	Video 2 subtitles
Progress of time	The reader’s scroll distance (1.2 seconds per screen)	Embed code setting
Aircraft type	The breakdown in the subtitles, and tracks pulled out by type. Illustrations of aircraft types are drawn into the video	Video 2 subtitles, Medium article
Missile range	Ranges drawn on the map	Video 3 subtitles
Device	Landscape 1920×1080 video (PC) and portrait 720×1080 video (width 767 px or less)	Embed code

A dark map of East Asia, with tracks of connected lines and circles overlaid from around Taiwan across Japan’s Nansei Islands, Kyushu, and the Sea of Japan

Source: 報導者 The Reporter / 〈【圖解】解放軍如何進逼第一島鏈〉 (China’s War Drums to the First Island Chain) (the article page’s share image)

In the article page’s share image, in addition to the lines around Taiwan, a shape of connected circles can be seen in the Pacific from east to south of Taiwan. The note in Video 3 tells us that the lines and circles carry over the drawing conventions of the original government maps.

每一線段或圓點可能包含多架共機、艦，圓圈為共艦點位或活動範圍，部分軌跡同時包含俄軍資料（聯合軍演）。

In short: a single line or point may contain multiple aircraft or ships; circles are ship positions or activity areas; and some tracks also include Russian military data (joint exercises). The Japanese materials sometimes list the same track repeatedly, and The Reporter draws it only once.

Two Numbers Left in the Dataset

The Reporter has published the retraced data on GitHub as “PLA_Path_Database” under the MIT License. The README is dated September 23, 2024, when data through July 31, 2024 was added.

The file used at the time of publication is Taiwan/20220801-20230915_fixed.geojson. The Taiwan README describes this data as “2024年9月修正錯誤” (errors corrected in September 2024). Its figures do not match those in the article.

What was counted	Maps / days	Tracks
Article appendix and README	317 maps	992 tracks
Tally of the corrected GeoJSON (by date and feature)	322 days	995 tracks

The article counts by number of maps, and the tally counts by number of dates. Neither the article nor the README explains the breakdown of the difference.

The Japan-side Japan/jp_data_cleaned.csv has 327 rows: 266 for ships and 61 for military aircraft. Each row records the shape of the original figure: arrow, double-headed arrow, ellipse, or group. The version column leaves the handling of duplicates to the user.

防衛省可能在新報告中，連帶標記日前公布過的軌跡，因此軌跡會重複描圖。如果欲去除重複，僅選擇new即可。

In short: the Ministry of Defense may include in a new report tracks it had published earlier, so tracks get traced twice; to remove duplicates, just select “new”.

From January 16, 2024, the Ministry of National Defense’s maps stopped showing tracks by aircraft type and switched to showing activity areas as polygons (README: “國防部不再公布各機型之航機，改以活動範圍多邊形註記”). Later data is stored as polygons.

How Third Parties Saw the Work

According to the Sigma entry page, the work was a 2024 winner. The judges praised its storytelling.

The project took dry military data and turned it into a compelling visual narrative to highlight the dangers of a war breaking out at one of the world’s key flashpoints. It made the risks of a confrontation clearer to an audience that may not have been focusing on the issues.

In short: it turned dry military data into a compelling visual narrative that shows the danger of war breaking out at one of the world’s flashpoints, and made the risk of confrontation clearer to an audience that may not have been paying attention to the issue. According to The Reporter’s announcement, it was one of 10 chosen from 591 entries from 322 organizations in 78 countries, and the only Chinese-language winner that year.

At the SOPA Awards 2024, it received an Award for Excellence in the Chinese-language category of Excellence in Infographics.

This piece adeptly demystified a highly intricate subject matter, successfully condensing an immense volume of data and information into a comprehensible and accessible format. The visual style maintains a consistent and elegant aesthetic, and the incorporation of animation further assists readers in grasping the narrative.

In short: it untangled a highly intricate subject and condensed an immense volume of data and information into an accessible form; the visual style is consistent and elegant, and the animation helps readers follow the narrative. The Honorable Mention in the same category went to CommonWealth Magazine’s (《天下雜誌》) 《2023 台北都更地圖》 (2023 Taipei Urban Renewal Map).

In Taiwan, the Central News Agency reported that it received the Innovation in News Narrative (不分媒體類新聞敘事創新獎) award at the 23rd Excellence in Journalism Awards (第23屆卓越新聞獎) on November 19, 2024.

The entry form self-reports that The Reporter worked with The Washington Post (《華盛頓郵報》), providing the dataset and translating the track analysis into English. I have not been able to read the contents of the English version.

All three evaluations focus on presentation: the narrative, the animation, and innovation in storytelling. The process of turning images back into geodata does not appear in their wording, and it can be read as remaining on the side of the appendix and the dataset.

How to Use This

What can be taken from this work is that even materials the government publishes only as images can be turned back into data if you follow the steps.

If the map extent is the same every time, cutting the margins and aligning the images lets you overlay them. Once you have the four-corner coordinates, you can convert the traced lines back to latitude and longitude. For maps whose extents differ, align them one by one using latitude and longitude intersections or terrain points as references.

Attach the limits to the data you restore. In the appendix, The Reporter wrote that the map should be viewed at a small scale and not zoomed in to be read. It also carries over, in notes, the convention of the original maps that one line can represent multiple aircraft. It hands the handling of duplicates to the user through a column in the dataset.

In presentation, it places the time breaks on events and leaves the progress of time to the reader’s scrolling. By adding tracks in three periods, it can be read as making the widening of the range easy to follow in sequence. Both the effort of restoring the data and the effort of showing it step by step make up the substance of this work.

References

Primary sources

報導者 The Reporter, 〈【圖解】解放軍如何進逼第一島鏈：台海衝突下牽動的美日軍事布局〉 (China’s War Drums to the First Island Chain), October 17, 2023 (article text, video subtitles, appendix “《報導者》團隊如何收集及呈現共軍軌跡資料”, data sources). twreporter.org
報導者 The Reporter, 《中國虎視下的島鏈──沖繩如何成為台海危機的熱點？》 (the special feature it belongs to). twreporter.org
Sigma Awards 2024 entry form (entry page; original and machine translation). sigma-awards.data-visualization.jp
data-reporter “PLA_Path_Database” (README, Taiwan/README, Japan/README, LICENSE, GeoJSON, CSV), MIT License. GitHub
data-reporter “PLA_Path_Database”, exploration map for the Japan side. GitHub Pages
曾偉婷, 〈捲動式影片2.0：最受歡迎的數位敘事模式再進化〉 (Scrollytelling Video 2.0: The Most Popular Digital Storytelling Format Evolves), 報導者 The Reporter (Medium), September 12, 2024. Medium (the text was read via Medium’s RSS)
報導者 The Reporter, summary of its 2024 awards. twreporter.org
報導者 The Reporter, 〈AI、Python與手工業，我們如何處理台灣+日本超過400張共軍軌跡圖？〉 (AI, Python, and Handicraft: How We Processed Over 400 PLA Track Maps from Taiwan and Japan; the making-of article that the README points to for methodology details. I could not check its text for this article). Medium

Secondary sources

Sigma Awards 2024 judges’ comment. sigma-awards.data-visualization.jp
SOPA 2024 Awards Winner List (p.24, Excellence in Infographics, Chinese-language category). PDF
Central News Agency, “第23屆卓越新聞獎” (23rd Excellence in Journalism Awards), November 19, 2024. cna.com.tw
The Washington Post, November 13, 2023 (the article the entry form names as a collaboration; not read for this article). washingtonpost.com

Turning 140 Water Marks into a Town's Water Surface — EL PAÍS's "Paiporta bajo el agua" (Paiporta under water)

Wed, 04 Nov 2026 00:00:00 +0900

From 140 water marks measured in Paiporta, EL PAÍS built the surface of the flood. It then overlaid that surface on a 3D model of the town’s buildings reconstructed from cadastral data. The setting is a town in the province of Valencia, Spain, where a river overflowed in the torrential rains of October 29, 2024. The 3D town is a video drawn in Blender, and it plays as the reader scrolls.

The piece appeared on December 29, 2024, in EL PAÍS Semanal, the Sunday magazine of EL PAÍS. It is signed by three authors: Borja Andrino, Luis Sevillano Pires, and Jacob Vicente. At the Sigma Awards 2025 it was selected for the shortlist. This article reads it alongside the piece’s main text and methodology section, the page’s scripts, Sevillano’s portfolio, and the Sigma entry form.

Turning 140 points measured three weeks later into a surface of water

Paiporta is a town of about 30,000 people, cut through by the Barranco del Poyo, a ravine that is normally dry. On October 29, heavy rain upstream made it overflow at around 6:30 in the evening. According to the piece’s text, 46 people died in Paiporta that day, 19 of them found in their own homes.

What EL PAÍS measured were the water marks left on walls, entrances, and garages. The methodology section begins like this.

Durante los días 18, 19 y 20 de noviembre se tomaron in si tu 140 mediciones de marcas de agua en todo el municipio de Paiporta. Es importante señalar que estas marcas corresponden al estado del agua estancada, es posible que en algunos momentos de la riada fuesen incluso superadas.

In short: on November 18 to 20, 140 water marks were measured on site across the town. The note adds that the marks show the state of the standing water, and that at some moment of the flood the level may have exceeded them. The water surface the piece draws is the height at which the marks remained, not the peak. According to the text cards, 73 of the 140 marks are above 2 m, and none is below 1.1 m.

The methodology section also describes the steps that turn points into a surface, in three stages.

Remove the riverbed from the digital terrain model (MDT) and smooth it to make a base surface.
Interpolate a grid by inverse distance weighting (IDW), using the depths of the 140 points and points on the outer edge of the flooded area (depth 0).
Add the interpolated depths to the base MDT to get the height of the water surface across the whole town.

The edge points were collected by researchers at the University of Valencia. Because of these 0 values at the edge, the water surface can be read as constructed so that it ends at the town’s boundary. What is interpolated is not the elevation of the water surface but the depth, which is then added to the terrain.

A bird’s-eye view of a gray town, with small magenta dots scattered at street corners and along the river

Source: Borja Andrino, Luis Sevillano Pires, Jacob Vicente / EL PAÍS, “Paiporta bajo el agua”

This is the poster image (for smartphones) of the video in card 7. The text of card 6 puts a magenta (#ff35ea) band behind the phrase “140 marcas de la inundación”. The magenta dots in the bird’s-eye view can be read as those 140 points. Card 7 continues with “El resultado es un municipio completamente anegado” (the result is a town completely under water).

Rebuilding the town from the cadastre and turning it into video in Blender

The town overlaid on the water surface is also built from data. The shapes of the buildings come from the Catastro (the Spanish cadastre). Heights were estimated from the difference between a surface model and a terrain model, giving an average height per floor. Extreme values were replaced.

The methodology states the estimated and hand-made parts one by one. Windows and entrances were placed at random, based on each building’s perimeter and number of floors. Tree positions were collected by hand, and three bridges and one footbridge were built by hand. The church and the buildings on two streets were built from Google street-level photos. Roads and streetlights come from Catastro vector data, which “pueden no estar completos” (may not be complete).

A 3D image looking down on the town. The streets and riverbed between white and pink buildings are filled with ochre-colored water

Source: Borja Andrino, Luis Sevillano Pires, Jacob Vicente / EL PAÍS, “Paiporta bajo el agua”

This 3D does not run inside the browser. The methodology says that the opening frames and the maps in the piece were all made in Blender. Sevillano also recalls it this way.

The reconstruction was rendered as video and embedded in a scroll-driven narrative format to guide readers through the progression of impact.

In other words, the reconstruction was rendered as video and embedded in a scroll-driven narrative. Reading the script, there are 12 text cards. Each card has a video for desktop and one for smartphones, and the video plays from the beginning when the card appears.

The Blender workspace. On the left, a perspective view of the 3D model of Paiporta; on the right, a bird’s-eye view over an aerial photo, and a list of layers for buildings and trees

Source: Luis Sevillano / luissevillano.net, “Paiporta under water: reconstructing a deadly flood”

Organized from the piece’s methodology and the page’s code, the correspondence between data and visuals looks like this.

Data	Processing / source	Appearance / material	Value
Water marks (points)	Measured on site, November 18–20	Numbers on text cards; magenta dots in the bird’s-eye view	140 points, 73 of them above 2 m
Flood depth (surface)	Points and edge zeros interpolated by IDW, added to the terrain	Ochre layer of water filling the spaces between buildings	Measured minimum 1.1 m
Terrain	Digital terrain model (MDT)	3D ground; line of the cross-section	Cross-section length 150 m
Buildings	Shapes from the Catastro; height from the difference of two models	Extruded boxes; windows and entrances random	—
Trees, bridges, church	Handwork, regional inventory, street-level photos	3D assets	Eucalyptus over 20 m
Ground-floor homes	No procedure given in the piece	Color-coded buildings in a still image	Over 2,000 homes, about 1,000 of them
Underground parking	No procedure given in the piece	Dark pink buildings in a still image	Over 400 buildings, about 5,000 spaces
Mobile phone disconnections	Analysis by Kido Dynamics	Two cumulative lines	57.7% at 20:11

On the text cards, the phrase “dos metros de altura” (two meters high) gets an ochre (#efa83d) highlight. It is in the same family of colors as the layer of water in the images, so the highlight in the text can be read as serving as the legend.

Ground-floor homes and underground parking

The main text uses the 3D town to approach the question of where the people who died were. More than 2,000 homes have rooms on the ground floor, and about half of them, around 1,000, have all their rooms on the ground floor. A home with every room on the ground floor has no upper floor to escape to.

A bird’s-eye view of the flooded town. Most buildings stay white, while some are colored in dark pink and orange

Source: Borja Andrino, Luis Sevillano Pires, Jacob Vicente / EL PAÍS, “Paiporta bajo el agua”

This still image is placed after the paragraph on ground-floor homes, with neither a caption nor a legend. The text puts a dark pink band behind two phrases: “homes with rooms on the ground floor” and “homes with all rooms at street level”. The two colors in the image can be read as corresponding to these two categories. However, it is not stated which color is which.

The next paragraph turns to the basements. More than 400 buildings have at least one underground parking space, and the spaces total about 5,000. Here too a still image appears, with buildings painted dark pink. The methodology does not say which data these numbers come from.

The entry form gives the meaning of this part as follows.

Estos espacios fueron determinantes en la catástrofe: la mitad de las víctimas fallecieron en ellos, atrapadas por la riada.

The claim is that these spaces were decisive in the catastrophe: half of the victims died in them, trapped by the flood. By overlaying the height of the water surface and the structure of the buildings in the same image, the piece can be read as showing in a single picture which homes’ ground floors went under.

The 20:11 alert and the numbers that disagree

The timing of the flooding is shown by different data in a different chart. It is an analysis of mobile phones by Kido Dynamics, counting the devices that dropped off the network from the evening to the night of the 29th and did not reconnect even days later.

A line chart of the cumulative share of last phone connections. The Paiporta line rises sharply after 18:28, while the line for a baseline antenna rises gently

Source: Kido Dynamics, EL PAÍS / EL PAÍS, “Paiporta bajo el agua” (Datawrapper)

The y-axis shows, among the devices whose last connection fell on the 29th or 30th, the share that had ended its last communication by that time. The two lines are the Paiporta antenna and a baseline antenna (Antena basal). In the data, Paiporta is at 26.3% at 18:28, 57.7% at 20:11, and 98.5% at 22:30. The baseline antenna was at 23.7% even at 20:11.

20:11 suena la alarma. Casi el 60% de los dispositivos que desaparecieron ese día lo hicieron antes de esta hora

The note says: at 20:11 the alarm sounds, and nearly 60% of the devices that disappeared that day did so before that time. The main text also gives the water depth at that moment.

Cuando la alarma de emergencias sonó en los móviles de los vecinos, a las 20.11, en muchas calles el agua corría con más de metro y medio de altura.

That is, when the alarm sounded at 20:11, the water was running more than 1.5 m high in many streets.

Here the numbers in the piece and in the entry form disagree. For the same time, the entry form writes “en varias calles el agua superaba los dos metros” (in several streets the water exceeded two meters), citing “numerosos testimonios” (numerous testimonies) as the basis. The number of measurements and the period also differ by source.

Item	The piece	Entry form	Sevillano’s portfolio
Water depth at 20:11	Over 1.5 m (many streets)	Over 2 m (several streets, from testimonies)	—
Number of measurements	140	más de 140 (more than 140)	hundreds (across several affected towns)
When measured	November 18–20 (three weeks later)	—	Days after the flood
Techniques	No mention of machine learning or scraping	Lists AI/Machine learning and Scraping	blender, r, qgis, aftereffects

For the water depth at 20:11, the number a reader saw in the piece is over 1.5 m. The entry form’s “over 2 m” cannot be confirmed within the piece as the depth at the time 20:11. Whether the hundreds of measurements include towns other than Paiporta is also not clear from the sources.

How third parties saw the piece

At the Sigma Awards 2025, 19 judges chose a shortlist of 50 from 498 entries, and this piece was one of them. It is not among the 10 winners. The listing site for Sigma works also carries the piece’s page as part of the 2025 shortlist. Per-entry comments have not been published, and I found no critique that discusses this piece. Here I set out two sources with different points of focus.

The first is a flood depth map made by the CGAT group at the Universitat Politècnica de València (UPV). For the same disaster, researchers and students measured wall marks in the surrounding area, reaching 5,097 points as of January 21, 2025. Like EL PAÍS, it interpolates points into a surface, but it also produces a separate surface with the buildings masked out.

the height of the masked sheet of water that presents the flooding only in the areas without buildings (streets, squares, open field, ….), where it has really occurred

This describes a surface that shows only the places without buildings, that is, where the flooding actually occurred. EL PAÍS does the opposite: it stands buildings up as 3D boxes and intersects them with the water surface. Where the surface cuts through a building, UPV removes it, while EL PAÍS shows it as the height of the water on a home’s ground floor, it can be read. According to press reports, Professor Josep Pardo Pascual, who led the survey, testified in court that the water exceeded 3 m. This agrees, independently, with the entry form’s statement of “over 3 m at some points”.

The second is a 2025 study by Montse Morcate and Rebeca Pardo, which analyzes front-page newspaper photographs in the weeks after the torrential rains.

visual metaphors for the dead were used frequently, especially symbolic images of piled‐up cars

The finding is that visual metaphors for the dead were used frequently, especially images of piled-up cars. This piece also shows piled-up cars in card 5. At the same time, it translates the reason for the deaths into building structure, ground-floor homes and underground parking, and shows it with numbers and 3D forms. Against the metaphors of photojournalism, the visualization can be read as trying to show the mechanism.

Takeaways

The first is to leave the procedure for making a surface from points in plain words: base surface, zeros at the edge, IDW, adding to the terrain. Because the steps are written down, readers can trace how much of the water surface is measured value and where interpolation begins.

The second is to write “measured values” and “made values” separately. Once things become a 3D image, the measured water surface and randomly placed windows melt into the same look. The piece states in the text that the marks are the height of standing water, that building heights are estimates, and that windows are random.

The third is to leave height and time to separate charts. The 3D town shows “how far it sank”, and the cumulative line shows “when they disappeared”. The two are connected by a single time, 20:11.

Finally, check the numbers in entry forms and retrospectives against the numbers in the piece. In this piece, both the water depth at 20:11 and the number of measurements differed from source to source. When introducing the piece, it is safer to put first the numbers the reader saw in the piece.

References

Primary sources

Borja Andrino, Luis Sevillano Pires, Jacob Vicente, “Paiporta bajo el agua,” EL PAÍS Semanal, 2024-12-29 (main text and methodology, “Metodología y recursos”): https://elpais.com/eps/2024-12-29/paiporta-bajo-el-agua.html
Script of the scroll section of the same piece (card text and video file names): https://ep00.epimg.net/infografias/2024/12/paiporta-zona-cero/assets/index.js?v=1
CSS of the scroll section of the same piece (highlight colors): https://ep00.epimg.net/infografias/2024/12/paiporta-zona-cero/main.css?v=1
Cross-section of the same piece (ai2html): https://ep00.epimg.net/infografias/2024/12/paiporta3D/perfil.js?v6666
Chart of mobile phone disconnections (Datawrapper, Kido Dynamics / EL PAÍS): https://datawrapper.dwcdn.net/vMAvT/1/ , data: https://datawrapper.dwcdn.net/vMAvT/1/dataset.csv
Luis Sevillano, “Paiporta under water: reconstructing a deadly flood” (portfolio): https://luissevillano.net/work/paiporta-under-water/
Sigma Awards 2025 entry form (listing site; Spanish original and machine translation): https://sigma-awards.data-visualization.jp/entries/2025-s-0032/

Secondary sources

GIJN, “Announcing the Sigma Awards 2025 Shortlist”: https://www.sigmaawards.org/announcing-the-sigma-awards-2025-shortlist/
GIJN, announcement of the 2025 Sigma Awards winners: https://www.sigmaawards.org/from-flammable-buildings-to-slaverys-hidden-legacy-to-tainted-groundwater-projects-from-10-countries-win-gijns-2025-sigma-awards/
UPV ETSIGCT, “The CGAT group of the DICGF, together with the ETSIGCT, creates a map of the flooding in l’Horta Sud after the episode of October 29, 2024,” 2025-01-28: https://www.upv.es/entidades/etsigct/en/2025/01/28/the-cgat-group-of-the-dicgf-together-with-the-etsigct-creates-a-map-of-the-flooding-in-lhorta-sud-after-the-episode-of-october-29-2024/
NacióDigital, report on the testimony of Professor Josep Pardo Pascual: https://naciodigital.cat/societat/un-expert-en-cartografia-declara-que-laigua-va-superar-els-tres-metres-dalcada-durant-la-dana.html
Montse Morcate & Rebeca Pardo (2025). Cut-Off Low (DANA) in Valencia: Visual Representation of Death and Grief in Photojournalism. Media and Communication 13 (CC BY 4.0). doi:10.17645/mac.10775 / https://www.cogitatiopress.com/mediaandcommunication/article/view/10775

Wind Roses for the Hours the Kilns Were Off: Grist's "Any Way the Wind Blows"

Wed, 04 Nov 2026 00:00:00 +0900

Grist’s investigation “Any Way the Wind Blows” (February 16, 2023) drew wind roses for the Oxbow Calcining plant in Port Arthur, Texas. The figures split the hours into those when all kilns were running and those when kilns 2, 3 and 5 were each shut down. In the wind roses for the hours the kilns were off, winds blowing toward the monitor to the north occur more often. The article presents this result as something that “suggests” a “likely violation” of the Clean Air Act.

Naveena Sadasivam did the reporting and writing, Clayton Aldern the data reporting, and Jessie Blaeser the data visualization. The piece won at the Sigma Awards 2024. The same suspicion had already been reported in 2022 by Public Health Watch and others. What Grist added, it can be read, was a statistical examination of the five-minute data filed in court, and wind roses covering one year from August 2018. Below, I follow Grist’s article and the analysis notebook Aldern published on GitHub, reading in the order the figures appear in the article.

January 10, 2017: it starts with five-minute numbers

The article opens with a scene from one night. Around 2 a.m. on January 10, 2017, a monitor began recording sulfur dioxide (SO2) above the federal standard of 75 ppb. The text lists times and numbers in short lines: “3:30 a.m.: 128.7 ppb,” “5:00 a.m.: 147.8 ppb.”

Beside the text, 60 five-minute readings from 1:30 a.m. to 5 a.m. appear one at a time as you scroll, in a typeface like a liquid-crystal display. Values above 75 ppb are red. Checked against the public data, this sequence matched the five-minute values filed in court. During this period the wind direction was 170 to 197°, a wind blowing from south to north.

The numbers come from a lawsuit. The contractor Port Arthur Steam Energy (PASE) sued Oxbow, and discovery produced thousands of pages of internal documents. At the center is a spreadsheet filed in a Harris County court.

The spreadsheet, which was filed in the Harris County court, contains wind direction, wind speed, sulfur dioxide monitor readings, and kiln behavior information at five-minute intervals from January 2017 through June 2019.

In short: wind direction, wind speed, SO2 and kiln operating status, every five minutes from January 2017 to June 2019. The SO2 values come from a monitor run by the state regulator TCEQ, and Oxbow pulled them in almost in real time. The article says that when the wind blew across the monitor within a 30° arc and the reading passed a threshold, managers received an email.

The opening numbers are a few rows of this table, shown as they are. Before any statistics, the reader sees the same sequence of five-minute values that Oxbow’s employees received, in order of time, it can be read.

Four panels teach how to read a wind rose

Before moving to a year of statistics, the article teaches how to read a wind rose. Six text boxes, each as tall as the screen, run down the page, and as you scroll, four fixed SVGs switch in turn. The first sentence explains how to read the figure.

The length of each bar shows how likely the wind is to blow from that direction.

In short: bar length shows how often the wind blows from that direction.

First panel of the wind rose. It notes that the direction of a bar is the direction the wind comes from, and the length of a bar is its frequency (the dashed circle is 5%)

Source: Grist / Jessie Blaeser / Clayton Aldern, from Grist, “How a Koch-owned chemical plant in Texas gamed the Clean Air Act”

The annotation on the first panel reads “Direction the wind is blowing from,” with “from” in italics. The second panel lays a northwest-pointing arrow over a long southeast bar, and the text goes on: “the long bar at the bottom means the wind often blows from south to north.” The third is a year of data for “All kilns on.” The text says the monitor in the north lies in the path of the sea breeze that blows from south to north.

A wind rose extends its bars toward the direction the wind blows “from.” What matters in this article is the direction the smoke is heading “to.” The long bar in the south becomes the wind that carries smoke to the monitor in the north. The design settles this reversal before the comparison, it can be read.

The table below summarizes what each figure in the article maps the data to.

Figure	Data	Visual mapping
Opening numeric display	Five-minute SO2 readings on January 10, 2017	Numbers shown one by one. Red above 75 ppb, light gray at or below
Wind rose	Wind direction, August 2018 to July 2019	Direction the wind comes from = bar direction. Frequency = bar length (dashed circles mark every 5%)
Same	Wind speed (split at 8, 16 and 25+ mph)	Color of the bands within a bar, from pale gray-green to dark green
Same	Kiln status	Separate figures: All kilns on, and Kiln 2 off, Kiln 3 off, Kiln 5 off
Before-and-after shutdown chart	Hourly median SO2	x-axis = hours before and after shutdown (±24 hours), y-axis = percent change from the time of shutdown. Points = medians, line = LOESS, band = 95% confidence interval
Kiln 4 comparison	Wind direction and speed, 2017 to 2019	Yearly wind roses placed along a time axis, with a dashed line marking the stack replacement

Drawing only the hours the kilns were off

The comparison comes in the fourth panel. At the top is the wind rose for “All kilns on,” and below it are the wind roses for “Kiln 2 off,” “Kiln 3 off” and “Kiln 5 off.”

Wind roses for all kilns running (top) and for kilns 2, 3 and 5 each shut down (bottom). With kilns 2 and 5 off, the bars from the south are long; with kilns 3 and 5, the dark bands are also more common on the south side

Source: Grist / Jessie Blaeser / Clayton Aldern, from Grist, “How a Koch-owned chemical plant in Texas gamed the Clean Air Act”

The text describes this figure as follows.

Compared to normal conditions, however, we noted that when kilns were shut down during this period in Port Arthur, a more extreme bias emerged in the wind roses. Kilns 2, 3, and 5 at the plant were more likely to be shut down when the wind was blowing from south to north.

In short: the wind roses for the hours the kilns were off show a stronger bias than under normal conditions, and kilns 2, 3 and 5 were more likely to be shut down when the wind blew from south to north. The next sentence says they were also more likely to be shut down in strong southerly winds. The scale circles differ from figure to figure: up to 15% for kiln 2, up to 10% for kiln 5, and only 5% for the others. So bar lengths should not be compared across figures; they have to be read by which percentage circle the tip of the bar reaches. The south bar of kiln 2 reaches the 15% circle, and the south-southeast bar of kiln 5 reaches the 10% circle. The south bar for all kilns on only slightly exceeds the 5% circle. The kiln 3 figure is almost the same shape as the all-kilns-on figure, and the bias is hard to read from it.

The period is August 2018 to July 2019, the year after the contract with PASE ended and after the period of “experiments” that Oxbow admitted in court. Kiln 4 was being rebuilt, so it was not used for the determination.

The number in the text is that when the wind blew to the north, at least one of kilns 2, 3 and 5 was shut down about 40% of the time. Tallying the public data for this article gives 40.0% for wind directions of 180° ± 15° and 34.7% for other directions. This is not a figure Grist published, but the difference is about 5 points.

According to the Methodology at the end of the article, five-minute values resemble their neighbors, so the data were re-aggregated by hour and by time windows of random length. A cross-validated random forest was then used to classify kiln status. In the notebook, the difference in wind-direction distributions is D=0.168 by a KS test (p<2.2e-16), and the model’s accuracy is 60.4%. Aldern writes:

Okay, so this a pretty poor model of kiln behavior. But it’s not worthless … We’re doing better than the random guesser.

In short: a rather weak model of kiln behavior, but better than guessing at random. The wind rose shows the difference as a shape, and how certain that difference is, it can be read, is left to the Methodology and the notebook.

The 24 hours around a shutdown, and three years of kiln 4

Next are time-series figures. Taking the moment a kiln was shut down as zero, they show the change in SO2 over the 24 hours before and after.

Percent change in SO2 in the 24 hours before and after a kiln shutdown. Flat before the shutdown, rising from about 10 hours after, and reaching about 600% at 24 hours

Source: Grist / Jessie Blaeser / Clayton Aldern, from Grist, “How a Koch-owned chemical plant in Texas gamed the Clean Air Act”

The subtitle says readings rise by about 600% within 24 hours of a shutdown decision. The text reads:

readings tended to spike in the 24 hours following a shutdown decision, while they were relatively stable in the preceding 24 hours — suggesting that shutdowns were executed in advance of known changes in environmental conditions.

In short: readings were relatively stable before the shutdown and spiked afterward, which suggests that shutdowns were carried out in anticipation of environmental changes. Here too the word is “suggesting.”

The last is kiln 4. Its stack was rebuilt 20 feet taller and about 3 feet narrower. The figure places wind roses for 2017 to 2019 along a time axis and draws a dashed line labeled “Kiln 4 replaced” between 2018 and 2019. The subtitle says that before the rebuild the kiln was more likely to be shut down in southerly winds, and after the rebuild that relationship disappeared.

Wind roses for kiln 4 for 2017, 2018 and 2019, placed along a time axis. In 2019, after the stack was rebuilt, the bars from the south are short

Source: Grist / Jessie Blaeser / Clayton Aldern, from Grist, “How a Koch-owned chemical plant in Texas gamed the Clean Air Act”

According to the text, from 2017 through the first half of 2018, kiln 4 was shut down 11% of the time when the wind blew to the north, against 8% for other directions. In 2019 the correlation was gone.

One night’s numbers, a year’s distribution, the time around a shutdown, three years of a single kiln. The scales differ, but in every figure wind direction is the axis. The wind rose format stays the same and only the conditions change from figure to figure, so readers need not relearn how to read it, it can be read.

How third parties saw the piece

The Sigma Awards 2024 judging committee praised the presentation and the combination of analysis and narrative.

The reporters did a great job of blending data analysis with narrative and visual detail for a compelling and important story.

In short: data analysis blended into narrative and visual detail. The Sigma listing site also carries it as a 2024 winner.

The suspicion itself did not begin with Grist. In “Case closed?” on February 17, 2022, Public Health Watch and the Investigative Reporting Workshop reported PASE’s claim that Oxbow cut production when the wind blew toward the monitor. The article also carries testimony that Oxbow knew from experiments that emissions reach the monitor only when the wind blows within an arc of about 30° toward it. What is new in Grist, it can be read, is that it tested this claim against a year of five-minute values and showed it in figures.

The Fund for Investigative Journalism, which funded the work, reports as an impact that an EPA spokesperson said the agency would respond to the coverage. As the funder, it is not an independent assessment.

There are rebuttals. Oxbow spokesperson Brad Goldstein called the analysis “flawed” and the conclusion “reckless and unsupportable” in the article. According to Public Health Watch on April 29, 2026, Oxbow called PASE’s claim “a conspiracy theory” and ultimately won the lawsuit. On the other hand, Oxbow is said to be voluntarily installing pollution controls on two kilns by April 2027.

The application Grist submitted to Sigma says it found “clear evidence.” That is stronger wording than the “suggests” in the article text. The article text, it can be read, keeps its language within what the figures can show.

Takeaways

There are four takeaways.

First, redraw the same kind of figure with only the conditions changed. The wind rose is a common chart in meteorology. Drawn only for the hours the kilns were off, it becomes a figure that asks how operations relate to wind. Not inventing a new form is the strength, it can be read.

Second, teach the reading before the comparison. A wind rose is easy to mistake for “the direction the smoke is heading,” it can be considered. Grist used three of its four panels for explanation and left the comparison for the last one.

Third, match the strength of the words to the figure. A wind rose shows a difference as a shape, but how certain the difference is cannot be read from the figure. Grist wrote “suggests” and “likely violation” in the text, left the method in the Methodology, and left the model’s weakness in the notebook. Do not claim in words more than the figure can show.

Fourth, publish the data and the code. The notebook and the CSV of five-minute values are on GitHub at Grist-Data-Desk/windblow. You can recount the 40% figure with a different angle width. The more a topic invites rebuttal, the more this openness counts, it can be read.

References

Primary sources

Grist, Naveena Sadasivam & Clayton Aldern, “How a Koch-owned chemical plant in Texas gamed the Clean Air Act” (Any Way the Wind Blows), 2023-02-16. https://grist.org/project/accountability/koch-oxbow-port-arthur-texas-clean-air-act-pollution/
Clayton Aldern, windblow.Rmd (“Any way the wind blows — Comparing kiln shutdowns to wind direction”, February 2023). https://raw.githubusercontent.com/Grist-Data-Desk/windblow/main/src/analysis/windblow.Rmd
Rendered HTML of the same notebook. https://raw.githubusercontent.com/Grist-Data-Desk/windblow/main/src/analysis/windblow.html
GitHub repository Grist-Data-Desk/windblow (README, data, LICENSE). https://github.com/Grist-Data-Desk/windblow
Sigma Awards listing site (Data Visualization Japan), entry page 2024-s-0002 (original text of the application). https://sigma-awards.data-visualization.jp/entries/2024-s-0002/
Grist, Republish page for the article (republication terms). https://grist.org/project/accountability/koch-oxbow-port-arthur-texas-clean-air-act-pollution/republish/

Secondary sources

The Sigma Awards, “Meet the winners of the Sigmas 2024 for data journalism”. https://www.sigmaawards.org/meet-the-winners-of-the-sigmas-2024-for-data-journalism/
Public Health Watch / Investigative Reporting Workshop, Savanna Strott & Riley Rogerson, “Case closed?”, 2022-02-17. https://archive.investigativereportingworkshop.org/investigation/case-closed
Fund for Investigative Journalism, “Grantees uncover efforts to minimize pollution detection at Koch-owned chemical plant in Texas; EPA vows action as a result of story”. https://fij.org/?p=15177
Public Health Watch, “A prodigious Port Arthur polluter will voluntarily install pollution controls — with a catch”, 2026-04-29. https://publichealthwatch.org/2026/04/29/a-prodigious-port-arthur-polluter-will-voluntarily-install-pollution-controls-with-a-catch/

Data Journalism on Visualizing.JP

Drawing Shelling Locations in Two Ways by Certainty: TEXTY's "Under attack"

February 24, 2022: It Starts with a Small Spreadsheet

Checking Place Names in Reports Against a Fire Map

Small Sharp Dots, Large Blurred Dots

Blurred Dots Make Up Most of the Map

How Did Third Parties See This Map?

Uses

References

How Many Mbps Were You Offered at the Same Price? The Markup's "Still Loading"

When the price is flat, speed is what you compare

The opening pair of houses: you can’t compare with your neighbor yourself

One share, laid out under three ways of grouping

Splitting the Kansas City map in two

How third parties saw the investigation

What to take from it

References

On a Day with a High of 36°C, the Heat Index Reached 69°C: The Hindu's 'Chennai's heat index'

On June 9, the high was 36°C

Handmade sensors and the NWS formula

Three numbers that change as you scroll

The four line charts do not share a yardstick

How do third parties see this method?

Takeaways

References

Showing a Ledger of About 80,000 Parcels at Different Scales: High Country News's "Land-Grab Universities"

Turning Land Called a “Donation” into a Ledger

The National Dot Map: Placing 80,000 Parcels on One Sheet

Maps by Cession and by University: Connecting Distant Land with Lines

Parcel Cards and the Caveat of Estimation

How the Views Map to the Data

How Third Parties Saw the Work

Takeaways

References

Showing How 227,000 Was Counted, Split into Maps of Place, Time and Numbers: Reuters' Hong Kong Protest Count

July 1, 2019: four numbers side by side

Showing where it counted, with a map and photos

Showing when it counted, with dots

Adding up the 15-minute bars gives 227,000

How third parties saw this piece

What to take from it

References

Teal before, yellow after: the 2,312 mosques in the FT's "How China is tearing down Islam"

It starts with one mosque in Beijing

From 4,450 locations to 1,714 mosques

Teal for before, yellow for after

The regional scatter plot, and totals that do not match

How third parties saw the piece

Takeaways

References

Total Area, Per Capita, Share of Land: Country Rankings Reshuffle in Green to Grey

Repeating “green to grey” in color

The dot map shows spread, not location

Three denominators, three bar charts

Method and uncertainty placed beneath the numbers

How third parties saw the work

How to use it

References

Tracing Track Maps by Hand to Turn Images Back into Maps: The Reporter (報導者), 〈解放軍如何進逼第一島鏈〉 (China's War Drums to the First Island Chain)

Chinese Aircraft Routes, Published Every Morning at 9

Cutting the White Margins, Aligning the Four Corners, Tracing One Line at a Time

Three Scroll-Driven Videos

Two Numbers Left in the Dataset

How Third Parties Saw the Work

How to Use This

References

Turning 140 Water Marks into a Town's Water Surface — EL PAÍS's "Paiporta bajo el agua" (Paiporta under water)

Turning 140 points measured three weeks later into a surface of water

Rebuilding the town from the cadastre and turning it into video in Blender

Ground-floor homes and underground parking

The 20:11 alert and the numbers that disagree

How third parties saw the piece

Takeaways

References

Wind Roses for the Hours the Kilns Were Off: Grist's "Any Way the Wind Blows"

January 10, 2017: it starts with five-minute numbers

Four panels teach how to read a wind rose

Drawing only the hours the kilns were off

The 24 hours around a shutdown, and three years of kiln 4

How third parties saw the piece