fliarbi/urban-heat-research-corpus

Urban Heat Research Corpus (UHRC) v1.0

What does the world study, invent and report about urban heat? This dataset puts three records of the same problem side by side: 20,422 research papers on urban heat islands and extreme heat in cities (1990–2025) with the claims their abstracts make, 106,458 news articles about heat (2021–2025) coded for 51 subjects, framings and terms, and 4,123 patent families for heat-mitigation technologies (2006–2024) — plus supplementary tables on the companies, funding rounds and Indian company accounts behind the paper's industry results.

Most bibliometric datasets record that a paper exists. This one records what its abstract claims: what the paper says causes urban heat, what it says urban heat causes, who it names as vulnerable, and any temperature figure it reports. The news and patent tables make every figure in the accompanying paper reproducible.

Attention narrows sharply on the way to the market: over a hundred thousand news articles and twenty thousand papers, but 140 companies and 21 funding rounds that state a heat or cooling purpose.

Twenty-one findings from the data

1. 1.Annual papers with an urban-heat term in the title rose from 36 (2000) to 864 (2024) — 24-fold, against 2.72-fold for all articles and reviews.
2. 2.Titles naming "extreme heat" or "heatwaves" overtook titles naming "heat island" in 2015–2019 and led 3,859 to 2,684 in 2020–2024.
3. 3.Among 2,409 papers making a causal attribution, 48% name urbanisation and land-use change and 23% impervious surfaces; the share naming anthropogenic heat fell from 18.4% before 2010 to 7.7% in 2018–2025.
4. 4.Beijing is the most-studied city (381 papers); nine East Asian cities account for 1,166 title mentions against 73 for Cairo, Lagos, Nairobi and Johannesburg combined.
5. 5.72% of news articles about heat are negative in tone, against 34% of all news; 74% are flagged high-priority against 50%.
6. 6.Meteorological agencies are named in 11.6% of heat-news leads, health bodies in 4.6% — heat is reported as weather 2.5 times as often as it is reported as health.
7. 7.Mitigation appears in 28% of research openings and 1.6% of news leads; forecasts and warnings in 22% of news leads and 1.4% of research openings.
8. 8.China holds 31% of the research and 2.3% of English-language coverage; India 9% and 25%.
9. 9.Patent families for cool and permeable pavements grew 11.6-fold between 2010–14 and 2020–24 while the research literature on them grew 2.4-fold; 83% of all mitigation patent families include a filing in China.
10. 10.Recorded heat incidents rose from 1,023 (2021) to 3,116 (2024); India's share of them rose from 8% to 35%.
11. 11.56% of heat coverage is published in June–August; the three busiest days in five years were 18–20 July 2022.
12. 12.Median reported cooling from urban trees is 2.0 °C across 194 studies (interquartile range 0.9–4.0).
13. 13.Between 2021–22 and 2024–25 the share of heat articles raising workers rose from 1.2% to 1.9%, schools from 1.8% to 3.1% and the elderly and vulnerable from 1.7% to 2.3%; wildfire fell from 8.9% to 4.9% and water and drought from 8.7% to 4.3% (country mix held fixed).
14. 14.Climate change is raised in about one heat article in ten and the share fell from 11.6% to 8.8% (0.76×, same direction in all four country groups); deaths did not move (8.4% to 8.4%); heat action plans are raised in 0.4% of heat articles and cooling centres in 0.3%.
15. 15.Sixteen of eighteen news subject rules agree with an independent second reader on 70% or more of sampled articles (median 87%); the two that fail carry no result. The reader judged 2.1% of sampled openings not to be about heat — the corpus inclusion rule's false-positive rate.
16. 16.The share of heat articles with a negative tone fell from 74% (2021) to 67% (2025) while all news moved from 35% to 33%.
17. 17.Wildfire is raised in 27% of heat articles set in Spain and 0.5% of those set in India; tourism in 15% of those set in Italy.
18. 18.140 of 5.05 million registered companies state a heat or cooling purpose in their own description; 5 describe themselves in terms of heat risk or the heat island.
19. 19.Of 21 funding rounds announced with a heat or cooling purpose, 5 infrastructure and debt financings total $2.7 billion and 15 venture, seed and grant rounds total $48 million.
20. 20.Universities and research institutes owned 9% of new heat-mitigation patent families in 2006–2010 and 21% in 2020–2024.
21. 21.Indian cooling companies with revenue above $10 million (n = 55) grew revenue at a median 25% a year over three years against 16% for 12,120 other Indian companies; over five years the rates are 17% and 16%.

Each number above can be recomputed from the files in this repository. The paper gives the method, the validation and the limits for every one of them.

The findings in charts

1. Research on urban heat grew far faster than science as a whole

Annual papers with an urban-heat term in the title rose from 36 (2000) to 864 (2024) — 24-fold, against 2.72-fold for all articles and reviews. Titles naming "extreme heat" or "heatwaves" overtook titles naming "heat island" in 2015–2019 and led 3,859 to 2,684 in 2020–2024.

2. What the literature says causes urban heat

Among 2,409 papers making a causal attribution, 48% name urbanisation and land-use change and 23% impervious surfaces; the share naming anthropogenic heat fell from 18.4% before 2010 to 7.7% in 2018–2025.

3. How much do trees, green roofs and cool roofs actually cool?

Median reported cooling from urban trees is 2.0 °C across 194 studies (interquartile range 0.9–4.0). Air-temperature reductions are smaller than surface-temperature reductions; the measurements table keeps the two apart.

4. Where the research is, and where the news is

Beijing is the most-studied city (381 papers); nine East Asian cities account for 1,166 title mentions against 73 for Cairo, Lagos, Nairobi and Johannesburg combined. China holds 31% of the research and 2.3% of English-language coverage; India 9% and 25%.

5. Research and news describe different problems

Mitigation appears in 28% of research openings and 1.6% of news leads; forecasts and warnings in 22% of news leads and 1.4% of research openings. Meteorological agencies are named in 11.6% of heat-news leads, health bodies in 4.6% — heat is reported as weather 2.5 times as often as it is reported as health.

6. What rose and what fell in heat news, 2021–2025

Between 2021–22 and 2024–25 the share of heat articles raising workers rose from 1.2% to 1.9%, schools from 1.8% to 3.1% and the elderly and vulnerable from 1.7% to 2.3%; wildfire fell from 8.9% to 4.9% and water and drought from 8.7% to 4.3% (country mix held fixed). Climate change is raised in about one heat article in ten and the share fell from 11.6% to 8.8% (0.76×, same direction in all four country groups); deaths did not move (8.4% to 8.4%); heat action plans are raised in 0.4% of heat articles and cooling centres in 0.3%. The share of heat articles with a negative tone fell from 74% (2021) to 67% (2025) while all news moved from 35% to 33%.

7. Every country reports a different heatwave

Wildfire is raised in 27% of heat articles set in Spain and 0.5% of those set in India; tourism in 15% of those set in Italy.

8. Invention: who patents heat mitigation

Patent families for cool and permeable pavements grew 11.6-fold between 2010–14 and 2020–24 while the research literature on them grew 2.4-fold; 83% of all mitigation patent families include a filing in China. Universities and research institutes owned 9% of new heat-mitigation patent families in 2006–2010 and 21% in 2020–2024.

9. Companies and capital

140 of 5.05 million registered companies state a heat or cooling purpose in their own description; 5 describe themselves in terms of heat risk or the heat island. Of 21 funding rounds announced with a heat or cooling purpose, 5 infrastructure and debt financings total $2.7 billion and 15 venture, seed and grant rounds total $48 million.

10. Indian cooling companies

Indian cooling companies with revenue above $10 million (n = 55) grew revenue at a median 25% a year over three years against 16% for 12,120 other Indian companies; over five years the rates are 17% and 16%. The aggregate revenue of the cohort is dominated by a single company and is deliberately not reported; see the paper.

What is in the dataset

News articles by year: 2021: 13,528, 2022: 24,723, 2023: 21,442, 2024: 24,495, 2025: 22,270. Patent families by technology: green roof 2,168, cool / permeable pavement 1,150, cool / reflective roof or coating 422, radiative cooling 206, heat island (general) 177.

The news table gives you the article, not a mystery id. Every row carries the url, the publisher domain, the headline as published and the language, so you can open any article and check any flag on its row. The article body is not redistributed; the flags were computed on the opening (median about 310 characters) and themes, fine_topics and tags describe what the article is about.

Two data-quality flags ship with the table, and the paper excludes both. src_boilerplate (6,484 rows) marks articles that matched the heat rule only through repeated site boilerplate — typically a site-wide weather-alert banner printed on an unrelated story. src_academic (3,248 rows) marks research abstracts from academic sites rather than journalism. Filtering both leaves 96,726 articles, which is the corpus every number in the paper is computed on. The flagged rows are shipped rather than deleted so the difference can be reproduced:

clean = news[(news.src_boilerplate == 0) & (news.src_academic == 0)]   # 96,726 rows

Quick start

from datasets import load_dataset
import pandas as pd

repo = "fliarbi/urban-heat-research-corpus"
papers = load_dataset(repo, "papers", split="train").to_pandas()
causes = load_dataset(repo, "attributions", split="train").to_pandas()
news   = load_dataset(repo, "news", split="train").to_pandas()

# What does the literature blame for urban heat?  (validated categories only)
ok = causes[causes.category_validation == "validated"]
print(ok.groupby("cause").work_id.nunique().sort_values(ascending=False))

# Reproduce the rise-and-fall chart: share of each year's heat articles raising a subject
subjects = [c for c in news.columns if c.startswith("subj_")]
print((news.groupby("year")[subjects].mean() * 100).round(2).T)

# Hold the country mix fixed, as the paper does
early, late = news[news.year <= 2022], news[news.year >= 2024]
w = pd.concat([early, late]).country_group.value_counts(normalize=True)
std = lambda d, c: (d.groupby("country_group")[c].mean() * w).sum() * 100
print({c: round(std(late, c) / std(early, c), 2) for c in subjects})

Questions this dataset answers

What causes urban heat islands, according to published research? Among 2,409 papers making a causal attribution, 48% name urbanisation and land-use change and 23% impervious surfaces; the share naming anthropogenic heat fell from 18.4% before 2010 to 7.7% in 2018–2025.

How much cooling do urban trees provide? Median reported cooling from urban trees is 2.0 °C across 194 studies (interquartile range 0.9–4.0).

Which cities are studied most? Beijing is the most-studied city (381 papers); nine East Asian cities account for 1,166 title mentions against 73 for Cairo, Lagos, Nairobi and Johannesburg combined.

Do the countries that study urban heat match the countries where it is news? No. China holds 31% of the research and 2.3% of English-language coverage; India 9% and 25%.

How do newspapers cover heatwaves? 72% of news articles about heat are negative in tone, against 34% of all news; 74% are flagged high-priority against 50%. Meteorological agencies are named in 11.6% of heat-news leads, health bodies in 4.6% — heat is reported as weather 2.5 times as often as it is reported as health.

Has heatwave coverage changed? Between 2021–22 and 2024–25 the share of heat articles raising workers rose from 1.2% to 1.9%, schools from 1.8% to 3.1% and the elderly and vulnerable from 1.7% to 2.3%; wildfire fell from 8.9% to 4.9% and water and drought from 8.7% to 4.3% (country mix held fixed). The share of heat articles with a negative tone fell from 74% (2021) to 67% (2025) while all news moved from 35% to 33%.

Is climate change mentioned in heatwave news? Climate change is raised in about one heat article in ten and the share fell from 11.6% to 8.8% (0.76×, same direction in all four country groups); deaths did not move (8.4% to 8.4%); heat action plans are raised in 0.4% of heat articles and cooling centres in 0.3%.

Who is inventing heat-mitigation technology? Patent families for cool and permeable pavements grew 11.6-fold between 2010–14 and 2020–24 while the research literature on them grew 2.4-fold; 83% of all mitigation patent families include a filing in China. Universities and research institutes owned 9% of new heat-mitigation patent families in 2006–2010 and 21% in 2020–2024.

Is there a heat-adaptation industry? 140 of 5.05 million registered companies state a heat or cooling purpose in their own description; 5 describe themselves in terms of heat risk or the heat island. Of 21 funding rounds announced with a heat or cooling purpose, 5 infrastructure and debt financings total $2.7 billion and 15 venture, seed and grant rounds total $48 million.

Are cooling companies in India growing faster? Indian cooling companies with revenue above $10 million (n = 55) grew revenue at a median 25% a year over three years against 16% for 12,120 other Indian companies; over five years the rates are 17% and 16%.

How it was built

A paper enters if "urban heat", "heat island", "urban overheating" or "urban cool(ing) island" appears in its title or abstract. Journal articles and reviews only; repository deposits excluded; de-duplicated on title and year. 17,101 papers carry an abstract of more than 200 characters and form the text-analysis set. Selecting by bibliographic topic tag instead returns 122,160 papers, dominated by building-energy and general climate research; text-based selection is the conservative rule and the one used here.

A sentence counts as a causal attribution only when the cause, a causal cue and the heat-island term appear in that order in one sentence (cause → leads to → heat island, or heat island → due to → cause). Counting mere co-occurrence is roughly half as precise: it scores "urban heat islands increase energy consumption" as evidence that energy use causes urban heat.

A news article enters if its opening mentions urban heat, heat islands, extreme heat, heatwaves, heat domes, heat action plans or heat stroke. All news flags are computed on the article opening (median about 320 characters), so they describe how heat stories open. About 2% of articles concern marine heatwaves; the ctx_marine flag lets you drop them.

A patent family enters if a member's title names a cool or reflective roof, green roof, cool or permeable pavement, radiative cooling or the heat island. Titles that merely say "heat-reflecting" must also name a building surface or coating — without that condition the rule also catches ovens, windshields and solar-concentrator mirrors.

Validation — read before using

Every cause category was checked against an independent second reader (a language model run over a stratified sample of 620 abstracts, 565 usable). The table reports agreement, not ground truth.

Three categories failed and carry category_validation = failed_validation in attributions. They are released so the extraction can be improved, not for making claims. The templates also miss about a third of attributions, so every count is a floor. The news subject flags were put through the same test: 18 rules, 912 sampled article openings, 875 usable. Sixteen reach 70% agreement or better; two fail and carry no result in the paper — subj_sport_events_pilgrimage (65.9%) and subj_trees_shade_cool_surfaces (52.6%, because shade is a weather-reporting convention rather than a cooling measure). The reader also judged 2.1% of sampled openings not to be about heat, which is the corpus inclusion rule's false-positive rate. Per-subject figures are in news_validation.csv. The consequences, vulnerable_groups and measurements tables have not yet been through this check; treat them as provisional.

Limits

1. 1.Abstracts, not full texts. A claim made only in a paper's body is invisible.
2. 2.Sentence text is not redistributed. Each claim carries a character offset, a length and a SHA-1 prefix of the sentence, so it can be verified against the paper's abstract.
3. 3.Study region is inferred from place names in title and abstract, not from author affiliation.
4. 4.`cited_by_count` is a snapshot taken on 18 September 2026.
5. 5.Reported temperature figures are not effect sizes. They mix measured and simulated values, peak and mean, air and surface.
6. 6.English-language bias in research and news alike. Country is where a story is set, and about half of articles carry none.
7. 7.The mix of news outlets changes over time. Compare years with the country mix held fixed (country_group), as in the example above.
8. 8.Use `term_in_title` for research growth series. The share of records with an abstract varies by year and inflates growth if ignored.
9. 9.Company, funding and patent counts are floors: they count only records whose own text states the purpose.

Citation

@misc{kumar2026uhrc,
  author    = {Kumar, Narendra},
  orcid     = {0009-0001-7468-6100},
  title     = {Urban Heat Research Corpus (UHRC) v1.1: research, news and patents on urban heat},
  year      = {2026},
  publisher = {Hugging Face},
  doi       = {10.57967/hf/10505},
  note      = {Fliar BI. Archived copy: doi:10.5281/zenodo.22840157. Accompanies the paper: What the World Studies, Invents and Reports About Urban Heat},
  url       = {https://doi.org/10.57967/hf/10505}
}

Mirrors: Zenodo (archival, versioned) and Kaggle (notebooks). More evidence maps built from the same records: fliarbi.com/prism/explore.