The Heat Cities Make

The heat

An old question about a modern problem

Cities are warmer than the countryside around them — the urban heat island. Part of that warmth is built in: dark roofs and pavements soak up the sun, and street canyons hold the heat overnight. But part of it is made, continuously, by the city itself. Furnaces and boilers, engines and air conditioners, factories and data centres, even the metabolism of millions of human bodies — all of it ends up as heat released into the urban air.

The idea is nearly two centuries old. Luke Howard, cataloguing London's unusual warmth in the 1830s, guessed that the city's fires and its inhabitants were partly responsible — and was struck less by the warming than by how small it seemed:

“The real matter of surprise, when we contemplate so many sources of heat in a city is, that the effect on the thermometer is not more considerable.”

Luke Howard, The Climate of London (1833)

Howard's surprise turns out to be the right question. In 1877, working largely from London's coal consumption, Henry Storks Eaton produced the first number — the oldest record in this review's evidence corpus: roughly 12 watts of waste heat for every square metre of the city. Left unchecked, he reckoned, that would warm London's air by more than a degree every hour. It plainly does not — so something must be carrying the heat away. How efficiently the atmosphere does that carrying, and what happens when it falters, is the scientific heart of this review.

A century and a half later, anthropogenic heat remains one of the least well-constrained terms in the urban energy budget — even as the stakes have risen with air conditioning, data centres, electric vehicles, and the rapid growth of cities in a warming climate.

The framework

Three questions for every study

The review reads 150 years of research through the language climate science uses for the planet as a whole: a forcing pushes on the system, the system responds, and feedbacks amplify or damp the response. Asked of a city rather than the globe, that becomes three questions:

Forcing

up to hundreds of W m⁻²

How much heat do human activities release — and where, and when?

Sensitivity

≈ 0.01 K per W m⁻²

How much do urban temperatures rise per unit of heat released?

Feedback

warmth that makes more warmth

How does the warming change the heat cities release — air conditioning above all — amplifying or damping the response?

The premise is that the climatic significance of anthropogenic heat depends not only on how much heat is released, but on how urban temperatures respond to it and how feedbacks modify that response. The same three questions organise the review's sections — and the evidence corpus you can explore on this site.

The story in a minute

Watch: the heat cities make

A one-minute hand-drawn film follows a single curl of waste heat through the three questions above: made by a city, carried off by the air, and caught for a while in the feedback loop on a still, hot night.

Best with sound; the captions carry it without. Every figure on screen comes from this page and the review’s evidence corpus.

Question one · Forcing

How much heat does a city release?

Averaged over the whole planet, human heat release is tiny — a rounding error next to the sun. But cities concentrate it ferociously. In quiet suburbs the flux is a few watts per square metre; in dense urban cores it reaches tens to hundreds. At the 250-m urban-grid scale, Ichinose et al. (1999) estimated a winter peak of about 1,590 W m⁻² in central Tokyo. At the much smaller scale of a single building footprint, Sailor et al. (2026) estimated that a 36-MW data centre rejected the equivalent of 2,800–6,200 W m⁻² — several times the roughly 1,000 W m⁻² peak solar irradiance on a clear day. The larger number is inferred from facility load and footprint, not measured as a city-grid flux; the contrast therefore shows why any claimed “maximum” depends on the area and time over which heat is averaged. And the forcing never sits still: it follows the morning commute, the office day, the weekly rhythm of work, and the seasonal swing between heating and cooling.

The review traces how a question once answered with a city's coal bill has become a global mapping effort — sector-by-sector inventories, satellite remote sensing, energy statistics and machine learning now combine to map human heat release worldwide, hour by hour. The frontier has moved from how much to how: the same total heat warms a street very differently if it leaves through a rooftop chiller, a car exhaust, or a factory stack.

From the globe to the street, and through the day

Click any city on the map to zoom into its heat at the dataset's native 1-kilometre grid — then scrub the months, flip to the 2050s scenario, and watch the two strips below the panel: how the heat swings across the year, and how it rises and falls within a single day. Hover anywhere to read the flux.

One instrument, two scales: the world overview and the selected city both drawn from the AH4GUC global 1-km dataset (Varquez et al., 2021, CC BY). The overview is max-pooled to ≈14 km so cities stay visible; each city panel is the native grid. The 2050s frames are a single scenario — RCP8.5 with SSP3, no adaptation or mitigation — not a forecast.

Two cautions worth carrying: these are monthly and annual averages from a global model, a different kind of number from the hourly wintertime Tokyo grid-cell peak and the building-footprint data-centre estimate above. Both are estimates rather than direct heat-flux measurements, and their averaging areas differ by orders of magnitude. A panel's hottest pixel is usually not its city centre but an industrial point source: Singapore's peak sits on the Jurong Island refineries, Tokyo's on the Kawasaki–Yokohama waterfront. Comparing datasets, scales and definitions is precisely what makes this forcing hard to pin down.

The lower strip is the day itself, and it comes from the same dataset's hourly product — a separate archive from the monthly maps above, held on the authors' own server rather than in the public repository. The curve is the mean over a half-degree box on the city centre — 56 kilometres north to south, and narrower east to west the further from the equator — on the clock the dataset itself keeps, which runs in whole hours, so Delhi's half-hour zone is rounded and Moscow sits an hour ahead of the time the city keeps today. The other eleven months are drawn faintly behind the one you have selected, and the opposite epoch dashed. Two things are worth watching. Every one of these cities troughs at about three in the morning and runs three to six times higher at its peak — a swing that dwarfs the seasonal one, which even in Moscow, the sharpest of them, spans only about seventy per cent between the darkest month and midsummer. And the hour of that peak moves with the season: in the dataset's construction the daily shape is chosen from four templates according to the month's mean temperature, so cool months peak around six in the evening and the warmest ones closer to four in the afternoon, as heating gives way to cooling. That also means the shape is a climatological template rather than a measurement of that particular city — what genuinely varies from place to place is the amplitude, the season, and how much the 2050s add.

Question two · Sensitivity

How much warmer does the heat make the city?

Release a steady extra watt per square metre into a city's streets — how much does the air warm? Across theory, observations and models, the answers centre on about one hundredth of a degree per watt per square metre: a district releasing 100 W m⁻² is roughly 1 °C warmer than it would otherwise be. The centre is firm — the middle half of studies sit within about a factor of two of it — but the tails run wide.

The neatest observational evidence comes from a natural experiment hiding in the working week. Commercial districts power down at weekends; residential ones barely change. Comparing the two across weekdays and weekends in Osaka and Tokyo (Kikegawa et al., 2014) isolates the warming caused by human heat — and lands on that same figure, about 0.01 K per W m⁻², rising to 0.012 at its late-afternoon maximum.

The number is not a constant, though — and its variations carry the physics. Sensitivity is highest at night, in winter, and in calm weather, when the atmosphere is poorly mixed and slow to carry heat away; it is lowest on breezy summer afternoons, when convection whisks the heat aloft. That is why anthropogenic heat leaves its clearest fingerprint on night-time and wintertime warmth — exactly when the sun is weakest. One complication the synthesis must respect: the literature measures this sensitivity in two ways — two quantities with the same units that are not the same thing.

Half a century of answers to the same question: 81 reviewed estimate scenarios from 44 studies, 1969–2026. One study-level mark per paper, at the geometric mean of the estimates its source genuinely supports, with a bar over the study's full range. Blue is forcing-based S_f, orange effective S_e; hover or tap a definition above to pick its studies out of the plot. The faint dashed lines mark each kind's median, and the effective one sits lower because a positive source feedback enlarges its denominator. The dark dashed line is the median across all studies, 0.017 K per W m⁻², and the shaded band its interquartile range, 0.0078–0.029; the reference line marks the observational benchmark from Kikegawa et al. (2014), the weekday–weekend natural experiment in Osaka and Tokyo that reports an observational sensitivity of about 0.01 K per W m⁻². Before 2000 the sparse record is grouped one slot per decade; from 2000 the axis runs year by year. Click any point to open its paper in the corpus explorer.

Question three · Feedback

When warming makes more heat

Here the story closes its loop. A warmer city runs more air conditioning; air conditioners pump the heat they remove — plus the electricity they consume — into the streets; the streets get warmer still. Heat begets heat. The review formalises this the way climate science treats the planet's feedbacks, separating three ingredients: the atmosphere's baseline ability to carry heat away (the strong restoring pull that answered Howard's surprise), restoring feedbacks that alter that ability as the city warms, and source feedbacks — air conditioning above all — that add heat as the city warms.

The balance matters most when it is closest to tipping. The baseline restoring pull is normally an order of magnitude stronger than the feedbacks — which is why cities do not run away to Eaton's degree-per-hour. But during heatwaves and on calm nights, in dense districts saturated with air conditioning, the amplifying loop gains on the restoring one, and the same added heat buys much more warming — precisely when heat is most dangerous. The review's frank conclusion is that this regime is also where the evidence is thinnest: only a handful of studies have put numbers on the feedback loop, and it stands as the field's most important open problem.

a the loop

Hover a feedback to trace its estimates in b ↓

b the estimates

The loop, and what the literature can put numbers on. a, human heat warms the air (red node); the warmth raises cooling demand and so heat release — the amplifying source feedback — while the atmosphere's restoring processes drain heat away (green). b, the review team's verified feedback estimates, plotted by magnitude on a log axis: dots with bars are seasonal medians and interquartile ranges across 5,410 US urban grid cells from Wang et al. (2023); bare dots are individually curated study values from building-energy and empirical work, and the dot-free bar is an empirically derived seasonal range. The neutral shaded band envelopes the summer and winter interquartile ranges of the baseline restoring pull. Solid marks are summer, open marks are winter; signs live in the tooltips — the baseline pull always damps, and the source feedback flips from amplifying in summer to damping in winter. In summer the restoring feedback nearly cancels (its range straddles zero), so its lone dot sits far left. The baseline runs ~100× stronger than any feedback, which is why cities stay stable. Every plotted value is a verified corpus record — click any mark to open its paper in the corpus explorer. Additional reviewed values remain in the corpus with their selection decision recorded, including the Takane et al. (2024) values not selected for this figure. Hover either feedback in a to pick out its lane in b.

The evidence

How we read 150 years of literature

Reviews are only as good as the literature they stand on — and this literature is scattered across a century and a half and 183 different venues, from mainstream climate journals to long-defunct regional bulletins. No single reader holds it all in view. So the team paired machine reading with human judgement, and documented every step.

3,035 records from a locked Web of Science search, tuned for recall across 150 years of terminology

screened by a language model,
checked by hand

a language model read every title and abstract and proposed include or exclude
262 borderline cases decided independently by two authors, reconciled by a third
the model's rejections audited against human readers; citation chasing recovered older, poorly-indexed work

577 full-text papers in the locked evidence corpus, each read against a structured extraction schema

The result is a locked corpus of 577 studies, spanning 1877 to 2026, with every screening decision, audit and extraction documented in the paper's supporting information. Each paper was then read in full and its quantitative evidence — heat-flux magnitudes, temperature sensitivities, feedback behaviour — captured in a structured, comparable form. That structured evidence is what powers the corpus explorer on this site.

A century and a half in one view: every paper in the 577-study corpus, placed by its year of publication. The axis breaks after the earliest isolated records — the field was carried for decades by a mere handful of studies — to give room to the modern surge, when satellite data, urban models and the air-conditioning boom drove publication from a trickle to a flood. The sparse mid-century decades are grouped into ten-year bins; from 2000 the record is dense enough to show year by year. The marked studies trace the through-line, from Eaton's first 1877 estimate to today's global datasets and feedback parameterisations. Hover a marker for the study; click to open it in the corpus explorer.

Explore

Where would you like to go next?

Go deeper into the evidence, meet the people behind the synthesis, or continue to the review itself.