Methods and sources

How this map was made, and what it rests on

What the globe on the front page shows, the rule behind each colour, how well each layer matches published work, where every number comes from, and where we could be wrong. It is written so that a climate scientist can check it.

Last updated November 2026. This method has not yet been reviewed by outside scientists. Tell us where we are wrong.

On this page

In short

  • The map is a long-run picture: where each level of warming leads if it is held for a very long time. It is not a forecast for a date, and we are not saying these temperatures will be reached.
  • Heat uses published three-hourly model results up to 4 °C (Vecellio and colleagues 2023 [51]). Above 4 °C the heat layer is our own extrapolation, and it probably shows too much land.
  • Food and water uses an aridity and growing-season screen on WorldClim climate, warmed with CMIP6 model patterns. It turns out cautious about drying and severe about heat, and the method it rests on is disputed.
  • Sea level uses the range the IPCC assessed for about 10,000 years after warming. The range is wide, and we show it. Above 5 °C the range is our own extension.
  • Only the sea-level layer shows a settled state. The heat layer and the food and water layer show the climate as models have it when a level is first reached.
  • Each layer has a table of how well it matches published work. The misses are listed with the hits.
  • Our dates for 2 °C and 3 °C are our own planning estimate. They are at the pessimistic edge, earlier than almost all published estimates, and may be too pessimistic.
  • Every paper, report and dataset the site uses is in the reference list. A number in square brackets, such as [51], leads to its entry.

1What the map shows

The globe has one control: global warming above the 1850–1900 average, from 1.5 to 6.0 °C. At each setting it shades land that fails one of three tests.

  • Lethal without reliable powered cooling. Heat and humidity pass the measured limit of the human body for a month or more of hours each year.
  • Food and water failure. The land is too dry, or its growing season too hot, to feed and water the people on it without supply from elsewhere.
  • Drowned by sea-level rise. Land below the sea level that the warming level commits the planet to in the long run.
  • Land that fails both the heat test and the food and water test is drawn in dark red.

Land that already fails the food and water test today, which is desert and some dry steppe, is drawn in a muted tone and is not counted as an effect of warming. The map is built from published data and published thresholds. Where we have made a choice or an extrapolation of our own, this page says so. The table gives the short version.

Each layer of the map: its rule, what is published, and what is our own.

LayerRulePublishedOur own
Lethal heat, solid redAt least 240 hours a year above the measured human heat limitThe limit, and the hours at 1 to 4 °C [52] [51]The 240-hour cut. Everything above 4 °C
Lethal heat, lighter redAt least 56 hours a year above the same limitThe sameThe 56-hour cut. Everything above 4 °C
Food and water failure, amber Aridity index below 0.20, or no three-month growing window, or every growing window at 36 °C or more The climate of 1970–2000 and the model projections [9][55]. The aridity classes, the growing-period mark and the crop heat limits behind the thresholds The choice of thresholds, the combination of rules, the scaling to each warming level and the river exemption. Above about 5.2 °C the pattern runs beyond the models
Deserts and dry steppe today, mutedThe same test, already failed at or below 1.5 °CThe sameDrawing it apart and leaving it out of the counts
Drowned coast, blueLand that the eventually higher sea can reach The sea-level ranges at 1.5, 2, 3, 4 and 5 °C [13]. Land heights [34]The central line, the wider band at 1.5 °C, and every value above 5 °C
Hotter than almost anywhere people live today, dotsMean annual temperature of 29 °C or more. Never counted as uninhabitableThe 29 °C mark [56] [27]The scaling to each warming level

2The definition

We use one definition throughout. In the long run, with the temperature held at a given level, a place is uninhabitable when either of two things is true.

  1. It can no longer supply food or water once supply chains break.
  2. Heat is lethal without powered cooling, and the power cannot be relied on.

The definition asks what a place can do for itself. That is deliberate. Our work is preparing towns for the failure of the systems they depend on, so the question we ask of a place is whether people could go on living there if imports and the grid failed.

It follows that the map shows exposure, not abandonment. People live today in places that fail these tests, on imported food, on river water from elsewhere and on air conditioning. The map shows where that dependence becomes total. It does not predict that people will leave, and it does not say that a place cannot be defended.

For the heat layer we use the label “lethal without reliable powered cooling” and not “uninhabitable”. With dependable power, people can live through such heat. The evidence on what happens without it comes from modelling of three cities in the United States: a blackout lasting several days during a heat wave more than doubles the estimated heat deaths in Atlanta, Detroit and Phoenix (Stone and colleagues 2023 [47]).

3A long-run picture, not a forecast

Each slider setting is a level of warming, not a year. The map answers one question: if the global temperature were held at this level for a very long time, where would these tests fail?

  • The map has no dates. We do not say when, or whether, any level will be reached. Section 11 covers what is known about the pace of warming.
  • The layers respond at different speeds. Heat and drought follow the temperature within decades. Sea level follows over centuries to thousands of years. The sea-level figures are for about ten thousand years after the warming.
  • Only the sea-level layer shows a settled state. The climate data behind the heat layer and the food and water layer are states of climate models at the moment a warming level is first crossed, under rising emissions. A climate that had settled at the same level would have a different contrast between land and ocean, different ice cover and a different pattern of rainfall. We have not measured how much this matters, and neither layer was tested against a settled climate. Land warms faster than the ocean while the climate is still changing, so we expect a settled climate to be somewhat less hot over land than these snapshots, at the same global temperature. That would tend to reduce the heat shown. It is our expectation, not a result, and we cannot say which way rainfall would differ. Sea level, by contrast, would be far higher in the settled state than in any snapshot, which is why that layer uses the long-run figure.

4From a global temperature to local conditions

Warming levels are measured against the 1850–1900 average. The present-day climate we start from is WorldClim 2.1 [9], which gives average monthly conditions for 1970–2000 on a grid of 10 arc-minutes (cells about 18 km across at the equator). That period was already about 0.56 °C warmer than 1850–1900 in Berkeley Earth’s earlier land and ocean record (Rohde and Hausfather 2020 [40], last updated December 2024; our calculation). The newer high-resolution record used for our charts [41] gives 0.51 °C. The difference is about half a step of the slider. A slider setting of 3.0 °C is therefore about 2.4 °C of global warming beyond the baseline climate.

A global figure hides large local differences. Land warms faster than the ocean, and rainfall rises in some regions and falls in others. We take these patterns from CMIP6 climate models, using the downscaled projections of monthly temperature and rainfall that WorldClim publishes [55]. We use 22 models. For each one we take the change between 1970–2000 and 2081–2100 under very high emissions (the scenario called SSP5-8.5) and divide it by that model’s own global warming. The global warming comes from the series in the IPCC Working Group I Atlas [24] [22], or from the CMIP6 archive [35] for the five models the Atlas lacks. That gives a local change for each degree of global warming. The pattern we use is the median of the 22 models, cell by cell and month by month.

Temperature at a slider setting is the baseline plus the pattern times the warming beyond the baseline. Rainfall is scaled the same way in logarithms, so it changes by a percentage and cannot fall below zero.

This is called pattern scaling. It assumes that local change grows in step with global warming. Two limits follow.

  • Models disagree on the sign of rainfall change in much of the tropics, and a median damps the change wherever they disagree. The rainfall pattern is the weakest part of this step.
  • The median model run reaches about 5.2 °C above 1850–1900 by the end of the century. Above that the pattern is extended in a straight line. So the food and water layer and the climate-niche layer above about 5.2 °C lie beyond what the models simulated.

5The heat layer

The limit

Wet-bulb temperature combines heat and humidity in one number. It is the lowest temperature a wet surface can reach by evaporation. When it is high, sweat no longer evaporates fast enough to cool the body.

For years the limit was taken from theory to be a wet-bulb temperature of 35 °C (Sherwood and Huber 2010 [44]). Measurements on people put it lower. Vecellio and colleagues (2022) [52] tested 24 healthy adults aged 18 to 34 at light activity. They found that the body can no longer hold a steady core temperature at about 30.6 °C wet-bulb in humid heat, and at 25.8 to 27.8 °C in hot dry air. No subject reached 35 °C. Heat beyond this point is called uncompensable: core temperature goes on rising for as long as the exposure lasts.

The data

We use the grids published with Vecellio, Kong, Kenney and Huber (2023) [51]. For each cell of a 0.25° grid they give the hours per year above the uncompensable limit, at global warming levels from 1 to 4 °C in steps of half a degree. The hours are the mean of 16 climate-model runs under very high emissions (SSP5-8.5). They are counted in three-hourly model output, with each model’s change added to the ERA5 weather of 1950–1976, a reconstruction of observed weather, to remove model bias. The limit is 30.6 °C wet-bulb when the air is at or below 40 °C, and falls in hotter, drier air, following the measurements.

The rule: a month, and a week

A cell gets the main heat colour when it has at least 240 hours a year above the limit: a month of eight-hour days. At that point everyone, not only the frail, needs powered cooling for a month of days every year. A lighter band marks cells with at least 56 hours a year: a week of eight-hour days. That is dangerous every year for older adults, whose measured limit is about 2.4 to 3 °C lower (Wolf and colleagues 2023 [54]).

The month and the week are our choices, taken from the bands that Vecellio and colleagues report. No study of deaths fixes a duration. Counting any single hour would claim too much: at 4 °C, 30.8% of land has at least one such hour in an average year (our count on the published grids). Using the old 35 °C limit would claim too little, by about two orders of magnitude at 4 °C.

From the published grids to the map

  1. The published files store counts of three-hour steps. We multiply them by three to get hours.
  2. Between the published half-degree levels we interpolate each cell in a straight line in the logarithm of hours plus one.
  3. We bring the result to the map’s own grid, with cells about 10 km across, and record for each cell the warming level at which its hours first reach 56 and 240. A cell that has met a rule stays marked at higher levels.

One repair to the published data

We changed the published grids in one place. The two ends of their longitude axis meet at the Greenwich meridian, and in eleven columns on either side of that seam, from 1.75° W to 0.75° E, the published mean lacks some or all of its 16 model runs at every warming level. The strip is about 300 km wide in the Sahel. Left alone, it draws a dark line through Mali, Burkina Faso and Ghana and a straight north–south edge at 1° E, where the solid red of the Sahel stops dead at 4 °C.

In the columns where most model runs are missing, just west of the meridian, we interpolate between the columns on either side and never go below the published value. In the other columns we keep the published pattern and scale it up, row by row, by the factor that closes the step to the first complete column beside it. We chose this repair by a blind test: we imposed the same damage on undamaged places and scored each candidate repair against the true values. The cause of the damage is inferred from its pattern, not confirmed, so every value in the strip is less certain than those around it (section 10). The check is in the list under the table below.

Where the heat falls

Heat alone shades little land. It falls where people live. At 4 °C, Vecellio and colleagues count about 1.5 billion people with a month or more, and 2.7 billion with a week or more, in a fixed mid-century population of 9.19 billion [51].

At 2 °C the month-long band is limited to small parts of the Indus valley, the shore of the Persian Gulf and the Red Sea coast. At 3 °C it covers parts of the Indus–Ganges plain and the Gulf and Red Sea coasts, with eastern China and the Sahel at the one-week level. At 4 °C all of these reach the one-month level, with the coast of the Bay of Bengal and parts of mainland South-East Asia. Europe and the eastern United States stay below both levels at 4 °C, and the Amazon and Congo basins almost entirely so.

Above 4 °C: our own extrapolation

Above 4 °C there is nothing to interpolate: we found no published gridded study of uncompensable heat above 4 °C. So we extend the published maps ourselves. We measured how fast the hours grew between 3 and 4 °C: about fourfold per degree where there were a few dozen hours a year, and less than twofold where there were already a thousand. Above 4 °C each cell’s hours are multiplied, in steps of 0.1 °C, by the factor that fits its current hours, and capped at the number of hours in a year. A cell with no hours at all at 4 °C stays at none.

We tested the rule by hiding published maps and predicting them from cooler ones. The table below gives the result: the rule finds most of the real zone, and it draws the month-long zone too large every time.

Everything on the heat layer above 4 °C is therefore our extrapolation, not a published result, and it probably errs on the large side. The globe says so when the slider is above 4 °C.

How well it matches published work

Up to 4 °C the layer should reproduce the published grids, and it does so closely. The misses are in the smallest zones and on the coast. Above 4 °C nothing published exists to test against.

The worst match is the count of people at 1.5 °C, the lowest setting of the slider. The map has about 242 million people with a week or more of such heat each year, against 140 million in the paper, and about 2.1 million with a month or more, against 30 million. The zones are very small at that level, and the two counts use different populations. Read the heat figures at the low end of the slider as rough.

The heat layer, checked against published figures. “Our map” is the map now on the front page.

What we comparedPublished figureOur mapVerdict
Share of land with a month or more of such heat each year, at 2, 3 and 4 °C [51]0.07%, 0.34% and 2.06% in the published grids, by our count0.05%, 0.30% and 2.07%CloseA little smaller at every level, because the published grid is redrawn on a finer map. The gap is widest at 2 °C, where the zone is a few scattered cells.
Share of land with a week or more each year, at 2, 3 and 4 °C [51]0.39%, 2.8% and 6.3% in the published grids, by our count0.40%, 2.78% and 6.27%CloseThe two agree to within a tenth of a percentage point.
People with a month or more each year at 4 °C [51]1.50 billion of a fixed mid-century population of 9.19 billion (16.3%)1.41 billion of the 7.84 billion people alive in 2020 (18.0%)MatchThe same order of size. The two counts use different populations, so they cannot agree exactly.
People with a week or more, and with a month or more, each year at 1.5 °C [51]About 140 million and 30 millionAbout 242 million and 2.1 millionMissThe worst match of the layer. Our count with a week or more is well above the paper’s, and our count with a month or more is more than ten times lower. We apply the threshold to the average of the climate models. The paper counts people model by model, which finds more of them where the hot zone is very small.
Red Sea coast: share of its land with a week or more, and with a month or more, at 4 °C [51]22.3% and 14.7%, by our count on the published grids18.8% and 11.1%MissThe hottest cells there are shoreline cells that are less than half land, and our land counts leave those out. Counting every cell with any land gives 22.9% and 15.0%. The Persian Gulf coast is low for the same reason.
Places the published maps leave without month-long heat up to 4 °C [51]Europe, the United States, the Amazon, the Congo basin and IndonesiaNo month-long heat in any of the five. The one-week band covers under 1% of eachMatchThe map does not shade places the published work leaves clear.
Test of our extrapolation: hide the published 4 °C map, predict it from the 3 °C map, and compare the month-long zone2.07% of land in the real 4 °C map2.40% of land predictedCloseOn the large side. The prediction found 92% of the real zone, and 21% of what it shaded was wrong.
The same test run six ways, each reaching a degree or more beyond the dataA sound rule would err in both directionsThe month-long zone came out too large in all six, by 15 to 67%MissAbove 4 °C the solid red area is probably too large by a similar margin. We did not correct for it. The one-week band showed no such bias.
The heat layer at 5 and 6 °CNothing published to compare withA month or more on 4.7% and 10.2% of landNot testedThis range is our own extrapolation. No published map exists to test it against.
All 23 checks on the heat layer, in the data build’s own words11 match, 7 close, 3 miss, 2 not tested

For specialists. The wording is technical. “Research note” means our working notes from the published literature, “4k grid” the map’s own grid of 4,096 by 2,048 cells, and “domain” the land cells north of 60° S.

  • MatchDownload integrity: Zenodo record 8347178 zip

    Expected
    md5 d58c1cdcbf847f0e141d0ed680273630 (published by Zenodo), 5,798,517 bytes
    Result
    md5 d58c1cdcbf847f0e141d0ed680273630, 5,798,517 bytes; open download, no access problem
  • MatchUnits and orientation: city hours at 4 °C (stored value x 3, lat north-up, lon 0..360 rolled to -180..180)

    Expected
    Lahore about 1,013 h, Dammam about 805 h, Chicago about 32 h
    Result
    Lahore 1,013.5 h, Dammam 804.7 h, Chicago 31.9 h (node 41.75 N, 87.5 W). Twelve cities at four levels reproduce the paper’s Table 2 to its rounding. Without the longitude roll, or flipped north-south, Lahore reads 0 h.
  • MatchLand share with ≥ 56 h at 1.5 / 2 / 2.5 / 3 / 3.5 / 4 °C, published cells counted as they are

    Expected
    0.25 / 0.39 / 1.6 / 2.8 / 4.3 / 6.3%
    Result
    0.230 / 0.369 / 1.619 / 2.784 / 4.279 / 6.301%
  • CloseLand share with ≥ 56 h at 1.5 / 2 / 2.5 / 3 / 3.5 / 4 °C, the layer (threshold field on the 4k grid)

    Expected
    0.25 / 0.39 / 1.6 / 2.8 / 4.3 / 6.3%
    Result
    0.220 / 0.398 / 1.633 / 2.784 / 4.260 / 6.267%. Largest difference 0.040 points. Bilinear resampling shrinks the area by about 2% at 4 °C; the repair of the seam adds 0.09 points.
  • MatchLand share with ≥ 240 h at 1.5 / 2 / 2.5 / 3 / 3.5 / 4 °C, published cells counted as they are

    Expected
    0.01 / 0.07 / 0.18 / 0.34 / 0.85 / 2.06%
    Result
    0.008 / 0.059 / 0.162 / 0.320 / 0.826 / 2.055%
  • CloseLand share with ≥ 240 h at 1.5 / 2 / 2.5 / 3 / 3.5 / 4 °C, the layer

    Expected
    0.01 / 0.07 / 0.18 / 0.34 / 0.85 / 2.06%
    Result
    0.004 / 0.048 / 0.152 / 0.304 / 0.799 / 2.066%. Within 0.01 points at 4 °C, where two things cancel: the repair of the seam adds 0.06 points and bilinear resampling takes 0.05 away. About a third low at 2 °C, where the footprint is a few isolated cells.
  • MatchPeople today at 4 °C (GHS-POP 2020, any land)

    Expected
    Same order of magnitude as the paper’s 2.71 bn (56 h) and 1.50 bn (240 h) of a 9.19 bn 2050 population
    Result
    2.557 bn (56 h) and 1.411 bn (240 h); 32.6% and 18.0% of today’s population against the paper’s 29.5% and 16.3%. On domain cells only: 2.526 and 1.398 bn.
  • ClosePeople today at 2 to 3.5 °C

    Expected
    Order of magnitude of the paper: 56 h 0.39 / 0.90 / 1.57 / 2.13 bn; 240 h 0.04 / 0.16 / 0.37 / 0.76 bn at 2 / 2.5 / 3 / 3.5 °C
    Result
    56 h 0.340 / 1.103 / 1.565 / 1.966 bn; 240 h 0.021 / 0.119 / 0.220 / 0.705 bn
  • MissPeople today at 1.5 °C

    Expected
    Order of magnitude of the paper: 0.14 bn (56 h), 0.03 bn (240 h)
    Result
    0.242 bn (56 h), 0.002 bn (240 h). The 240 h count is more than ten times below the paper’s. A threshold on ensemble-mean hours gives fewer people in a narrow hot tail than per-model counts.
  • CloseRegional boxes, 11 boxes x 4 levels (1.5, 2, 3, 4 °C), share of each region’s land

    Expected
    Table in section 4.1, e.g. at 4 °C Indus-Ganges 59.8 / 41.2, Sahel 55.5 / 16.0, Eastern China 61.7 / 27.2
    Result
    At 4 °C Indus-Ganges 58.7 / 40.5, Sahel 56.7 / 17.6, Eastern China 61.5 / 27.3. Mean absolute difference over all pairs 0.79 points (56 h) and 0.44 points (240 h). The Sahel and the Guinea coast (16.5 / 0.1 against 13.3 / 0) are above the note because its figures were computed with the seam defect in place.
  • MissRegional box: Red Sea coast and Tihama at 3 and 4 °C

    Expected
    16.1 / 9.3 at 3 °C; 22.3 / 14.7 at 4 °C
    Result
    12.5 / 5.0 at 3 °C; 18.8 / 11.1 at 4 °C. Cause: the hottest cells are shoreline cells under half land and so outside domain_4k. Counting on any land gives 22.9 / 15.0 at 4 °C. Persian Gulf (up to 2.5 points low) and Sonoran (1.5 low) behave the same way.
  • MatchShould not light up through 4 °C: Europe, United States, Amazon, Congo basin, Indonesia

    Expected
    0 at the 240 h class; under 1% at 56 h
    Result
    240 h: 0 in all five at 2, 3 and 4 °C. 56 h at 4 °C: Europe 0.00%, United States 0.01%, Indonesia 0.33%, Amazon box 0.34%, Congo box 0.51%. East African highlands 0.02%.
  • MatchHold-out 3.0 -> 4.0 °C, IoU of the ≥ 240 h footprint

    Expected
    At least 0.5 for one method, otherwise label as low-confidence
    Result
    Damped factors (used) 0.742; probit-linear 0.694; the research note’s recipe as written 0.713. For 56 h: 0.861, 0.838, 0.693.
  • CloseHold-out 3.0 -> 4.0 °C, predicted land share (method used)

    Expected
    Real 4 °C grid on the 4k domain: 6.27% (56 h), 2.07% (240 h)
    Result
    6.21% (-0.8%) and 2.40% (+16.3%). 92% of the real 240 h area found; 21% of the predicted area wrong. Regional boxes: mean absolute error 1.1 / 1.3 points, largest 4.1 / 9.9 (Eastern China 240 h 37.1 against 27.3).
  • CloseHold-out 3.5 -> 4.0 °C, predicted land share (method used)

    Expected
    6.27% (56 h), 2.07% (240 h)
    Result
    6.05% (-3.5%), IoU 0.870; 2.04% (-1.3%), IoU 0.831. Probit-linear is better on this test: IoU 0.884 and 0.860.
  • CloseHold-out 2.0 -> 4.0 °C (added by the heat builder; same reach as the slider’s 6 °C beyond the data)

    Expected
    6.27% (56 h), 2.07% (240 h)
    Result
    Method used: 6.34% (+1.1%), IoU 0.780; 2.50% (+20.9%), IoU 0.667. Probit-linear: 7.41% and 2.95%, IoU 0.739 and 0.607.
  • MissDirection of error in the six hold-out tests that reach 1 °C or more beyond the data (240 h area, method used)

    Expected
    no systematic bias
    Result
    Too large in all six, by 15 to 67%. The 56 h area showed no such bias (0.90 to 1.08 times the real area).
  • Not testedLand share and people today at 5 °C (extrapolation)

    Expected
    No published target
    Result
    56 h: 14.68% of land, 3.50 bn people. 240 h: 4.65%, 2.19 bn. Other methods give 14.4 to 19.7% and 4.6 to 5.7%.
  • Not testedLand share and people today at 6 °C (extrapolation)

    Expected
    No published target
    Result
    56 h: 24.91% of land, 4.18 bn people. 240 h: 10.22%, 3.09 bn. Other methods give 23.2 to 26.6% and 10.0 to 15.7%.
  • MatchSeam of the published grids at the Greenwich meridian (eleven columns from 1.75 W to 0.75 E lack some or all of the 16 model runs)

    Expected
    No straight north-south edge in the layer; a repair that does better than leaving the columns alone when the same damage is imposed on sound places
    Result
    Rows between 8 and 22 N in which the 240 h level differs by 0.25 °C or more between cells a quarter of a degree apart: at most 27 of 159 inside 2 W to 1 E (0.25 E to 0.5 E, all between 8 and 12 N, where the published field itself is uneven), against 40 for the worst pair elsewhere between 12 W and 20 E. With the first repair, which touched only 1.0 to 0.25 W, the pair 0.75 E to 1.0 E had 75. Blind test, West and central Africa: error of log(hours + 1) 0.110 for the repair used, 0.228 for the first repair, 0.353 for plain interpolation of all eleven columns; mismatch of the 240 h footprint 0.060, 0.654 and 0.173. The repair moves the 240 h share at 4 °C from 2.01 to 2.07% of land and the Sahel box from 15.9 to 17.6%.
  • MatchResampling self-check: heat.py’s bilinear against common.resample_to_4k (GDAL) at 4.0 °C

    Expected
    No difference
    Result
    Largest difference in log(hours + 1) on the domain 0.00e+00
  • MatchThreshold field conventions

    Expected
    float32 (2048, 4096); values in 1.0 .. 7.0 or +inf; NaN exactly where there is no land; heat56 <= heat240
    Result
    Both fields float32 (2048, 4096); finite range 1.000 .. 7.000; NaN exactly where land_frac_4k == 0; heat56 <= heat240 everywhere. heat240: 314,372 finite, 2,527,132 +inf. heat56: 674,828 finite, 2,166,676 +inf.
  • MatchDeterminism and run from scratch

    Expected
    Identical outputs on repeat runs
    Result
    Two runs of the final script, one from an empty raw/vecellio2023 folder including the download, gave identical SHA-256 for all three files and identical logs.

6The food and water layer

This layer asks whether the land could feed and water the people on it from local rain. It uses three rules. A cell fails if any one of them is met.

  1. Too dry. The aridity index is below 0.20. The index is a year’s rainfall divided by a year’s potential evapotranspiration, which is the water the air would take up from a well-watered surface. Below 0.20 covers the “arid” and “hyper-arid” classes of the United Nations Environment Programme’s scheme [19] [57]. This rule alone does not flag semi-arid land, between 0.20 and 0.50, which supports rain-fed millet and sorghum. The second rule catches some of it.
  2. No moist season. No three-month window in the year has rainfall of at least half of potential evapotranspiration. Half of evaporative demand is the mark the Food and Agriculture Organization uses for the start of a growing period [7]. Our test looks at moisture only, so a frozen winter window passes it.
  3. Growing season too hot. Every three-month window that is wet enough has an average daily high of 36 °C or more. Above about 35 °C maize pollen fails, rice flowers become sterile and wheat stops filling grain (Luo 2011 [30]; Hatfield and Prueger 2015 [16]).

Cold is not tested.

How the layer changes with warming

Local temperature and rainfall are changed with the CMIP6 patterns described in section 4. Potential evapotranspiration in the baseline climate is computed month by month with the FAO-56 Penman–Monteith formula, from WorldClim’s temperature, vapour pressure, sunshine and wind. WorldClim does not state the height of its wind data; we took it as wind at 10 m. Under warming, potential evapotranspiration is raised by about 2.5% for each degree of local warming, which is the temperature term of the Hargreaves formula. That is a moderate choice, and one reason our map dries less than published studies. The three rules are then applied again at every 0.1 °C.

Deserts and dry steppe today

By 1.5 °C, 25% of the land outside Antarctica fails these rules, almost all of it already in the 1970–2000 climate, and 556 million people live on it today, on rivers, wells and food brought in from elsewhere. That is more land than the desert class of the Köppen climate scheme [2], which covers 19.1% (our calculation from WorldClim 2.1). The moist-season rule takes no account of water stored in the soil, so it also catches semi-arid steppe where dryland farming and grazing exist, such as the High Plains of the United States: about 12% of the class is such steppe, not desert. That is why the legend says “deserts and dry steppe”.

A map that shaded all dry land would show the Sahara at the lowest setting and make it look like an effect of warming. So land that fails at or below 1.5 °C, the lowest setting of the slider, is drawn in a muted tone, labelled “Deserts and dry steppe today”, and left out of the counts. Only land that first fails at a higher level gets the strong colour.

River-fed drylands

Arid land along the Nile, the Indus, the Tigris and Euphrates and the Colorado is inhabited because water arrives from somewhere else. We do not model rivers, aquifers or glaciers cell by cell. Land within 25 km of a major river [33] is treated as depending on water from upstream and is not counted as failed, unless even its coolest season is too hot for grain. Smaller rivers and canal systems are not exempt. That is generous to the places beside large rivers, because the flow is assumed to continue at every level of warming. Immerzeel and colleagues (2020) [20] count 1.9 billion people living in or directly downstream of mountain “water towers”, and find that the most important of them are among the most vulnerable.

Where this layer is most open to challenge

The aridity index is disputed as a measure of future drying. In published studies it projects that drylands expand. Measures built on soil moisture, runoff and vegetation project little or no expansion globally, partly because plants use less water when there is more CO2 in the air (Berg and McColl 2021 [3]). Drying in the Mediterranean, southern Africa and subtropical South America appears on both kinds of measure.

Published estimates differ widely. Spinoni and colleagues (2021) [46] find about 5.2 million km² of land shifting to an arid climate at 3 °C. Huang and colleagues (2016) [19] project drylands of all kinds covering 56% of land by 2100 under high emissions. Feng and Fu (2013) [8] find drylands 5.8 million km² larger by 2100 under high emissions than in 1961–1990, and Koutroulis (2019) [25] finds 11.2% of land moving to a drier class at 4 °C.

Above 4 °C the layer is severe about heat instead. Most of its new land then comes from the third rule, the 36 °C limit of today’s maize, rice, wheat and sorghum, and by 6 °C that rule reaches much of the Amazon basin. No new crop varieties, shifted seasons or irrigation are assumed.

How well it matches published work

We aimed to land between two published estimates of drying. We did not: the map dries less than both. Three reasons are known, and we adjusted none of them.

  • Raising potential evapotranspiration by 2.5% per degree is a moderate choice.
  • The median of 22 models damps rainfall change wherever the models disagree on its direction.
  • The 0.20 limit is far from most semi-arid land. Madrid has an aridity index of 0.35 and is never marked.

The food and water layer, checked against published figures. “Our map” is the map now on the front page.

What we comparedPublished figureOur mapVerdict
How much faster land warms than the globe, in the pattern we useAbout 1.3 to 1.6 °C for each degree of global warming1.38 °C for each degreeMatchInside the expected range.
Does the pattern hold at a lower level of warming? We predict the models’ own mid-century climate from itA close fit: a correlation near 1Correlation 0.98 for temperature and 0.93 for rainfallMatchThe pattern holds. Rainfall is the weaker of the two.
Arid and hyper-arid land in the 1970–2000 climateRoughly a fifth of the land outside Antarctica23.1%CloseOn the dry side of the published figure.
Land that becomes arid at 2, 3 and 4 °C [46]2.6, 5.2 and 6.8 million km²0.59, 1.08 and 1.55 million km²MissOur map finds 4 to 6 times less newly arid land.
Growth of all drylands by 2100 under high emissions [8]5.8 million km² more3.0 million km² more at 4 °CMissAbout 59% of the published growth.
Share of land that is dryland by 2100 [19]50% under medium emissions and 56% under high emissions, up from about 45%48.6% at 2.7 °C and 49.7% at 4.5 °C, up from 47.2%MissOur drylands start larger and grow by a quarter to a third as much. Matching the two emissions paths to 2.7 and 4.5 °C is our assumption.
Regions where published studies agree on dryingThe Mediterranean, south-west North America, southern Africa, Australia and subtropical South AmericaAll five gain failed land by 4 °C: Australia the most, the Mediterranean and subtropical South America the leastCloseThe right regions dry, but the drying in the Mediterranean and subtropical South America is weak on our map.
New food and water failure at 5 and 6 °C, most of it from the crop-heat ruleNothing published to compare with4.0% and 8.8% of landNot testedNo published map applies a crop-heat limit in this way, so this part of the layer is untested.
All 19 checks on the food and water layer, in the data build’s own words4 match, 4 close, 8 miss, 3 not tested

For specialists. The wording is technical. “Research note” means our working notes from the published literature, “4k grid” the map’s own grid of 4,096 by 2,048 cells, and “domain” the land cells north of 60° S.

  • MatchLand-mean warming per degree of global warming (median pattern, land north of 60 S)

    Expected
    roughly 1.3 to 1.6
    Result
    1.378 (daily maximum 1.343, daily minimum 1.413); single models 1.263 to 1.549; tropics 1.14, north of 50 N 1.83; 1.377 on the 4k domain
  • MatchLinearity: predict ssp245 2041-2060 fields from the ssp585 pattern

    Expected
    high pattern correlation, small bias
    Result
    Each model’s own pattern: temperature r median 0.961 (0.889 to 0.979), bias -0.03 °C; precipitation log ratio r median 0.870 (0.716 to 0.930), bias -1.2%. Published median pattern at 1.76 °C against the multi-model median field: r 0.976 and 0.930, bias -0.01 °C and -1.2%. BCC-CSM2-MR is an outlier at -16.7% precipitation bias.
  • CloseBaseline aridity index below 0.20 (hyper-arid plus arid)

    Expected
    roughly a fifth of land outside Antarctica
    Result
    23.13% of the 4k domain (31.10 million km²); 22.53% on the 10-minute grid; 23.67% if wind is taken as a 2 m wind
  • CloseBaseline drylands (index below 0.65)

    Expected
    roughly 41 to 47%
    Result
    47.16% of the 4k domain (63.42 million km²); 46.12% on the 10-minute grid
  • Not testedLand flagged “desert today” (threshold at or below 1.5 °C)

    Expected
    no like-for-like target; research note gives Koppen desert 19.1% of land
    Result
    34.22 million km², 25.45% of land, 541 million people today (domain cells). 76% of it meets tests (a) and (b) together, 12% (a) only, 10% (b) only, 1% (c). With test (b) left out it would be 30.93 million km², 23.00% of land, 488 million people.
  • MissWhat the “desert today” class holds: land that is not arid (aridity index of the 1970-2000 climate 0.20 or more)

    Expected
    none. The research note keeps the index limit at 0.20 because “semi-arid land supports rain-fed millet and sorghum, so flagging it would over-claim”
    Result
    4.14 million km² of semi-arid land (index 0.20 to 0.50): 12.1% of the class, 3.08% of all land, 85 million people on domain cells. 3.38 million km² of it is marked by test (b) alone, 0.45 through test (c), 0.32 by test (a) between 0.56 and 1.5 °C. Largest: United States 1,000 thousand km², Australia 870, China 469, Mongolia 355, Argentina 280, Mexico 252, South Africa 212. In the contiguous United States the class covers 21.5% of the land (1.71 million km², 24.0 million people) and reaches east to 98.8 W between 33 and 39 N; Lubbock, Amarillo, Denver and Goodland in Kansas are in it. The rule is the research note’s and was not changed; the class is not only desert and should not be labelled as if it were.
  • MissTest (b) as a growing-season test (FAO: the growing period is the time when temperature and water both permit crop growth)

    Expected
    a moist window that counts is one in which a crop can grow
    Result
    Test (b) looks at moisture only, as the research note states it. In the 1970-2000 climate 2.09 million km² of semi-arid land (10 million people) has a moist window only where the window’s mean temperature is below 0 °C, and 3.64 million km² (27 million people) only below 5 °C: Kazakhstan 847 thousand km², United States 448, Canada 344 at the 0 °C limit. Such land is not marked, while warmer steppe with more rain is: Goodland in Kansas (456 mm a year, index 0.319, best window April to June, ratio 0.42) is desert today; Havre in Montana (324 mm, index 0.309, best window December to February at -7.6 °C, ratio 0.87) is not. Adding a temperature limit to the window would enlarge the class, not shrink it: 28.81% of land with a 0 °C limit and 31.71% with 5 °C, against 25.45%, and most of the added land is ice and tundra (Greenland 1.8 to 2.0 million km²) and the wheat belts of Kazakhstan and the prairies. The rule was left as the research note states it; the inconsistency is reported.
  • MissLand newly flagged at 2 / 3 / 4 / 5 / 6 °C (all tests, above the desert-today class)

    Expected
    between Spinoni and Huang
    Result
    0.40 / 1.29 / 2.53 / 5.35 / 11.79 million km² = 0.30 / 0.96 / 1.88 / 3.98 / 8.76% of land; 14 / 54 / 121 / 229 / 503 million people today (domain cells). Below both Spinoni and Huang up to 4 °C.
  • MissLand newly arid since 1981-2010 (Spinoni 2021) at 1.5 / 2 / 3 / 4 °C

    Expected
    2.0 / 2.6 / 5.2 / 6.8 million km²
    Result
    Index below 0.20 at the level but not at 0.77 °C: 0.35 / 0.59 / 1.08 / 1.55 million km² (net 0.27 / 0.45 / 0.83 / 1.16). 4 to 6 times lower.
  • MissArid plus hyper-arid share by 2100 (Huang 2016)

    Expected
    22.8% (RCP4.5) to 27.5% (RCP8.5)
    Result
    23.13% at baseline, 23.72% at 2.7 °C, 24.16% at 4.5 °C. The level is inside the range only because the baseline is already 23.1%; the increase is 0.6 to 1.0 points. Mapping RCP4.5 to 2.7 °C and RCP8.5 to 4.5 °C is the builder’s assumption.
  • MissDrylands by 2100 (Huang 2016)

    Expected
    50% (RCP4.5) and 56% (RCP8.5), from about 45%
    Result
    47.16% at baseline, 48.60% at 2.7 °C, 49.68% at 4.5 °C, 50.43% at 6 °C. Growth about a quarter to a third of Huang’s.
  • MissDryland growth by 2100 under high emissions (Feng and Fu 2013; Koutroulis 2019)

    Expected
    +5.8 million km²; +7% of land
    Result
    +3.01 million km² at 4 °C, +3.40 at 4.5 °C (+2.5% of land): 59% of Feng and Fu, about a third of Koutroulis
  • MissLand moving to a drier / wetter aridity class at 4 °C (Koutroulis 2019)

    Expected
    11.2% / 4.24%
    Result
    6.04% / 1.40% against the 0.77 °C climate: about half and a third
  • CloseRegions that dry

    Expected
    Mediterranean, south-west North America, southern Africa, Australia, subtropical South America
    Result
    New land by 4 °C in million km²: Australia 0.85, southern Africa 0.25, south-west North America 0.22, subtropical South America 0.12, Mediterranean 0.10. The five hold 70% of new land at 3 °C and 61% at 4 °C. Also Middle East and Central Asia 0.30 and Sahel 0.26. Mediterranean and subtropical South America are weak.
  • Not testedSplit of newly flagged land among tests (a) aridity index, (b) no moist season, (c) crop heat

    Expected
    no published target
    Result
    By 4 °C (2.53 million km²): (a) only 24%, (b) only 16%, (c) 54%, river heat test 5%. By 6 °C (11.79 million km²): (a) only 9%, (b) only 5%, (c) 76%, river heat test 9%; 3.06 million km² of the (c) land is in Brazil.
  • CloseCoolest-quarter crop-heat screen against the research note’s illustration

    Expected
    0.1 / 0.9 / 3.2 / 10.6 / 19.5% of land at 2 / 3 / 4 / 5 / 6 °C, assuming all land warms 1.4 °C per degree
    Result
    With that assumption 0.1 / 0.8 / 3.0 / 10.1 / 19.1%. With the CMIP6 pattern used for the map: 0.02 / 0.30 / 1.00 / 2.81 / 7.55%, because the daily maximum of the coolest quarter in the hottest lands warms only 1.06 °C per degree.
  • Not testedRiver-fed drylands (exempt from the aridity tests)

    Expected
    Nile, Indus, Tigris-Euphrates, Colorado and similar corridors left unflagged
    Result
    Corridor (25 km around Natural Earth rivers of scalerank 5 or lower) 9.31 million km²; river-fed dryland today 1.175 million km² with 192 million people (Egypt 82.1 million, Pakistan 30.6, Iraq 27.1, Sudan 15.2). Rivers of scalerank 6 and above and canal systems are not exempt.
  • MatchFloat fields against encoded bytes

    Expected
    met_at on the float field equals met_at_byte on the bytes at all 46 slider levels
    Result
    0 disagreements for both fields (thresholds are kept one byte step inside their 0.1 °C step)
  • MatchDeterminism and run from scratch

    Expected
    Identical outputs on repeat runs
    Result
    Repeated runs from scratch gave identical SHA-256 for the output files and identical printed numbers

7The sea-level layer

Ice sheets and the deep ocean respond to warming over thousands of years. The sea level that a given warming eventually produces is called the committed rise. We use the committed rise about 10,000 years after the warming. That is the longest period for which the IPCC has assessed a range [13].

Committed global sea-level rise by warming level, about 10,000 years on. Metres above the 1995–2014 level. The globe draws the central value.
WarmingCentralLowHigh
1.5 °Cc6.5 m5 m10 m
2.0 °C10.5 m8 m13 m
2.5 °Ca14 m9 m19 m
3.0 °C18 m10 m24 m
3.5 °Ca22 m15 m30 m
4.0 °C26 m19 m33 m
4.5 °Ca29.5 m23 m35 m
5.0 °C33.5 m28 m37 m
5.5 °Cb37.5 m32.5 m40 m
6.0 °Cb41 m36 m51 m
  • Unmarked rows: range from IPCC AR6 Working Group I (2021), Chapter 9, Table 9.10 [13].
  • a Range read by us from the model data behind the IPCC’s Figure 9.30 [23].
  • b Our extension. The IPCC gives no assessed range above 5 °C.
  • c The IPCC range is 6 to 7 m. We widened it, for the reason given below.

Where the numbers come from

The IPCC’s table gives ranges of 6 to 7 m at 1.5 °C, 8 to 13 m at 2 °C, 10 to 24 m at 3 °C, 19 to 33 m at 4 °C and 28 to 37 m at 5 °C. They rest on two modelling studies (Clark and colleagues 2016 [5]; Van Breedam and colleagues 2020 [49]), and the IPCC labels them “medium agreement, limited evidence”. Our central line is the middle of those ranges, smoothed to a straight line of about 7.7 m for each degree.

We changed the IPCC ranges in two places.

  • At 1.5 °C we widened the band to 5 to 10 m. The IPCC’s 6 to 7 m comes from a single study. The IPCC also reports that seas stood 5 to 10 m higher about 125,000 years ago, when the world was 0.5 to 1.5 °C warmer than 1850–1900 [21].
  • Above 5 °C we extended the line to 6 °C using the same simulations the IPCC used [23], which include runs on both sides of 6 °C. Read the 6 °C figure as roughly 40 m, give or take 10 or more.

Three further points for specialists.

  • The values are not quite equilibrium. In the simulations the sea is still rising slowly at year 10,000, and in most of them the temperature has drifted down from its peak. In those models, a temperature held for ever would give more, not less.
  • Other models give less. Levermann and colleagues (2013) [28], using a different family of ice-sheet models, give about 12.5 m at 3 °C and 13.9 m at 4 °C with no time limit. Our central values are 18 m and 26 m.
  • Past warm periods sit near the central line. The IPCC reports seas very likely 5 to 25 m higher about three million years ago, when the world was 2.5 to 4 °C warmer [21].

This century is a different question

None of this is a forecast for this century. For 2100 the IPCC’s likely range under very high emissions is 0.63 to 1.01 m above the 1995–2014 level, and it says a rise approaching 2 m cannot be ruled out [13]. Over 2,000 years it gives 2 to 3 m at 1.5 °C and 19 to 22 m at 5 °C [21]. What arrives by 2100 is a few percent of what is eventually committed. The IPCC says it is uncertain how quickly the long-term committed sea level will be reached.

How flooding is mapped

  • Elevation comes from ETOPO 2022 [34] at 60 arc-seconds: cells about 2 km across, with land heights measured at the ground, not at treetops or rooftops.
  • The sea is raised by the same amount everywhere.
  • A cell floods only if the sea can reach it through neighbouring cells that are also below the new level. Land that is below sea level and dry today, such as Death Valley or the shore of the Dead Sea, stays dry unless the sea can get in.
  • Today’s coastline comes from Natural Earth [33] and not from the elevation data, so that land below sea level behind dykes, such as the Dutch polders, still counts as land today. Where the sea can reach such land, it counts as lost at every setting: dykes and pumps are not in the map.
  • The ice sheets of Greenland and Antarctica are drawn at their present size and are not flooded. Over these lengths of time much of that ice would be gone.
  • Three inland basins change with the slider. The Salton Trough in California joins the sea once it has risen about 11 m, at a slider setting of 2.1 °C. The Danakil depression in Ethiopia and Eritrea joins at about 15 m (2.7 °C). The lowland around the Caspian Sea joins at about 30 m (4.6 °C) in the elevation data, across a divide only a few metres high. Read any change there as indicative.

Global elevation data are wrong by a metre or more on low coasts. That matters a great deal for this century and much less here. Our levels run from 6.5 to 41 m, and the uncertainty in the level itself is far larger than the error in the elevation.

The share of land lost is small. The share of people is not: about one billion people live on land less than 10 m above today’s high-tide line (Kulp and Strauss 2019 [26]).

How well it matches published work

The height of the sea is taken from the IPCC, so the check here is of the flooding: whether the right land floods, and whether the people counted on low land agree with a published count.

The sea-level layer, checked against published figures. “Our map” is the map now on the front page.

What we comparedPublished figureOur mapVerdict
Flooded land with the sea 6.5, 18 and 41 m higher (the levels for 1.5, 3 and 6 °C)2.18, 4.72 and 9.22 million km². Not a published figure: an earlier calculation of our own, with a different definition of land2.14, 4.54 and 9.21 million km²CloseUp to 4% lower. The two calculations draw today’s coastline from different data.
People living less than 10 m above the sea today [26]About 1.0 billion975 millionCloseWithin a few percent.
People living less than 5 m above the sea today [26]About 770 million566 millionMiss26% lower. That study uses a different elevation model and population year, and measures height from the local high-tide line where we measure from mean sea level. Nothing was adjusted.
Low-lying places that must flood at the 3 °C sea levelThe Netherlands, the Ganges–Brahmaputra, Mekong, Nile and Po deltas, South Florida, and Shanghai with the lower Yangtze plainAll flood: between 84% and 100% of the land of eachMatchEvery one of them floods on the map.
Basins below sea level that are walled off from the ocean and must stay dryThe Dead Sea, the Qattara Depression, Death Valley, Lake Eyre and the Turpan DepressionNone floods at any setting of the sliderMatchThe sea reaches only land it can get to.
People who live today on land drowned at 1.5, 3 and 6 °CNo published figure for a commitment over 10,000 years712 million, 1.39 billion and 2.23 billionNot testedNothing to test against. The counts below 10 m and 5 m above are the nearest check.
All 15 checks on the sea-level layer, in the data build’s own words9 match, 2 close, 1 miss, 3 not tested

For specialists. The wording is technical. “Research note” means our working notes from the published literature, “4k grid” the map’s own grid of 4,096 by 2,048 cells, and “domain” the land cells north of 60° S.

  • CloseFlooded land at the ten central nodes against section C.5 of the research note

    Expected
    2.18 / 3.24 / 3.97 / 4.72 / 5.45 / 6.18 / 6.83 / 7.75 / 8.55 / 9.22 million km² at 6.5 / 10.5 / 14 / 18 / 22 / 26 / 29.5 / 33.5 / 37.5 / 41 m
    Result
    2.144 / 3.119 / 3.808 / 4.538 / 5.236 / 5.938 / 6.570 / 7.769 / 8.554 / 9.209 million km²: -1.6% at 6.5 m, -3.7 to -4.1% from 10.5 to 29.5 m, within 0.2% from 33.5 m. Cause: the pipeline counts Natural Earth land, the note counted ETOPO cells above 0 m.
  • MatchThe research note’s own definition recomputed with this script’s flood routine (ocean = cells at or below 0 m connected to the Pacific or the Black Sea; land = cells above 0 m)

    Expected
    the fifteen figures of section C.5 (2.18 ... 9.22; band 1.69, 3.13, 5.82, 7.66, 11.17)
    Result
    2.180 / 3.240 / 3.967 / 4.720 / 5.454 / 6.179 / 6.826 / 7.752 / 8.553 / 9.220; band 1.692 / 3.127 / 5.822 / 7.660 / 11.168. All fifteen within 0.1%.
  • MatchFlood routine against an independent method (scipy.ndimage.label of the cells at or below the level, or seed cells)

    Expected
    the same connected cells
    Result
    0 cells differ at 18 m and 41 m; 1 cell at 29.5 m, on the date line at 65.2 N, which the priority flood reaches by wrapping and label cannot
  • MatchGrid registration at 60 arc-seconds

    Expected
    best at zero shift
    Result
    98.948% of people on land cells at zero shift against 97.76 - 98.51% when moved by one cell; ETOPO above 0 m agrees with the land mask on 99.692% of cells against 99.573 - 99.652%
  • MatchPlaces that must flood at 3 °C (18 m): share of present-day land flooded

    Expected
    Netherlands, Ganges-Brahmaputra delta, Mekong delta, Nile delta, South Florida, Shanghai and the lower Yangtze plain, Po delta
    Result
    Netherlands 84% (62% at 1.5 °C; 14.8 million people), Ganges-Brahmaputra delta 100% (74%; 98.7 million), Mekong delta 99% (28.7 million), Nile delta 96% (47.1 million), South Florida 99% (7.3 million), Shanghai and lower Yangtze 97% (68.4 million), Po delta 100%. All seventeen named points flood at or below 3 °C (Dhaka at 1.81 °C, Mansoura at 1.93 °C, the others at or below 1.5 °C).
  • MatchPlaces that must not flood (Dead Sea, Qattara Depression, Death Valley, Lake Eyre, Turpan Depression)

    Expected
    dry at every level the site uses (central 41 m at 6 °C, 48 m at 7 °C; high curve 51 m at 6 °C)
    Result
    Dead Sea connects at 60.4 m, Qattara at 63.8 m, the other three at no level up to 80 m. None floods.
  • MatchBasin connection levels against the research note

    Expected
    Salton Trough 11 m, Danakil 15.5 m, Caspian 29.7 m, Dead Sea 61 m, Qattara 64 m, Aral Sea bed 66 m
    Result
    10.95 / 15.38 / 29.69 / 60.37 / 63.82 / 65.85 m. With 4-neighbour connectivity the Caspian connects at 31.11 m (note 31.1).
  • MatchBlack Sea

    Expected
    cut off at 0 m in the elevation grid alone, joins between 10 and 10.5 m; leaving its seed out loses 24 thousand km² at 6.5 m
    Result
    joins at 10.44 m; 24.3 thousand km² lost at 6.5 m and nothing at 10.5 m without the seed. In the pipeline the Black Sea is seeded from Natural Earth.
  • ClosePeople below 10 m (Kulp and Strauss 2019, as quoted in the research note)

    Expected
    about 1.0 billion
    Result
    975 million on cells connected at or below 10 m (1,001 million on all cells at or below 10 m)
  • MissPeople below 5 m (Kulp and Strauss 2019)

    Expected
    about 770 million
    Result
    566 million connected (587 million on all cells): 26% lower. Their study uses CoastalDEM, 2010 population and heights above the local high-tide line. Nothing was adjusted.
  • MatchLand now above 0 m that floods through the Caspian sill at 41 m

    Expected
    221 thousand km² (research note)
    Result
    223 thousand km²
  • Not testedRegional areas at 41 m in the research note (Amazon basin, West Siberia and others)

    Expected
    research note section C.5
    Result
    not reproduced: the note gives no boundaries for these regions
  • Not testedPeople on flooded land at 1.5 / 2 / 3 / 4 / 5 / 6 °C, central curve

    Expected
    no published target for a 10,000-year commitment
    Result
    712 / 1,007 / 1,392 / 1,722 / 1,997 / 2,228 million (9.1 to 28.4% of today’s people). Counting land-mask cells only: 657 / 944 / 1,322 / 1,647 / 1,919 / 2,150 million. Low to high curve at 3 °C: 975 to 1,644 million.
  • Not testedSensitivity of flooded land (4-neighbour connectivity; glaciers impassable; Caspian left out; small seed pieces not seeded)

    Expected
    small against the uncertainty of the sea-level curve
    Result
    4-neighbour -1.0 to -1.7%; glaciers impassable -0.01%; Caspian basin left out -6.2% at 5 °C and -5.6% at 6 °C; small seed pieces at most -7.6 thousand km². Low and high curves at 3 °C: 3.01 and 5.59 million km² around 4.54.
  • MatchDeterminism

    Expected
    Identical outputs on repeat runs
    Result
    Four runs gave identical SHA-256 hashes for all four output files

8The climate-niche layer

The globe has an optional dotted layer, switched off at first, for land with a mean annual temperature of 29 °C or more. Xu and colleagues (2020) [56] found that people have clustered in places with a mean annual temperature of about 11 to 15 °C, with a second cluster at about 20 to 25 °C. A mean of 29 °C or more is found on 0.8% of land today, mostly in the Sahara. Lenton and colleagues (2023) [27] estimate that at 2.7 °C of warming about 22% of a world of 9.5 billion people would live in such heat.

The rule is simple: the mean annual temperature of the 1970–2000 climate, warmed with the same pattern as the food and water layer (section 4), is 29 °C or more. The threshold is published. The scaling is ours.

This describes where people have lived, which reflects history and economics as well as climate. It is not a survival limit. It marks the Amazon by 2.7 °C. The measured heat limit in section 5 does not, even at 4 °C. So we label the layer “hotter than almost anywhere people live today”, and we never count it as uninhabitable.

How well it matches published work

The layer matches the published area in today’s climate and falls short of it at higher warming. Most of the gap is a difference in labelling, set out in the last two rows.

The climate-niche layer, checked against published figures. “Our map” is the map now on the front page.

What we comparedPublished figureOur mapVerdict
Land with a mean annual temperature of 29 °C or more in the 1970–2000 climate [56]0.8% of land0.97% of landMatchSlightly more than published.
Area of such land at 1.5 °C [27]6 million km²5.1 million km²CloseSomewhat less than published.
Area of such land at 2.7 °C [27]22 million km²15.3 million km²MissAbout 30% less. Our map reaches the published area at 3.2 °C.
Share of people living in such heat at 1.5, 2.7 and 4.4 °C [27]5%, 22% and 45 ± 7% of 9.5 billion people in 20702.7%, 11.5% and 33.6% of the people alive in 2020MissAbout half the published share at 1.5 and 2.7 °C. Two reasons are known. The paper counts a larger future population that grows most in hot countries. And it labels each emissions path with its expected warming, where we use the warming each model actually reached.
The same check without our scaling: the model files for the period the paper labels 4.4 °C, used directly [27]45 ± 7% of people38 to 39% of the people alive in 2020MatchInside the published range. In the 22 models we use, that period is 5.24 °C warmer than 1850–1900, not 4.4 °C.
All 6 checks on the climate-niche layer, in the data build’s own words2 match, 2 close, 2 miss

For specialists. The wording is technical. “Research note” means our working notes from the published literature, “4k grid” the map’s own grid of 4,096 by 2,048 cells, and “domain” the land cells north of 60° S.

  • MatchClimate niche, baseline climate (mean annual temperature ≥ 29 °C)

    Expected
    0.8 to 1.0% of land (Xu 2020: 0.8%; research note 0.97%)
    Result
    0.97% of land (1.311 million km²), 39.1 million people today
  • CloseClimate niche area at 1.5 °C (Lenton 2023)

    Expected
    6 million km²
    Result
    5.09 million km² (3.78% of land), 203 million people on domain cells
  • MissClimate niche area at 2.7 °C (Lenton 2023)

    Expected
    22 million km²
    Result
    15.31 million km² (11.39% of land). This map reaches 22 million km² at 3.23 °C.
  • MissClimate niche, share of people at 1.0 / 1.5 / 1.8 / 2.7 / 3.6 / 4.4 °C (Lenton 2023, of 9.5 bn in 2070)

    Expected
    0.9 ± 0.4 / 5 / nearly 10 / 22 / 34 ± 10 / 45 ± 7%
    Result
    1.2 / 2.7 / 3.7 / 11.5 / 23.5 / 33.6% of today’s 7.84 bn (coastal cells take the nearest domain cell). About half of Lenton’s at 1.5 and 2.7 °C, just below the range at 3.6 °C, below it at 4.4 °C.
  • MatchClimate niche from the ssp585 2081-2100 files directly, no pattern scaling (the period Lenton labels 4.4 °C)

    Expected
    45 ± 7% of people
    Result
    Multi-model median (mean) field: 40.14 (40.84) million km² and 38.1% (39.2%) of today’s people. In the 22 models that period is 5.24 °C above 1850-1900, not 4.4 °C; the pattern-scaled map at 5.24 °C gives 41.23 million km² and 39.5%.
  • CloseClimate niche, named places at 2.7 °C (Lenton 2023)

    Expected
    Burkina Faso and Mali near 100% of land; Brazil the largest area; India, Nigeria, Indonesia, Philippines, Pakistan, Sudan, Niger by population
    Result
    Burkina Faso 99%, Mali 100%; Brazil largest (1,583 thousand km²); people: India 324 million, Nigeria 66, Sudan 37, Pakistan 35, Indonesia 27, Philippines 26, Niger 24. Lenton’s counts are several times higher.

9Counting land and people

Shares of land are shares of the world’s land north of 60° S, which leaves out Antarctica and includes Greenland: 134.5 million km². Each cell is weighted by its true area. For heat, food and water a cell counts when at least half of it is land. Deserts and dry steppe that fail the test today are reported separately and are not part of the figure for land that newly fails. A piece of land that fails more than one test is counted once.

People are counted where they live today. We use GHS-POP R2023A [43][36], the 2020 population grid from the European Commission’s Joint Research Centre, which totals 7.84 billion people. The readout says “people who live there today” on purpose. We do not project population growth or migration, and the number is not a forecast of how many people will be affected. Published studies often use a fixed mid-century population of 9.19 to 9.5 billion instead, so their counts are not directly comparable with ours.

How well the foundations check out

Before any layer is drawn, the land, the people and the picture of the Earth have to sit in the right place. These are the checks on that.

The land, the people and the picture of the Earth, checked against published figures. “Our map” is the map now on the front page.

What we comparedPublished figureOur mapVerdict
Area of the land the map covers: all land north of 60° S135.8 million km². Not a published figure: the area in our research notes, measured on a coarser grid134.49 million km²Miss1.0% lower, and outside the range we had set. The difference is the coarser grid of the notes. Nothing was adjusted.
Total number of people once the population grid is brought to the map [43]7,840,952,947 in the source grid7,840,952,947MatchNobody is lost or counted twice.
Population and area of eight large countries, against national figures [33]Population within a few percent of the national estimates for 2019; area within 1%Populations 1 to 4% above the 2019 estimates; areas within 0.6%MatchCountries hold the people and the land they should.
Does the picture of the Earth line up with the coastline data? [45]Within one cell of the map, about 10 kmNo shift of a whole cell; the picture’s coast sits about 1.2 km northMatchThe shading sits on the land it belongs to.
All 5 checks on the land, the people and the picture of the Earth, in the data build’s own words4 match, 1 miss

For specialists. The wording is technical. “Research note” means our working notes from the published literature, “4k grid” the map’s own grid of 4,096 by 2,048 cells, and “domain” the land cells north of 60° S.

  • MissArea of the analysis land domain (Natural Earth 10m land north of 60 S, cells at least half land)

    Expected
    about 135 to 136 million km² (research note: 135.8, Natural Earth 50m land on a 0.25 degree grid)
    Result
    134.4883 million km² (1,991,832 cells). 1.0% below 135.8. The 50m polygons clipped at 60 S give 134.37 million km² when measured directly, so the difference is the research note’s coarse grid. Nothing was adjusted.
  • MatchGHS-POP 2020 total on the 4k grid

    Expected
    7,840,952,947
    Result
    7,840,952,947.0 (difference 0.0000 per million). 97.900% on domain cells, 2.072% on cells with some land but less than half, 0.028% on cells with no Natural Earth land.
  • MatchPopulation grid registration

    Expected
    best at zero shift
    Result
    99.972% of people fall on cells with some land; moving the grid by one 4k cell lowers that to 97.67 - 99.15%
  • MatchTexture alignment: Solar System Scope coastlines against Natural Earth on the 4k grid

    Expected
    best shift within one cell
    Result
    best shift (0, 0) cells; agreement 99.680% at zero shift; sub-cell offset at most 0.17 of a cell (texture coast about 1.2 km north)
  • MatchCountry totals against Natural Earth attributes (India, China, United States, Nigeria, Brazil, Russia, Australia, Greenland)

    Expected
    population within a few percent of POP_EST 2019; area within 1% of the polygon area
    Result
    India 1,397.2 million (1,366.4), China 1,425.6 (1,397.7), United States 337.3 (328.2), Nigeria 208.4 (201.0), Brazil 213.5 (211.1); areas within 0.6%

10Limits and caveats

These are the limits we know of. Several could move the result a long way. We list them all so that the map can be weighed properly.

In brief, with this build’s numbers

These lines are written by the data build each time the map is made, so the figures in them are those of the map now on the front page.

  • This is a long-run picture built from published data, not a forecast: it shows where each level of warming leads if it is held for a very long time (about ten thousand years for the sea), and says nothing about when, or whether, that level is reached.
  • Above 4 °C the heat layer is our own extrapolation of published maps that stop at 4 °C; in tests on hidden data it found about nine tenths of the real zone but drew the month-long zone 15 to 67% too large, so those areas probably err on the large side.
  • The heat limit was measured on 24 healthy young adults at rest or light activity in shade, and shading a place at 240 hours a year above it is our own choice; older and ill people reach their limit at a wet-bulb temperature about 2.4 to 3 °C lower, and a 56-hour rule shades about three times the land at 4 °C.
  • The water and food layer is cautious about drying and severe about heat: it finds four to six times less newly arid land than one published study, and above 4 °C most of its new land comes from a 36 °C limit for today’s maize, rice, wheat and sorghum, with no new varieties, shifted seasons or irrigation.
  • The land drawn in a muted colour, which already fails the water and food test today, is wider than desert: 12% of it (3.1% of all land, home to 86 million people) is semi-arid steppe such as the High Plains of the United States, marked because no three months in a row bring rain worth half of what evaporation could take; that test has no soil moisture, snowmelt or fallow in it and ignores temperature, so colder steppe whose wettest months are frozen passes it. In all, 556 million people live on this land now, on rivers, wells and imported food: the map shows where that dependence becomes total, not where people will leave.
  • Sea-level figures are close to the middle of the range assessed by the IPCC in 2021 for about ten thousand years after the warming (10 to 24 m around our 18 m at 3 °C); another class of ice-sheet models gives a little over half as much above 4 °C, and above 5 °C there is no assessed range.
  • The sea is raised by the same amount everywhere on an elevation map with cells about 2 km across and errors of a metre or more, with no sea walls, sinking deltas, rebounding land or melted ice sheets, so the blue shows regions, not towns.
  • People are counted where they live today (GHS-POP 2020), with no population growth and no migration, and rivers, groundwater, glaciers and trade are not modelled, although they are what keeps many shaded places inhabited now.

The full list follows.

The framing

  1. The map shows exposure, not a forecast of abandonment. People live today in places that fail these tests, using imported food, river water and air conditioning. The map shows where that dependence becomes total.
  2. Equilibrium is assumed but not modelled. The climate data behind the heat layer and the food and water layer are model states at the moment a warming level is crossed, under rising emissions. A climate settled at the same level would differ in its contrast between land and ocean, its ice cover and its rainfall. We have not measured by how much. Only the sea-level layer describes the long run (section 3).
  3. Warming levels are measured from 1850–1900. The baseline climate used for the map, 1970–2000, was already about 0.56 °C warmer.
  4. Other hazards are outside the map. Fire, river floods, storms, disease, conflict and migration are not modelled. Unshaded land is not safe land.

Heat

  1. Nothing above about 4 °C comes from a published gridded study. The heat layer from 4 to 6 °C is our extrapolation of the trend below it, and in our tests it drew the month-long zone too large. The best-known estimate for much greater warming comes from one coarse model in 2010 (Sherwood and Huber [44]), in which a wet-bulb temperature of 35 °C is first passed at about 7 °C of global warming.
  2. The human limits were measured on few people. They come from 24 young adults in one climate in the United States, not used to heat, at rest or light activity, in shade with moving air [52]. Limits are lower with age, illness, pregnancy, some medicines, work and direct sun. They may be a little higher in people who are used to heat. The limits for older adults come from 49 people [54].
  3. Survival limits are modelled, not measured. Published survival limits (Vanos and colleagues 2023 [50]) come from a heat-balance model that ends at a core temperature of 43 °C. They cannot be tested on people. We use the measured uncompensable limit instead.
  4. The duration threshold is our choice. “A month of daytime hours a year” is a design decision. No study of deaths fixes it. The lighter one-week band shows how much the choice matters: roughly three times the land at 4 °C.
  5. Monthly averages cannot stand in for the extremes that kill. We tried to rebuild the heat layer from monthly climate. It was off by up to about 3 °C of wet-bulb temperature in the Sahel, on the Gulf coast and in eastern China, and it failed our hidden-data test. Nothing in the heat layer comes from monthly climate: it uses the published three-hourly results to 4 °C and our extrapolation of them above.
  6. Wet-bulb values depend on the formula. Common formulas differ by up to 1 °C in hot dry air, which is as much as the difference between warming levels. Thresholds from different papers are not directly comparable. The map uses one paper’s hours and limit together, so it is consistent with itself.
  7. Grid cells average over coasts, cities and irrigated fields. The heat data have cells of about 28 km. Weather stations on the Persian Gulf record extremes about 3 °C above the gridded record (Raymond and colleagues 2020 [37]). Cities and irrigation raise local values. The map misses both.
  8. Humidity over land is the least certain input. A fall of 5% in relative humidity roughly doubles the area that crosses the young-adult limit at 2 °C, because dry heat becomes the limit (Matthews and colleagues 2025 [31]).
  9. A strip along the Greenwich meridian is repaired, not published. The published data are damaged there, from 1.75° W to 0.75° E. Part of the strip is interpolated from the columns on either side, and in the rest the published values are scaled up by a measured factor (section 5). The cause of the damage is inferred, not confirmed. In our blind test the scaled columns came out up to 3% too high, and the small solid-red patch on Lake Volta at 4 °C is uncertain. In Ghana, between about 6.5 and 10.5° N, the shading inside the strip is reconstructed: it is probably too large west of the lake, and the warming level at which it appears is uncertain by up to about 0.5 °C. The straight east side of that shading, at about 0.4° E, is the edge of the lake in the source data.
  10. “Deadly heat days” in older studies mean something weaker. The finding that about 30% of people already face 20 or more deadly days a year (Mora and colleagues 2017 [32]) uses a statistical threshold drawn from past heat waves, not a survival limit. We do not mix the two on one map.

Food and water

  1. Aridity is disputed. In published studies the aridity index projects that drylands expand. Measures based on soil moisture, runoff and vegetation project little or no expansion globally [3]. We use the aridity index, and our map finds several times less newly arid land than published studies that use it. Drying in the Mediterranean, southern Africa and subtropical South America appears in both.
  2. Rainfall change is the largest unknown. Models disagree on the sign in much of the tropics, and a median hides that. Temperature alone moves the desert boundary by only a few percent of land between 2 and 6 °C.
  3. Deserts and dry steppe are shown separately because they came before the warming. Land that fails the food and water rules at or below 1.5 °C covers 25% of the land outside Antarctica, almost all of it already in the 1970–2000 climate. The test is strict and also catches semi-arid steppe.
  4. Rivers, aquifers and glaciers are not modelled cell by cell. River-fed drylands are marked as depending on water from upstream. The loss of glacier and snow melt is not included. Groundwater is not considered at all.
  5. Crop limits are for today’s varieties. The thresholds describe maize, rice, wheat and sorghum as they are now, using monthly averages of daily highs. Changes in irrigation, planting dates, new varieties and protected growing are not included. Nor is any benefit to crops from more CO2. This rule carries most of the new land above 4 °C.
  6. Above about 5.2 °C the pattern of change is outside the models. The median model run reaches about 5.2 °C above 1850–1900 by the end of the century. Above that the pattern is extended in a straight line. The same holds for the climate-niche layer.
  7. Monthly climate only. The layer sees an average year. It has no drought years and no extremes within a month.
  8. One climate model has faulty rainfall files. BCC-CSM2-MR is about 36% wetter than the baseline climate in every month and in both scenarios, which cannot be a response to warming. We kept it, as one of 22 models in a median. Leaving it out changes the new failed land at 4 °C by under 2%.

Climate niche and people

  1. The climate niche is a correlation. It describes where people have lived, which also reflects history and economics. It marks the Amazon by 2.7 °C. The measured heat limit does not, even at 4 °C.
  2. People are counted where they live today. There is no population growth and no migration in the figures. The published studies we compare with use a fixed mid-century population.

Sea level

  1. The assessed range is wide and rests on two modelling studies. At 3 °C it runs from 10 to 24 m. The IPCC labels it “medium agreement, limited evidence” [13].
  2. A different family of ice-sheet models gives much less at the hot end. About 14 m against our 26 m at 4 °C [28].
  3. Above 5 °C there is no assessed range. We extended the line ourselves.
  4. The time scale is thousands of years. The layer is not a forecast for this century, for which the IPCC’s likely range is 0.63 to 1.01 m under very high emissions.
  5. The sea is raised by the same amount everywhere. In reality the rise is larger far from the melting ice and smaller near it. Land still rising after the last ice age, around Hudson Bay and the northern Baltic, would fare better than shown. Sinking deltas and cities would fare worse.
  6. Sea walls, erosion and new sediment are not included.
  7. The elevation data are coarse. Heights are averages over cells about 2 km across, with errors of a metre or more, measured from mean sea level and not from the high-tide line. The map shows regions, not towns or streets.
  8. Ice sheets are drawn at their present size. Over these lengths of time much of that ice would be gone.
  9. Inland basins. The lowland around the Caspian Sea joins the ocean at about 30 m in the elevation data. That depends on a low divide a few metres high, which the central curve passes by a small margin. Read it as indicative.

Not a town assessment

The map is a coarse global screen. It is not the assessment we would carry out for a single town, which looks at that town’s own water sources, buildings, power supply and terrain.

11Our warming estimate and the literature

The map has no dates. Our plan does. This section keeps three things apart: what the data show, what published scientists argue, and what we have chosen to plan for.

1What the data show

  • Berkeley Earth’s update for September 2026 [39] gives 1.65 ± 0.11 °C for the month and 1.458 ± 0.074 °C for the 12 months to September 2026, and puts the chance that 2026 becomes the warmest year on record at about 96%. It expects record and near-record warmth to continue through the end of the year “due to the short-term influence of a record-breaking El Niño”, and it makes no statement about the warming rate.
  • Berkeley Earth’s 2025 annual report [38] says its long-term trend has passed 1.4 °C above 1850–1900. Single years sit above or below the trend: in the multi-dataset average used by Forster and colleagues (2026) [10], observed warming was 1.51 °C in 2024 and 1.39 °C in 2025.
  • Human-caused warming was 1.37 °C in 2025 and is rising at about 0.27 °C per decade, with a range of 0.2 to 0.4. That is the fastest rate in the instrumental record (Forster and colleagues 2026 [10]).
  • In Berkeley Earth’s monthly data [41] the world warmed at about 0.19 °C per decade from 1970 to 2009, and at about 0.37 °C per decade from 2010 to October 2026. These are our own straight-line fits. The recent rate is uncertain by about 0.15 either way, and the record ends during a record El Niño, which raises it.
  • With El Niño, volcanoes and the solar cycle removed, the speed-up since about 2014 is statistically significant in five major temperature records (Foster and Rahmstorf 2026 [12]).
  • Earth’s energy imbalance, the rate at which the planet gains heat, about doubled in twenty years, from 0.5 to 1.0 W/m² (Loeb and colleagues 2024 [29]).

2What published scientists argue

  • Against the models. Recent warming is faster than the central projections and faster than nearly every model of the two earlier generations. It sits in the upper part of the range of the current models. It has not outrun the hottest current models, which warm faster than the real world has (Hausfather 2025 [17]; Foster and Rahmstorf 2026 [12]). Earth’s energy imbalance has risen faster than most models simulate, and the cause is not agreed (Hausfather 2026 [18]).
  • Evidence against a speed-up. With data to the end of 2023, Beaulieu and colleagues (2024) [1] detected no change in the warming rate beyond the 1970s in most temperature records. The study predates the data for 2024 to 2026.
  • 1.5 °C. The annual Indicators of Global Climate Change assessment expects human-caused warming to reach 1.5 °C around 2030 (Forster and colleagues 2026 [10]).
  • 2 °C. For medium to very high emissions, the IPCC’s central estimates put 2 °C in the 2040s or early 2050s, with the early 2030s as the earliest edge of the range [4]. In its two lowest-emission scenarios the central estimate stays below 2 °C all century [21]. James Hansen and colleagues argue that the climate is more sensitive than the IPCC’s best estimate and that 2 °C will arrive in the 2030s [15]; their peer-reviewed paper of 2025 gave 2045 [14]. That is a minority position. It has not been ruled out.
  • 3 °C. The United Nations Environment Programme’s 2026 report [48] puts current policies on course for about 2.8 °C by 2100, with a one-in-ten chance of more than 3.4 °C. In 2025 the German Meteorological Society and the German Physical Society said that 3 °C by 2050 can no longer be ruled out [6]. Other climate scientists called that very unlikely [53].

3Our planning estimate

We plan for about 2.0 °C by around 2030 and 3.0 °C between 2045 and 2050. That is at the pessimistic edge. It is roughly what continuing the pace of the last six years would give, and that window runs from a La Niña into a record El Niño, so most scientists expect the long-term average to rise more slowly: published estimates put 2 °C in the 2030s (Hansen and colleagues) to the 2040s or early 2050s (central IPCC estimates for medium to very high emissions), and current policies at about 2.8 °C by 2100 (UNEP). With deep cuts in emissions, the IPCC’s central estimate stays below 2 °C.

We may be too pessimistic. We plan for the early dates because preparing too late costs far more than preparing too early.

This estimate is ours alone. It is not Berkeley Earth’s, the IPCC’s or any other source’s.

How far our dates are from the published rates

A straight line through Berkeley Earth’s yearly averages for 2021 to 2026, with 2026 taken to October, rises at 0.87 °C per decade. It reaches 2.0 °C in 2031 and 3.0 °C in 2042. Starting the line in 2020 gives 2032 and 2047. Lines fitted over ten years or more give 2 °C around 2038 to 2040 and 3 °C in the 2060s. The table shows what steady rates from the literature would give. It is arithmetic, not a forecast.

Year each level is reached at a steady rate of warming, starting from human-caused warming of 1.37 °C in 2025. Our arithmetic.
Rate°C per decadeWhose rate2.0 °C3.0 °C
0.20Long-term trend since 1980 [38]20572107
0.27Forster and colleagues, central [10]20492086
0.36Foster and Rahmstorf, three records [12]20432071
0.42Foster and Rahmstorf, highest record [12]20402064

Our dates need more than that. From the same starting point, reaching 2.0 °C by 2030 needs about 1.3 to 1.5 °C per decade, and reaching 3.0 °C between 2045 and 2050 needs 0.65 to 0.84 °C per decade held for two decades.

The front page measures the same gap from a different starting point: the long-term level in Berkeley Earth’s record, which was 1.42 °C in October 2026. From there, reaching 2.0 °C by the middle of 2030 needs about 1.6 °C per decade, and reaching 3.0 °C between the start of 2045 and the start of 2050 needs 0.68 to 0.87 °C per decade. Either way the rates are far above any in the published work. We state this so that nobody mistakes the planning estimate for a forecast.

How our arithmetic on the record was done

  • Berkeley Earth’s data file measures from the 1951 to 1980 average. We shift every value to measure from the file’s own 1850–1900 average. Values re-based this way run 0.02 to 0.04 °C above the figures Berkeley Earth prints in its reports. We have not established why. With Berkeley Earth’s own conversion, the dates the fitted lines give for 2 °C and 3 °C would fall one to two and a half years later. Rates and rankings do not change.
  • Rates are straight lines fitted to monthly values. Each ± is two standard errors, widened to allow for autocorrelation by the more cautious of two published corrections (Santer and colleagues 2008 [42]; Foster and Rahmstorf 2011 [11]).
  • We did not remove El Niño, volcanic or solar effects. Foster and Rahmstorf (2026) [12] did.

The slider on the globe runs from 1.5 °C because that is about where single years are now; the long-term level is about 1.4 °C. It opens at 3.0 °C, the level in our name: the one we want communities to be ready before.

12All the checks

Each layer is checked against published results that it should reproduce: the share of land above the heat limit in the published grids, published estimates of new arid land, a published count of people on low coasts, and the land with a mean annual temperature of 29 °C or more.

The data build ran 68 checks on the map now on the front page. Each one compares a number the map produces with a published figure, or with a place that must or must not be shaded.

Match30
the published figure is reproduced
Close15
near it, with the difference explained
Miss15
the map disagrees with the published figure
Not tested8
there is no published figure to test against

The table under each layer gives the checks that matter most, in plain words. Under each table, a closed list holds every check on that layer in the data build’s own words, including the ones the map fails.

13Licences and reuse

This site is non-commercial. That matters, because two of the datasets it rests on may be used only on those terms.

Data for non-commercial use only

  • Berkeley Earth [41] [40]. Its data and reports are under CC BY-NC 4.0 (opens in a new tab): they may be copied and adapted for non-commercial purposes, with credit to Berkeley Earth, a link to the licence and a note of any changes. We re-based the record to 1850–1900, fitted the trends and drew the charts. Berkeley Earth is not responsible for our fits or extrapolations.
  • WorldClim [9] [55]. Free for academic and other non-commercial use. The data may not be passed on.

Only derived maps are published

We publish maps made from the data, not the data. The images the globe reads hold, for each cell, only the warming level at which a rule is first met. The climate grids, the heat grids, the elevation grid and the population grid are not on this site. To get them, go to their sources in the reference list.

The other sources

  • Heat data [51]. The only licence in the data archive is the GNU General Public License, version 3, which is a software licence. Zenodo lists the record as “other-open”. If our heat layer counts as a work based on those files, that licence asks for the same licence and for public source code. We have not settled that question, and the code that builds the map is not published yet (section 15).
  • IPCC Atlas series [24] and the GHS-POP population grid [43]: CC BY 4.0 (opens in a new tab).
  • ETOPO 2022 elevation [34]: CC0 1.0. Not to be used for navigation.
  • Natural Earth [33]: public domain.
  • Earth imagery [45]: CC BY 4.0 (opens in a new tab), based on NASA elevation and imagery data. We resized and re-encoded it.

How to credit

If you reuse a map or a chart from this site, the terms of the data behind it still apply: non-commercial use only, with credit to the sources. These lines cover it.

For the map
Map: community3.earth. Data: Vecellio et al. 2023 (heat); WorldClim 2.1 (climate); IPCC AR6 Working Group I (sea level); NOAA ETOPO 2022 (elevation); Natural Earth (coastlines); GHS-POP R2023A, European Commission JRC, CC BY 4.0 (population); Solar System Scope, CC BY 4.0, based on NASA imagery, modified (Earth imagery). Non-commercial use only.
For a temperature chart
Data: Berkeley Earth (www.berkeleyearth.org), CC BY-NC 4.0. Chart and re-baselining to 1850–1900 by community3.earth.

Please link to this page when you can, so that readers can find the method and its limits.

CMIP6 acknowledgement

The climate model results come from CMIP6. We repeat the acknowledgement that the IPCC Atlas repository asks for:

We acknowledge the World Climate Research Programme, which, through its Working Group on Coupled Modelling (WGCM), the Coupled Model Intercomparison Project (CMIP) and the Coordinated Regional Downscaling Experiment (CORDEX), coordinated and promoted the production of the global (CMIP5 and CMIP6) and regional (CORDEX-CMIP5) climate projection datasets. We thank the climate modelling groups for producing and making available their model output, the Earth System Grid Federation (ESGF) for archiving the data and providing access, and the multiple funding agencies who support this work.

14References

Every paper, report and dataset used anywhere on this site, 57 in all, in alphabetical order. Numbers in square brackets in the text point here. Links open in a new tab.

We give a title only where we have it from the publication itself, and we say where we read only an abstract or a copy. Datasets are marked, with their licence.

  1. Beaulieu, C. et al. (2024). A recent surge in global warming is not detectable yet. Communications Earth & Environment 5, 576. doi:10.1038/s43247-024-01711-1 (opens in a new tab).

    PaperUsed for: The main published evidence against a detectable speed-up, with data to the end of 2023.

  2. Beck and colleagues (2018). Open full text. PubMed Central PMC6207062 (opens in a new tab).

    PaperUsed for: The definition of the desert class in the Köppen–Geiger climate scheme, which we use for one comparison.

  3. Berg, A. and McColl, K. A. (2021). Nature Climate Change. doi:10.1038/s41558-021-01007-8 (opens in a new tab).

    PaperUsed for: The case against the aridity index: measures built on soil moisture, runoff and vegetation show little or no global expansion of drylands.

  4. Carbon Brief. Analysis of what the IPCC’s 2021 report says about when the world may pass 1.5 °C and 2 °C. carbonbrief.org (opens in a new tab).

    ArticleUsed for: The years in which the IPCC’s scenarios cross 2 °C, and our reading of the IPCC’s Table SPM.1. We read both here because the IPCC’s own site refused automated access.

  5. Clark, P. U. et al. (2016). Consequences of twenty-first-century policy for multi-millennial climate and sea-level change. Nature Climate Change 6, 360–369. doi:10.1038/nclimate2923 (opens in a new tab).

    PaperUsed for: One of the two modelling studies behind the IPCC’s long-run sea-level ranges. We used its simulations through the IPCC’s data files; we could not open the paper itself.

  6. Deutsche Meteorologische Gesellschaft and Deutsche Physikalische Gesellschaft (2025). Globale Erwärmung beschleunigt sich: Ein Aufruf zu entschlossenem Handeln. Joint call, September 2025; read in the annex to a session of the German Bundestag’s advisory council on 15 October 2025. bundestag.de (PDF, in German) (opens in a new tab).

    ReportUsed for: The statement that 3 °C by 2050 can no longer be ruled out. It is a call to action by two learned societies, not a peer-reviewed projection.

  7. Food and Agriculture Organization of the United Nations. Growing-period classes, two web pages. fao.org/4/x5756e (opens in a new tab); fao.org/4/t8300e (opens in a new tab).

    ReportUsed for: The rule that a growing period starts when rainfall exceeds half of potential evapotranspiration.

  8. Feng and Fu (2013). Atmospheric Chemistry and Physics 13, 10081. acp.copernicus.org (opens in a new tab).

    PaperUsed for: A published estimate of how far drylands grow by 2100 under high emissions, used to check the food and water layer. We read the abstract only.

  9. Fick, S. E. and Hijmans, R. J. (2017). WorldClim 2: new 1 km spatial resolution climate surfaces for global land areas. International Journal of Climatology 37 (12), 4302–4315. Data: WorldClim 2.1, climate of 1970–2000 at 10 arc-minutes. worldclim.org (opens in a new tab).

    Dataset Licence: WorldClim terms: free for academic and other non-commercial use; redistribution is not allowed. Used for: The present-day climate behind the food and water layer and the climate-niche layer. We publish maps made from the data and do not pass on the data themselves.

  10. Forster, P. M. et al. (2026). Indicators of Global Climate Change 2025. Earth System Science Data 18, 3889–3933. doi:10.5194/essd-18-3889-2026 (opens in a new tab).

    PaperUsed for: Human-caused warming in 2025, its rate per decade, observed warming in 2024 and 2025, and the expectation of 1.5 °C around 2030.

  11. Foster and Rahmstorf (2011). Environmental Research Letters 6 (4), 044022. doi:10.1088/1748-9326/6/4/044022 (opens in a new tab).

    PaperUsed for: One of the two corrections for autocorrelation behind the ± values of our trend fits.

  12. Foster, G. and Rahmstorf, S. (2026). Global Warming Has Accelerated Significantly. Geophysical Research Letters 53, e2025GL118804. doi:10.1029/2025GL118804 (opens in a new tab).

    PaperUsed for: The test of the speed-up with El Niño, volcanic and solar effects removed, and the rates of 0.34 to 0.42 °C per decade.

  13. Fox-Kemper, B. et al. (2021). Ocean, Cryosphere and Sea Level Change. Chapter 9 of Climate Change 2021: The Physical Science Basis, Working Group I contribution to the Sixth Assessment Report of the IPCC, pp. 1211–1362. Cambridge University Press. doi:10.1017/9781009157896.011 (opens in a new tab).

    ReportUsed for: Table 9.10: the committed sea-level rise about 10,000 years after 1.5, 2, 3, 4 and 5 °C of warming. We took the middle of each range, smoothed it to a straight line, widened the band at 1.5 °C and extended it above 5 °C. Also the IPCC’s figures for 2100.

  14. Hansen, J. E. et al. (2025). Global Warming Has Accelerated: Are the United Nations and the Public Well-Informed? Environment: Science and Policy for Sustainable Development 67 (1), 6–44. doi:10.1080/00139157.2025.2434494 (opens in a new tab).

    PaperUsed for: The peer-reviewed statement of the higher climate sensitivity, with 2 °C by 2045.

  15. Hansen, J., Kharecha, P., Morgan, D. and Vest, J. (2026). Another El Nino Already? What Can We Learn from It? 6 February 2026. jimehansen.substack.com (opens in a new tab).

    ArticleUsed for: The argument that 2 °C is likely to be reached in the 2030s. The authors measure from 1880–1920.

  16. Hatfield, J. L. and Prueger, J. H. (2015). Weather and Climate Extremes. doi:10.1016/j.wace.2015.08.001 (opens in a new tab).

    PaperUsed for: Heat limits of rice, sorghum and maize, behind the crop-heat rule.

  17. Hausfather, Z. (2025). Comparing climate models with observations. The Climate Brink, December 2025. theclimatebrink.com (opens in a new tab).

    ArticleUsed for: How recent warming compares with three generations of climate models, including the hottest current ones.

  18. Hausfather, Z. (2026). State of the climate: 2026 likely to tie with 2024 as the warmest year on record. Carbon Brief, 3 November 2026. carbonbrief.org (opens in a new tab).

    ArticleUsed for: The statement that Earth’s energy imbalance appears to have risen faster than most climate models simulate.

  19. Huang, J. et al. (2016). Nature Climate Change. Read in the authors’ copy. hjp.lzu.edu.cn (PDF) (opens in a new tab).

    PaperUsed for: A published projection of drylands by 2100, used to check the food and water layer, and the aridity classes attributed to the UN Environment Programme.

  20. Immerzeel, W. W. et al. (2020). Nature. doi:10.1038/s41586-019-1822-y (opens in a new tab).

    PaperUsed for: The number of people living in or directly downstream of mountain water towers. We read the abstract only.

  21. IPCC (2021). Summary for Policymakers. In Climate Change 2021: The Physical Science Basis, Working Group I contribution to the Sixth Assessment Report of the IPCC. Cambridge University Press. doi:10.1017/9781009157896.001 (opens in a new tab).

    ReportUsed for: Sea level in past warm periods and over the next 2,000 years (paragraph B.5.4), and the warming projections of Table SPM.1.

  22. IPCC (2023). Atlas. In Climate Change 2021: The Physical Science Basis, Working Group I contribution to the Sixth Assessment Report of the IPCC, pp. 1927–2058. Cambridge University Press. doi:10.1017/9781009157896.021 (opens in a new tab).

    ReportUsed for: Cited with the Atlas repository, as its licence asks.

  23. IPCC Working Group I. Chapter 9 figure code and data, Figure 9.30 (sea-level commitment). github.com/IPCC-WG1/Chapter-9 (opens in a new tab).

    Dataset Licence: none recorded in our notes; no IPCC data file is passed on by this site. Used for: The model runs behind the IPCC’s sea-level table, from which we read the values at 2.5, 3.5 and 4.5 °C and extended the curve to 6 °C.

  24. Iturbide, M., Fernández, J., Gutiérrez, J. M. et al. (2022). Implementation of FAIR principles in the IPCC: the WGI AR6 Atlas repository. Scientific Data 9, 629. Data: IPCC Working Group I Atlas repository, global mean temperature series of the CMIP6 models. doi:10.1038/s41597-022-01739-y (opens in a new tab); github.com/IPCC-WG1/Atlas (opens in a new tab).

    Dataset Licence: CC BY 4.0 (opens in a new tab). Used for: How much global warming each climate-model run stands for: the divisor of the local patterns.

  25. Koutroulis (2019). Science of the Total Environment. doi:10.1016/j.scitotenv.2018.11.215 (opens in a new tab).

    PaperUsed for: Published estimates of dryland growth and of land changing aridity class at 4 °C, used to check the food and water layer. We read the abstract only.

  26. Kulp, S. A. and Strauss, B. H. (2019). New elevation data triple estimates of global vulnerability to sea-level rise and coastal flooding. Nature Communications 10, 4844. PubMed Central PMC6820795 (opens in a new tab).

    PaperUsed for: People living on land less than 5 m and 10 m above the high-tide line, used to check the sea-level layer.

  27. Lenton, T. M. et al. (2023). Nature Sustainability. doi:10.1038/s41893-023-01132-6 (opens in a new tab).

    PaperUsed for: Land and people with a mean annual temperature of 29 °C or more at each level of warming, used to check the climate-niche layer.

  28. Levermann, A. et al. (2013). The multimillennial sea-level commitment of global warming. Proceedings of the National Academy of Sciences 110, 13745–13750. PubMed Central PMC3752235 (opens in a new tab).

    PaperUsed for: A different family of ice-sheet models, which gives a lower long-run sea level at 3 and 4 °C than the curve we use.

  29. Loeb, N. G. et al. (2024). Observational Assessment of Changes in Earth’s Energy Imbalance Since 2000. Surveys in Geophysics 45. doi:10.1007/s10712-024-09838-8 (opens in a new tab).

    PaperUsed for: The doubling of Earth’s energy imbalance in twenty years.

  30. Luo, Q. (2011). Climatic Change (a review). pagesperso.locean-ipsl.upmc.fr (PDF) (opens in a new tab).

    PaperUsed for: Heat limits of wheat, maize, rice and sorghum, behind the 36 °C crop-heat rule.

  31. Matthews, T. et al. (2025). Nature Reviews Earth & Environment. doi:10.1038/s43017-024-00635-w (opens in a new tab).

    PaperUsed for: How strongly the area of uncompensable heat depends on humidity over land.

  32. Mora, C. et al. (2017). Nature Climate Change. We read the abstract only. researchonline.jcu.edu.au/59791 (opens in a new tab).

    PaperUsed for: The “deadly heat days” figure, which uses a much weaker threshold than this map. Named so that the two are not confused.

  33. Natural Earth. 1:10 million land, ocean, lakes, rivers, glaciated areas and countries. Made with Natural Earth: free vector and raster map data. naturalearthdata.com (opens in a new tab).

    Dataset Licence: public domain. Used for: Today’s coastline and land area, the ocean from which the flood starts, the river corridors, present-day ice, and country borders for the checks.

  34. NOAA National Centers for Environmental Information (2022). ETOPO 2022 15 Arc-Second Global Relief Model. NOAA National Centers for Environmental Information. Accessed 16 November 2026. Not to be used for navigation. doi:10.25921/fd45-gt74 (opens in a new tab).

    Dataset Licence: CC0 1.0. Used for: Land heights, in the 60 arc-second ice-surface version, to find which land a higher sea can reach.

  35. Pangeo and Google Cloud. CMIP6 archive. storage.googleapis.com/cmip6 (opens in a new tab).

    Dataset Licence: CMIP6 terms of use. Used for: Global temperature of the five climate models that the IPCC Atlas repository lacks.

  36. Pesaresi, M. et al. (2024). Advances on the Global Human Settlement Layer by Joint Assessment of Earth Observation and Population Survey Data. International Journal of Digital Earth 17 (1). doi:10.1080/17538947.2024.2390454 (opens in a new tab).

    PaperUsed for: The paper to cite with the GHS-POP population grid.

  37. Raymond, C., Matthews, T. and Horton, R. M. (2020). Science Advances. doi:10.1126/sciadv.aaw1838 (opens in a new tab).

    PaperUsed for: Weather stations on the Persian Gulf record humid-heat extremes about 3 °C above gridded weather data.

  38. Rohde, R. (2026). Global Temperature Report for 2025. Berkeley Earth, 14 January 2026, updated 24 March 2026. berkeleyearth.org (opens in a new tab).

    Report Licence: CC BY-NC 4.0 (opens in a new tab). Used for: The three quotations on the front page about the warming spike of 2023 to 2025, and the statement that Berkeley Earth’s long-term trend has passed 1.4 °C.

  39. Rohde, R. and Hausfather, Z. (2026). September 2026 Temperature Update. Berkeley Earth, 22 October 2026. berkeleyearth.org (opens in a new tab).

    Report Licence: CC BY-NC 4.0 (opens in a new tab). Used for: Berkeley Earth’s published figures on the front page: September 2026, the 12 months to September 2026, and the chance that 2026 is the warmest year.

  40. Rohde, R. A. and Hausfather, Z. (2020). The Berkeley Earth Land/Ocean Temperature Record. Earth System Science Data 12, 3469–3479. Data: Berkeley Earth land and ocean record at 1°, last updated December 2024. doi:10.5194/essd-12-3469-2020 (opens in a new tab); berkeleyearth.org/data (opens in a new tab).

    Dataset Licence: CC BY-NC 4.0 (opens in a new tab). Used for: The difference between the 1970–2000 climate and 1850–1900 that the map uses.

  41. Rohde, R. A., Calvert, B. and Hausfather, Z. (2026). Berkeley Earth Surface Temperature – High-Resolution (BEST-HR): a 0.25° Global Gridded Temperature Data Set for Climate Monitoring. Earth System Science Data Discussions, preprint, in review. Data: high-resolution global land and ocean series, beta release. doi:10.5194/essd-2026-412 (opens in a new tab); berkeleyearth.org/data (opens in a new tab).

    Dataset Licence: CC BY-NC 4.0 (opens in a new tab). Used for: The temperature record behind the three charts on the front page. We re-based it to 1850–1900, fitted the trends and drew the charts.Retrieved 16 November 2026.

  42. Santer and colleagues (2008). International Journal of Climatology. doi:10.1002/joc.1756 (opens in a new tab).

    PaperUsed for: One of the two corrections for autocorrelation behind the ± values of our trend fits.

  43. Schiavina, M., Freire, S., Carioli, A. and MacManus, K. (2023). GHS-POP R2023A – GHS population grid multitemporal (1975–2030). European Commission, Joint Research Centre. Used: population in 2020 at 30 arc-seconds. doi:10.2905/2FF68A52-5B5B-4A22-8F40-C41DA8332CFE (opens in a new tab).

    Dataset Licence: CC BY 4.0 (opens in a new tab). Used for: Every count of people who live on shaded land today. We summed the grid to the map’s cells.

  44. Sherwood, S. C. and Huber, M. (2010). Proceedings of the National Academy of Sciences. PubMed Central PMC2906879 (opens in a new tab).

    PaperUsed for: The theoretical limit of 35 °C wet-bulb, and the only published estimate of when it is first passed, at about 7 °C of warming.

  45. Solar System Scope. Earth textures, based on NASA elevation and imagery data. Modified by us: resized and re-encoded. solarsystemscope.com/textures (opens in a new tab).

    Dataset Licence: CC BY 4.0 (opens in a new tab). Used for: The picture of the Earth under the shading, on the globe and in the flat maps.

  46. Spinoni, J. et al. (2021). Global and Planetary Change. JRC publications repository, JRC123579 (opens in a new tab).

    PaperUsed for: Land shifting to an arid climate at 1.5, 2, 3 and 4 °C, used to check the food and water layer. We read the abstract only.

  47. Stone, B. et al. (2023). Environmental Science & Technology. PubMed Central PMC10249403 (opens in a new tab).

    PaperUsed for: What a blackout lasting several days does to heat deaths in Atlanta, Detroit and Phoenix.

  48. United Nations Environment Programme (2026). Emissions Gap Report 2026: The point of return. Figures as reported by Carbon Brief, 21 October 2026; the UN’s own site refused automated access. carbonbrief.org (opens in a new tab).

    ReportUsed for: Where current policies lead: about 2.8 °C this century, with a one-in-ten chance of more than 3.4 °C.

  49. Van Breedam, J., Goelzer, H. and Huybrechts, P. (2020). Semi-equilibrated global sea-level change projections for the next 10 000 years. Earth System Dynamics 11, 953–976. esd.copernicus.org (opens in a new tab).

    PaperUsed for: The other of the two modelling studies behind the IPCC’s long-run sea-level ranges.

  50. Vanos, J. et al. (2023). Nature Communications. doi:10.1038/s41467-023-43121-5 (opens in a new tab).

    PaperUsed for: Modelled survival limits, which we name and do not use.

  51. Vecellio, D. J., Kong, Q., Kenney, W. L. and Huber, M. (2023). Greatly enhanced risk to humans as a consequence of empirically determined lower moist heat stress tolerance. Proceedings of the National Academy of Sciences. Data: Zenodo record 8347178, version Vecellio2023_PNAS_001. doi:10.1073/pnas.2305427120 (opens in a new tab); data: doi:10.5281/zenodo.8347178 (opens in a new tab).

    Dataset Licence: the data archive carries the GNU General Public License, version 3; Zenodo lists the record as “other-open”. Used for: Hours a year above the measured human heat limit at 1 to 4 °C of warming: the heat layer up to 4 °C, and the starting point of our extrapolation above it. We interpolated between the published levels, repaired one damaged strip and extrapolated above 4 °C.

  52. Vecellio, D. J., Wolf, S. T., Cottle, R. M. and Kenney, W. L. (2022). Journal of Applied Physiology. PubMed Central PMC8799385 (opens in a new tab).

    PaperUsed for: The measured heat limit of healthy young adults: about 30.6 °C wet-bulb in humid heat, and lower in hot dry air.

  53. web.de (2025). News report of 4 October 2025 with the responses of three climate scientists to the German societies’ call, in German. web.de (opens in a new tab).

    ArticleUsed for: The view of other climate scientists that 3 °C by 2050 is very unlikely.

  54. Wolf, S. T. et al. (2023). Communications Earth & Environment. PubMed Central PMC10826365 (opens in a new tab).

    PaperUsed for: The measured heat limit of older adults, which is lower.

  55. WorldClim. WorldClim 2.1 downscaled CMIP6 future climate at 10 arc-minutes: monthly temperature and rainfall of 22 climate models for 2081–2100 under SSP5-8.5 and 2041–2060 under SSP2-4.5. Cite as Fick and Hijmans (2017). worldclim.org (opens in a new tab).

    Dataset Licence: WorldClim terms, as for the baseline climate; CMIP6 terms of use. Used for: The pattern of local warming and rainfall change for each degree of global warming.

  56. Xu, C. et al. (2020). Proceedings of the National Academy of Sciences. PubMed Central PMC7260949 (opens in a new tab).

    PaperUsed for: The temperatures at which people have lived, and the 29 °C mean annual temperature used for the climate-niche layer.

  57. Zomer and colleagues (2022). Open full text. PubMed Central PMC9287331 (opens in a new tab).

    PaperUsed for: The aridity classes attributed to the UN Environment Programme, and a published aridity dataset built on the same climate with the same formula for potential evapotranspiration.

15Tell us where we are wrong

This method has not yet been reviewed by outside scientists. We want it to be. If you work on climate, heat physiology, water, crops, ice or sea level, and you see an error, a better dataset or a threshold we should change, please tell us. The tables on this page show where the map already misses published work. Those are good places to start.

Use the form on the front page and choose “Climate and earth-system scientists”. It asks four short questions, all optional. We read every message. Corrections will be recorded here with their date. The code that builds the map is not published yet; if you want to check a step, ask us through the same form.

Open the form for scientists

Change log

  • November 2026: first version.
  • November 2026: corrections after our own review of this page. The heat layer never falls back on monthly climate; the first version said it could. The 0.56 °C between 1970–2000 and 1850–1900 comes from Berkeley Earth’s earlier record, not from the high-resolution one. Three inland basins join the sea as the slider moves, not two. The reference list was completed and numbered, and a comparison with published work was added for each layer.
  • November 2026: the repair of the heat data at the Greenwich meridian was widened after an independent check of the data build, and is described here as it now stands. The muted class on the map is now called “deserts and dry steppe today”, because part of it is semi-arid steppe and not desert (section 6).