How Climate Twins works

The question

A climate analog is a location whose present climate resembles the projected future climate of another location. Analogs translate abstract projections, such as “3 °C warmer summers”, into a place people can picture, and they have a long history in ecology and adaptation planning, from the study of novel and disappearing climates (Williams et al. 2007; Williams & Jackson 2007) to maps of analogs for North American cities (Fitzpatrick & Dunn 2019).

For each place, scenario and period, the map reports the grid cell whose 1991–2020 climate is closest to the place's projected climate, four further good matches, the land area within 2σ, and how strongly the individual climate models agree. When no cell is close, the map says so rather than drawing a misleading arrow.

Places

North America (790 places): 635 in the contiguous United States, 23 in Alaska, 56 in Canada and 76 in Mexico, drawn from the GeoNames gazetteer. Places were selected greedily by population with a minimum spacing that widens as population falls (in the United States: 100,000 people at 25 km spacing, then 25,000 at 45 km, 8,000 at 75 km and 1,000 at 120 km; in Canada and Mexico: 250,000 at 35 km, 40,000 at 110 km and 5,000 at 220 km), so that large cities are all present and rural regions still have coverage. Two North Carolina places were added by hand.

World (1,596 cities): every city of at least one million people outside the United States, Canada and Mexico; all 202 national capitals (Natural Earth); Honolulu, which lies outside the North American grid; and cities chosen by population with minimum spacing to cover sparsely populated regions. Together they span 208 countries and territories. Country names come from Natural Earth, shortened to common forms.

The world list was later extended from GeoNames with two rules. First, every country or territory has at least three places wherever GeoNames has them. Second, further places were added greedily by population, each at least a minimum distance from every place already listed. Elsewhere that distance runs from 80 km for cities of a million people to 480 km for towns of 40,000; in Africa, Latin America, Central Asia, Oceania and the Middle East, where the first list was sparse, it runs from 70 km for cities of 300,000 to 400 km for towns of 5,000. A place is kept only if its 0.5° cell, or a neighbouring one, is a land cell of the world grid, and the grid stops at 60° S. Places on very small islands are included only to reach the three-per-country minimum.

Present-day climate

Climate is described by sixteen numbers: the mean daily maximum temperature, mean daily minimum temperature, mean dewpoint and total precipitation for each meteorological season (December–February, March–May, June–August, September–November). Precipitation enters the comparison as log(mm + 1), so that a given relative change counts similarly in wet and dry climates.

North America

Present climate comes from the AdaptWest 1991–2020 monthly normals at 1 km, generated with ClimateNA v7.30 (AdaptWest Project 2022; Wang et al. 2016). For the comparison grid, 1 km cells were averaged into 15 km blocks on a Lambert azimuthal equal-area projection centred at 45° N, 100° W, keeping blocks that are at least 25% land: 97,488 cells. Each place's own present climate is the mean of the 1 km cells within 10 km of it.

Worldwide

Present climate comes from TerraClimate (Abatzoglou et al. 2018), 1991–2020 monthly values at 1/24° (about 4 km). For the global comparison grid, values were averaged to 0.5° cells that are at least 25% land, covering all land except Antarctica and south of 60° S: 64,958 cells, including North America. Each world city's present climate is the mean of TerraClimate land cells within 10 km. Because the city and the grid use the same dataset, a city compared with its own surroundings scores close to 0σ (median 0.00σ).

Year-to-year variability

Distances are scaled by how much each season's climate varies from year to year at the place itself, estimated from 30 seasonal values (1991–2020), each series linearly detrended so that the warming trend is not counted as variability. In the contiguous United States these come from PRISM monthly 4 km data (PRISM Climate Group; Daly et al. 2008). Elsewhere, including Alaska, Canada, Mexico and all world cities, they come from TerraClimate. Where both are available, the two agree closely (TerraClimate-to-PRISM ratio of standard deviations: 1.05 for summer maximum temperature, 0.90 for precipitation).

Humidity

Humidity is described by the dewpoint, the temperature at which the air would become saturated. It tracks the moisture people feel better than relative humidity, which swings with the time of day. Present-day dewpoints come from TerraClimate monthly vapour pressure (Abatzoglou et al. 2018) everywhere, including North America, where the 1/24° values are averaged over the 3 × 3 pixels nearest each 15 km cell; a season's dewpoint is computed from its mean vapour pressure with the Magnus formula (Alduchov & Eskridge 1996). Of the four seasonal dewpoints, December–February and June–August enter the comparison, covering the humid season in both hemispheres; the other two are shown but not matched, so humidity carries weight comparable to its importance without swamping temperature and precipitation. Fourteen measures are matched in all.

Future humidity applies each model's humidity change, as a ratio limited to 0.5–3, to the observed vapour pressure: from NEX-GDDP-CMIP6 relative humidity and temperature where that dataset is used (see Projected climate), otherwise the change in near-surface specific humidity (huss). One model provides no humidity output for one scenario (FGOALS-g3 under SSP1-2.6); there, vapour pressure is scaled with that model's temperature change at constant relative humidity, following the Clausius–Clapeyron relation.

Projected climate

Projected climate follows the change-factor (delta) approach (e.g. Maraun & Widmann 2018): each model's simulated change is added to the observed present climate at the place, which removes the model's own bias in the baseline.

• Periods. 2041–2060 (“2050”) and 2081–2100 (“2100”), the IPCC AR6 mid- and end-of-century periods, compared with 1991–2020. The model baseline joins each model's historical run (1991–2014) to the chosen scenario (2015–2020).
• Scenarios. SSP1-2.6, SSP2-4.5, SSP3-7.0 and SSP5-8.5 from ScenarioMIP (O'Neill et al. 2016; Riahi et al. 2017).
• Variables. Monthly near-surface daily maximum and minimum temperature and precipitation (CMIP6 tasmax, tasmin, pr), as monthly climatologies for each window.
• Source of the change: downscaled where available. For every model and scenario that NASA's NEX-GDDP-CMIP6 publishes (Thrasher et al. 2022; 20 of the 24 models, all four scenarios), the change is taken from that dataset instead of the model's own coarse grid: daily output statistically downscaled to 0.25° (about 25 km) and bias-corrected against observations, averaged to monthly climatologies over the same windows (every second year sampled) at the 0.25° cell of the place. Temperature changes are added and precipitation ratios limited exactly as below; the humidity change is the ratio of vapour pressure computed from NEX-GDDP relative humidity and temperature (limited to 0.5–3). A 25 km cell follows coasts and mountains far better than a 100–250 km model cell, so the change fits the place better; the downscaling does not change a model's large-scale warming. Where NEX-GDDP lacks a model (AWI-CM-1-1-MR, CNRM-CM6-1-HR, CanESM5-CanOE, EC-Earth3-Veg), a scenario, humidity (BCC-CSM2-MR) or a land cell at the place, the model's native-grid CMIP6 change described in the next bullet is used for that piece. The ensembles keep the same models either way. NEX-GDDP uses one simulation per model, sometimes a different one from the members averaged here, and joins historical years to SSP2-4.5 for 2015–2020 in the baseline. The build report states which source each model used and how far the answer moves compared with CMIP6 changes alone; data/manifest.json records it per model and scenario. The warming-level view keeps the native-grid CMIP6 changes, because NEX-GDDP starts in 1950 and cannot give the 1850–1900 reference.
• Extraction (native-grid CMIP6). Model values at a place are a land-weighted Gaussian average of nearby model grid cells, with a length scale of one grid spacing, cut off at two spacings, and weights multiplied by each cell's land fraction so that ocean cells do not dilute land climate.
• Applying the change. Temperature changes are added month by month. Precipitation changes are applied as ratios, limited to the range 0.2–5. Where a model's baseline month has no precipitation at all, the ratio is set to 1. The minimum temperature is kept at least 0.1 °C below the maximum.
• Ensemble mean. The map's main answer uses the mean of the models' projected climates. Precipitation is averaged in millimetres before the log transform.

Here “obs” is the observed 1991–2020 value at the place, and “fut” and “base” are the model's own future and 1991–2020 means for the same month.

Warming levels

The map can also be read by how much the world has warmed instead of by year and scenario (the “Warming level” choice under View, at the top of the panel). This follows the IPCC AR6 idea of global warming levels (Lee et al. 2021): a warming level of +2 °C is the same amount of global warming whichever scenario, and whichever year, produces it. Regional climate at a given level depends far less on the scenario than the year does, so the view pools scenarios and is less tied to a particular emissions story.

• Levels and reference. +1.5, +2, +3 and +4 °C of global warming above the 1850–1900 average, the AR6 convention. Warming is measured on each model's own global mean annual near-surface air temperature (tas, area-weighted), against that model's own 1850–1900 mean. The 1991–2020 baseline used for “here now” is itself already about +0.8 °C above 1850–1900 in the models, so +1.5 °C is a modest step beyond today.
• When a level is reached. For each model and scenario, the first year in which the centred 20-year running mean of global temperature (the 10 years before and the 9 years after) is at or above the level. That year is taken as the centre of a 20-year window, for example 2039–2058. The first-crossing rule is used and later dips are ignored. Windows that would run past 2100 are not used, so a level a model reaches only in its last decade is treated as not reached.
• The change. Each model's regional change over that window minus its own 1991–2020 baseline, in exactly the same way as for the fixed periods (temperature differences, precipitation and humidity ratios, same limits). Windows that straddle 2014 and 2015 join the historical run to the scenario run, weighted by years.
• Pooling. For each model, the changes from every scenario that reaches the level are averaged with equal weight. The ensemble is then the mean over the models that reach the level; a level is offered only if at least three of the ensemble's models reach it. Under the default likely-range screen, +4 °C is usually not offered because few of those models warm that much this century; including the hotter models adds it. The panel states how many models contribute.
• Everything else is unchanged: change factors added to observed normals, sigma dissimilarity, the same pools and match rules. Sea-level projections are only shown in the year-and-scenario view, because they are tied to scenario and year.

Choices worth knowing: warming levels come from model temperatures, not observations, and models warm at different speeds, so “+2 °C” arrives in different years in different models (the panel shows the range). Averaging several members per model reduces, but does not remove, the noise from natural variability in 20-year windows, most for the +1.5 °C window where the change is small.

Climate models

The map uses 24 CMIP6 models (Eyring et al. 2016), each represented by the average of up to 3 initial-condition simulations (“members”) where the archive has them, read from the Pangeo CMIP6 cloud archive (Abernathey et al. 2021). Several CMIP6 models warm faster in response to CO₂ than observations and other lines of evidence support. Following Hausfather et al. (2022), the default ensemble keeps the 15 models whose transient climate response (TCR) lies in the IPCC AR6 likely range of 1.4–2.2 °C (Forster et al. 2021, Table 7.SM.5). The remaining nine can be added under Advanced on the map: seven above the range, one below it (INM-CM4-8) and one without a TCR in that table (EC-Earth3-Veg-LR). Including them mostly shifts results toward hotter outcomes.

Measuring similarity

Two climates are compared relative to the natural year-to-year variability of the place being studied. A difference of 1 °C in summer highs matters more at a place where summers barely vary than at one where they swing widely, and a mismatch in a pattern that rarely varies matters more than one in a pattern that varies a lot.

For place t, each of the 14 matched measures is standardised by its detrended interannual standard deviation s. The correlations among the 14 detrended series, estimated from only 30 years, are stabilised with Ledoit–Wolf shrinkage toward the identity matrix (Ledoit & Wolf 2004), with the shrinkage intensity α estimated from each place's own data (median 0.44 in North America, range 0.15–0.70; median 0.40 for world cities, range 0.16–0.76). The distance between the place's projected climate x and a candidate cell's present climate y is the Mahalanobis distance in this metric:

Following Mahony et al. (2017), D is converted to sigma dissimilarity: its percentile in a χ distribution with k = 14 degrees of freedom is re-expressed as the number of standard deviations of a one-dimensional normal distribution with the same two-sided tail probability. This makes distances comparable across places and dimensions, and gives the unit a familiar meaning.

Seasons that almost never rain

In very dry seasons (for example in Lima, Cairo or Khartoum) precipitation barely varies between years, which makes the distance undefined. For world cities, any standardised series with a standard deviation below 0.10 (in log precipitation, roughly 10%; or 0.10 °C) receives a fixed, seeded, zero-mean random series that raises its standard deviation exactly to that floor without correlating with the other measures. This affected 26 cities, mostly in a single measure. No North American place needed it.

Choosing and presenting matches

• Best match: the cell with the smallest σ for the ensemble-mean projection.
• Further matches: the next best cells at least 150 km (North America) or 300 km (worldwide) from every match already chosen, up to five in total. Matches under 1σ are about equally good.
• Area within 2σ: the summed land area of cells under 2σ, outlined on the map (true cell areas on the global grid, using latitude and land fraction).
• Model agreement: the share of individual models whose own best match lies within 500 km of the ensemble's (1,000 km for worldwide searches).
• A second method: as a robustness check, distances are recomputed with principal-component truncation instead of shrinkage, dropping components that carry under 1% of the year-to-year variance (as in Mahony et al. 2017).
• Confidence: high when at least 60% of models agree and the second method's best match is within 500 km; low when under 40% agree, or when the second method disagrees and fewer than 60% agree; medium otherwise.
• No match: when the best σ is 4 or more, or 2 or more with the best cell at the place itself, the map states that nowhere in the comparison area has this climate today and draws a ring instead of an arrow.
• Worldwide matches for North American places: when the best North American match is 2σ or worse, the place is also searched against the global grid, using its TerraClimate present climate so that both sides come from the same dataset.
• Labels: each cell is named after its nearest GeoNames populated place (worldwide: of at least 1,000 people), with the distance when it is not close.

All of this is computed in the reader's browser at the moment a place is chosen, from per-place metric matrices shipped with the page, so every cell on the map carries its own σ. The browser results match the Python pipeline cell for cell.

Why many tropical places have no match

Places near the equator are far more likely to have no present-day match. Around 2100 under SSP2-4.5 (models in the likely range):

This is mainly physics, not a gap in the data or an effect of wealth: rich tropical and desert cities such as Singapore, Dubai and Darwin have no match either. Two things combine.

• There is nowhere warmer to point to. A mid-latitude city's future climate usually exists today somewhere closer to the equator. The hottest places on Earth have no hotter place to match, so their future climates are new to the planet. Williams et al. (2007) found such novel climates concentrated in the tropics and subtropics.
• Tropical climates vary little from year to year. σ measures change against a place's own normal swings, so a given warming stands out far more where years are steady. This is why tropical climates are expected to leave their historical range first (Mora et al. 2013). It also means the change is large relative to what local ecosystems and people are adapted to.

Data play a smaller part. Where weather stations are sparse, TerraClimate can understate year-to-year variability, which inflates σ; the variability floor (above) limits this, but some tropical results are more uncertain than mid-latitude ones.

The pattern also has a moral dimension. The places facing climates with no present-day equivalent are largely those that have contributed least to the emissions causing the change (King & Harrington 2018).

Checks and results

• Dataset consistency. For North American places, the present climate from TerraClimate and from AdaptWest differ by a median of 0.00σ (90th percentile 0.72σ).
• Self-check. Each world city's present climate compared with the global grid has a median best distance of 0.00σ. The exceptions are 53 small islands and narrow coastal strips (for example Malé, Funafuti, Majuro, Aden, Dakar and Saipan) whose climate no 0.5° cell resembles; their arrows should be read loosely.
• Model processing. Change factors from the global model fields reproduce the regional North American extraction at 5 places × 24 models × 4 scenarios, with a 99th-percentile difference of 0.01 °C (storage rounding).
• Browser against pipeline. The page and the Python batch pick the same best cell for North American places and for all 12 world cities tested. One known near-tie (Raleigh, 0.050σ against 0.053σ for two neighbouring cells) resolves differently because of rounding.

Most tropical lowland cities have no match by 2100 under higher emissions, because their projected heat has no present-day counterpart anywhere, consistent with earlier global analyses of novel climates (Williams et al. 2007).

Cross-check against independent datasets

Every place's present climate comes from one dataset (AdaptWest in North America, TerraClimate elsewhere), so a fault in that dataset would go unnoticed. As an independent second opinion the build computes the same sixteen seasonal measures for 1991–2020 from ERA5 (Hersbach et al. 2020), a reanalysis that shares no observations-to-grid processing with either of them. Daily highs and lows, precipitation and dewpoint come from the 0.25° ERA5 daily fields distributed by Google (ARCO-ERA5, aggregated to days by WeatherBench 2; Carver & Merose 2023; Rasp et al. 2024). The year-to-year variability is the same detrended standard deviation used everywhere else.

Like with like. A 0.25° ERA5 cell is about 28 km across and describes the average of its land and sea; the place's baseline describes a 10 km neighbourhood (1 km in North America). Compared directly, small islands and coasts disagree by many σ for no reason other than that difference in footprint: the first full build put a third of all places over 3.5σ that way. The check therefore compares the two datasets over the same area. Each place gets its own ERA5 land cell: the 0.25° cell nearest to it whose land fraction in ERA5's land-sea mask is at least one half, within 50 km. The place's baseline is then carried to that cell's footprint: TerraClimate is averaged over its land pixels inside the cell (the 36 pixels of 1/24° that make up one ERA5 cell), and for North American places the AdaptWest baseline is shifted by TerraClimate's difference between that cell average and the 10 km neighbourhood (temperatures and dewpoints by the difference, precipitation by the ratio). Places with no ERA5 land cell within 50 km (atolls and small islands that ERA5 treats as sea) are not checked against ERA5 at all; they rely on the other sources below. The land-cell rule and distance are in config.toml under [era5] (match = "point" restores the original point comparison).

Systematic offsets. ERA5 differs from the gridded observational datasets in the same direction almost everywhere: nights are about 1.5 °C warmer, days about 0.5 °C cooler, and precipitation differs by season. Left in, these offsets alone would put most places several σ apart. The build therefore estimates the typical ERA5 offset for each measure (the median over places, separately for the AdaptWest and TerraClimate places; precipitation as a log ratio) and removes it, so that the agreement below reflects how far a place departs from the typical offset. The offsets themselves are reported in the table.

Baseline agreement is the σ distance between a place's baseline (carried to its ERA5 cell as above) and the offset-adjusted ERA5 baseline of that cell, measured with the place's own year-to-year variability and the fourteen matched measures, exactly as for a match. Below 2σ the two datasets are treated as agreeing and the page says so. Between 2σ and 3.5σ the page notes that they differ somewhat. Above 3.5σ the confidence badge is lowered by one level (high to medium, medium or low to low) and the panel says why. Thresholds are in config.toml under [era5]; they are judgement calls, not a statistical test.

What a disagreement means. It does not say which dataset is right. ERA5 is known to be too cold in parts of the high latitudes and mountains, too wet in many tropical and mountain regions, and its terrain is smoother than its grid, so in steep mountains even the cell averages can differ by a degree or more (no lapse-rate correction is applied). Days are UTC days, which shifts daily extremes slightly away from local days. ERA5 dewpoint is blended over water and ice at low temperatures, whereas the other data use the over-water formula. The check therefore flags places where the answer rests on a single, unconfirmed source or on terrain finer than the grids.

In this release the median disagreement is 1.22σ across 2,331 places (90th percentile 9.4σ); 42% of places are above 2σ and 32% above 3.5σ. 55 places have no ERA5 land cell within reach and are not checked against ERA5.

Two more sources. ERA5 is not the only opinion. CHIRPS v2.0 (Funk et al. 2015) blends satellite infrared rainfall estimates with rain-gauge records at 0.05° between 50°S and 50°N; the build takes its monthly totals for 1991–2020 at every place (nearest valid cell) and forms the four seasonal precipitation totals exactly as for the baseline. It checks precipitation only, and precipitation is where the main datasets are least sure, so it is the most useful second opinion there. CHELSA v2.1 (Karger et al. 2017) is a 30 arc-second downscaling of reanalysis with terrain and cloud corrections; the build uses its monthly highs, lows and precipitation. Its published normals cover 1981–2010, not 1991–2020, so the difference between the periods (about half a degree of warming, and some regional shifts in rainfall) appears as an offset and is removed together with the dataset's own bias.

Combined check. For every source the build computes the same σ distance as above, using only the measures that source covers: the place's variability model is restricted to those measures (its covariance is the matching model's covariance for them) and the source's typical offset is removed for each measure. CHELSA is shown in the table but does not affect the badge: it is a downscaling of ERA5, so the two share ERA5’s biases and cannot confirm each other (in the first build with both, they “agreed” on a poor match at a quarter of places, nearly all coastal). A place is flagged only when two independent sources, ERA5 and CHIRPS, both disagree with its baseline: its agreement is the second-largest of these distances, and the page names both sources. One source alone never lowers the badge. This rule was adopted after the first full rebuild: even compared cell to cell, ERA5 keeps a real warm-night bias of 2–3 °C at many coasts and small islands (Honolulu, Kingston, Dakar), which would otherwise have lowered confidence at about a third of places for a reason unrelated to the match. The thresholds are the ones above: over 2σ the panel says the sources differ somewhat; over 3.5σ confidence drops one level. Because each source is tested only where it has data, the check is not identical from place to place: CHIRPS is absent poleward of 50° and covers precipitation only, so outside the tropics and subtropics few places can be flagged; the shares in the table show how many were.

Year-to-year variability of precipitation is also compared: the detrended standard deviation of log seasonal totals in CHIRPS, divided by the value in each place's own variability model, has a median of 0.82 (winter), 0.85 (spring), 0.79 (summer) and 0.84 (autumn) across places. Values near 1 mean the two agree on how much the rain varies; this comparison is reported only and does not change any confidence level.

Hindcast test

Can the method find, from a 1961–1990 starting point, the place whose climate a location came to have in 1991–2020? With ERA5 we know what happened, so the observed change can stand in for a projection. For each of about 100 places (half chosen to be as different as possible in climate, half random) we take the 1961–1990 climate, add the change of the surrounding 5° box as a coarse model would (temperature added; precipitation and humidity ratios clipped, as in the real method), and search the 1961–1990 climates on a 0.5° land grid using the place's own 1961–1990 variability. We then measure how far the climate that actually occurred was from (a) the no-change assumption, (b) the projection, and (c) the analog the method chose, and compare with (d) the best analog any search could have found and (e) a naive alternative, the cell on the same meridian, equatorward of the place, with the closest annual mean temperature.

100 places sampled from 2,377 (half farthest-point in climate space, half random). Median warming of the annual mean between the windows was 0.76 °C.

In the method's own units, the actual 1991–2020 climate lay a median of 0.04σ from the predicted analog (80% of places under 1σ, 88% under 2σ, 90th percentile 2.2σ). The predicted analog was within 500 km of the best possible analog for 97% of places (median 0 km apart), and the projection was closer to the truth than “nothing changes” at 94% of places. Predicted analogs lay a median of 109 km from the place, equatorward for 53% of them. Skill against the latitude-shift baseline (1 minus the ratio of median distances) was 56%, and against the place's own 0.5° cell 36%. The predicted analog was only 12% closer to the truth than the place's own 1961–1990 climate (Table H, “No change”), which is what a small change against large year-to-year swings implies: the test shows that the search finds essentially the best place there is, not that moving to it is warranted.

Limitations. Between these two periods the world warmed by only about 0.5 °C, against 2–4 °C in the scenarios, so most places moved less than their own year-to-year swings and the test is weak: it can show that nothing is badly wrong, not that the method works at the much larger changes in the projections. The change is added from a 5° box of the same dataset, which is closer to the truth than a climate model's own error would be. ERA5 is not homogeneous over 1961–2020 (the satellite era brings observing-system shifts, and early decades are less constrained in the tropics and the Southern Hemisphere), so part of the apparent change may not be climate. A 0.25° reanalysis under-represents mountains and coasts, and the same dataset supplies both the projection and the truth, so its biases cancel in ways they will not for real projections. The variability model is the one built from 1961–1990 only, thirty years for fourteen measures, as in the real method.

Sensitivity to model resolution

Climate models run on grids of roughly 100 to 250 km, so a place's projected change is really the change of one large grid cell. Three independent downscaled versions of the same CMIP6 changes exist, and once a year we rebuild each place's future climate a second time with each of them (temperature added, precipitation scaled, humidity scaled where the source has it, ratios limited to the same ranges as the main analysis) and compare it with the projection used on the page:

• NASA NEX-GDDP-CMIP6 (Thrasher et al. 2022), 25 km, bias-corrected, the same monthly changes as the extreme-day counts (also humidity). Since the main projection now uses these changes wherever they exist (see Projected climate), comparing them with the page would compare a source with itself; the check instead compares the page's projection with the same models' native-grid CMIP6 changes, which keeps measuring how much the answer depends on resolution. Before the NEX-GDDP changes are available it compares NEX-GDDP with the native-grid projection directly.
• WorldClim 2.1 CMIP6 (Fick & Hijmans 2017), global, statistically downscaled monthly minimum and maximum temperature and precipitation for many models. We use the change from its own 1970–2000 baseline to 2041–2060 and 2081–2100 (temperature difference, precipitation ratio).
• AdaptWest downscaled CMIP6 (Wang et al. 2016 method; AdaptWest Project 2022), 1 km, North America only: the project's 8-model ensemble mean for 2041–2060 and 2071–2100 (its monthly files use 2071–2100, not 2081–2100), as a change from the AdaptWest 1991–2020 normals. Because it is one ensemble product it is compared with the mean of this analysis's likely-sensitivity models rather than model by model.

The comparison uses the models a source shares with this analysis; where fewer than six are shared (for example an AdaptWest multi-model ensemble) the source's own multi-model mean is compared with the mean of this analysis's likely-sensitivity models. A place is marked moderate if the two projected climates differ by 0.5σ or more in its own year-to-year variability (or the best match would move more than 500 km), and high if they differ by 1σ or more (or 0.5σ or more with a best match that moves more than 500 km). The worst case over both periods and both scenarios counts for each source, and a place is flagged only when two sources agree: WorldClim measures its changes from a 1970–2000 baseline, so on its own it would flag about 40% of places for two decades of warming that are not a resolution effect at all (a place with a single source is judged by that source). Only “high” is shown on the page, as one line under the answer that names the source(s) that flagged it; the full list, with each source's own flag, is in data/nexcheck.json.

Limits. This tests one thing: whether the answer would change with a finer version of the same models' changes. It is not a check against observations and not a validation of the method. The NEX-GDDP / native-grid comparison covers the scenarios processed for both; WorldClim and AdaptWest cover SSP2-4.5 and SSP5-8.5 only, and the worst case over all of them is shown for every scenario. NEX-GDDP has one simulation per model, while the native-grid changes average up to three simulations per model, so part of a difference can be internal variability rather than resolution. The sources use different baselines (WorldClim 1970–2000, AdaptWest 1991–2020, this analysis 1995–2014): a WorldClim change therefore includes some warming that has already happened (a few tenths of a degree), which pushes its difference slightly above zero; humidity is held at constant relative humidity for WorldClim and AdaptWest. Places without data in a source's grid are not tested by it. Taking the worst case over sources makes the flag deliberately cautious. The thresholds are judgement calls and are set in config.toml.

Fick, S. E., & 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. https://doi.org/10.1002/joc.5086. Wang, T., Hamann, A., Spittlehouse, D., & Carroll, C. (2016). Locally downscaled and spatially customizable climate data for historical and future periods for North America. PLoS ONE, 11(6), e0156720. https://doi.org/10.1371/journal.pone.0156720. AdaptWest Project (2022). Gridded current and projected climate data for North America at 1 km resolution, interpolated using the ClimateNA v7.30 software. https://adaptwest.databasin.org

Climate type, hardiness zone and growing season

These are computed from monthly normals by the same rules for a place today, the same place in the future (the ensemble-mean monthly climate) and every cell of both comparison grids, so the three columns in the comparison are always like with like.

Climate type follows the Köppen–Geiger criteria of Beck et al. (2018), which use a 0 °C boundary between temperate and cold climates (Peel et al. 2007), with summer taken as the warmer half of the year.

Hardiness zone is estimated, and is not the official USDA Plant Hardiness Zone Map. The quantity behind the zones, the average of each year's coldest temperature, is estimated from the coldest month's mean low, the annual temperature range and annual precipitation by a regression fitted to the official 1991–2020 grid from the PRISM Climate Group and USDA Agricultural Research Service (2023) across the contiguous United States (R² = 0.96; median error 1.0 °C). The estimate falls in the same half-zone as the official grid for 56% of 15 km cells and within one half-zone for 95%. Outside the contiguous United States the same formula is applied without local calibration.

Growing season is the estimated number of days between the last spring frost and the first fall frost. It sums, over the months, the days in each month times a normal probability that the month is frost-free given its mean low, fitted to ClimateNA's frost-free period over North America (Wang et al. 2016; R² = 0.96; median error 7 days).

Snowfall is not estimated. Monthly averages cannot separate rain from snow reliably: in months whose average temperature is above freezing, cold spells still bring snow. Even ClimateNA's precipitation-as-snow puts Denver's snow near 3% of its precipitation, far below what falls there.

Already happening

For each place, the average of the most recent ten complete years in TerraClimate (2016–2025) is compared with its 1991–2020 average from the same dataset, so the difference reflects climate rather than differences between datasets. The page reports the change in annual mean temperature and, where the projected change is large enough to compare with, the share of the warming projected by around 2050 under SSP2-4.5 that has already occurred. Ten years is short, so part of any difference is natural variability.

Sea level

For places within 50 km of a projection site, the page shows relative sea-level rise from the IPCC Sixth Assessment Report (Fox-Kemper et al. 2021), using the medium-confidence projections at tide gauges and on a 1° coastal grid (Garner et al. 2021). Values are the median and the likely range (17th–83rd percentiles) for 2050 and 2100 relative to 1995–2014, and include vertical land motion such as subsidence. 1013 places have projections. Low-likelihood, high-impact outcomes involving rapid ice-sheet loss are not included and could add substantially by 2100.

Extreme days

The “Extreme days” row counts, for each place, the days a year that are dangerously hot, humid or cold, and how dry and how wet the extremes get. It comes from a separate dataset and a separate yearly job from the rest of the page: NASA's NEX-GDDP-CMIP6 (Thrasher et al. 2022), which gives daily maximum and minimum temperature, precipitation and relative humidity on a 0.25° (about 25 km) grid, bias-corrected against observations, for historical years and for the emissions scenarios. The daily values needed for these counts are not in the monthly CMIP6 data used everywhere else.

Indicators. Days with a daily high of 35 °C (95 °F) or more, and 30 °C (86 °F) or more; tropical nights (daily low of 20 °C or more); frost days (daily low below 0 °C); hot-humid days (wet-bulb temperature of 28 °C or more); the longest run of days with under 1 mm of precipitation in a year; and the wettest single day. Each is computed for every sampled year and averaged over a window: 1991–2020 for “Now” (historical years to 2014, SSP2-4.5 for 2015–2020), and 2041–2060 and 2081–2100 for the future. Counts are scaled to a 365-day year so that models with 360-day calendars compare.

Wet-bulb temperature is calculated with the empirical formula of Stull (2011) from the daily high temperature and relative humidity. The dataset provides only the daily mean humidity, which is much higher than the humidity in the afternoon when the high occurs, so the humidity at the time of the high is estimated by holding the vapour pressure (dewpoint) fixed while temperature rises from the daily mean, taken as the average of the high and low, to the high. This is an approximation and the threshold is a rough marker of very hard conditions, not a measured heat-stress limit; wet-bulb temperatures near 35 °C are the physiological limit for healthy people, and 28 °C is already dangerous for hard work.

From grid to place. Each place takes the values of the nearest 0.25° cell that has data (up to three cells away for coastal places). A 25 km cell is a spatial average, so hot spots, city heat islands and mountain valleys are not resolved, and counts of extreme days at one site can differ from the cell value. Ensembles: the median across the same likely-sensitivity models used elsewhere, where NEX-GDDP publishes them (13 of them; the exact list is in the data file; BCC-CSM2-MR has no humidity, so it is left out of the hot-humid count), with the range (lowest to highest) shown beneath. To keep the download affordable, every second year of each window is sampled; this adds a small amount of noise to counts of rare events. The same job also processes the other models NEX-GDDP publishes, for the projected changes (see Projected climate).

Scenarios. All four scenarios are processed. If a scenario is missing from the data file, the page shows the nearest processed one and says so. Places with no land cell in the grid (some small islands) show nothing. The counts are model output, not observations, and are not forecasts for a particular year.

Thrasher, B., Wang, W., Michaelis, A., Melton, F., Lee, T., & Nemani, R. (2022). NASA Global Daily Downscaled Projections, CMIP6. Scientific Data, 9, 262. https://doi.org/10.1038/s41597-022-01393-4. Stull, R. (2011). Wet-bulb temperature from relative humidity and air temperature. Journal of Applied Meteorology and Climatology, 50(11), 2267–2269. https://doi.org/10.1175/JAMC-D-11-0143.1

What the maps show

For each organism the page shows the places where the model finds the climate suitable now (1991–2020), in the future (2041–2060 or 2081–2100, under the moderate SSP2-4.5 or the high-emissions SSP5-8.5 scenario), and the change: land that is suitable now but not later (lost, hatched orange), suitable in both (kept, grey) or suitable only later (gained, dotted blue). Colour is never the only cue: lost is hatched and gained is dotted. Suitability is drawn in three classes, not as a smooth probability.

The summary gives the area with suitable climate, its change, and how far the centre of that area moves. The centre shift is the movement of an area-weighted average position; it can move even when no single place does.

Data

• Where the organism has been recorded: GBIF occurrence records with a CC0, CC-BY or CC BY-NC (non-commercial) licence, cleaned and thinned to one record per 0.125° cell (about 14 km) for counting; each organism cites the download DOI (table below). Records with other licences are not used. Organisms with fewer than 100 such cells are not modelled.
• Where the organism has actually been recorded: models learn from every recorded location, wherever it is, including places a species has spread to. We do not restrict records to a “native range”, because published native ranges are often out of date and the climate a species lives in today is what the records show. Records tagged as introduced, naturalised, invasive, managed or cultivated are still removed, which keeps out gardens, farms, zoos and stocked populations. Migratory birds are the exception: their breeding-season fits still use a breeding range. Organisms fitted this way say “Every recorded location” on their page. About a third keep the earlier fit made inside a published native range (their records barely changed when the restriction was dropped, or the all-records fit did not meet the publication rule); their page says “Records inside the organism’s known native range”.
• Present climate: TerraClimate, 1991–2020, on a grid of 1/24° (about 4.6 km). Ten candidate measures: annual mean temperature, temperature range and seasonality, warmest-month and coldest-month temperature, annual precipitation, precipitation seasonality, driest-quarter precipitation, growing degree-days, climatic water deficit and actual evapotranspiration.
• Future climate: the same present-day climate plus the change in each of 14 climate models from NASA’s NEX-GDDP-CMIP6 for SSP2-4.5 and SSP5-8.5.
• Checks against independent data: the North American Breeding Bird Survey (for birds), and, for trees, comparison with the US Forest Service’s published tree projections (RDS-2024-0020).

The model

For each organism, a gradient-boosted tree model (LightGBM) learns how the recorded places differ in climate from a sample of background places. The background is drawn where observers of the same group of organisms (for example, all birds) have recorded things, which corrects part of the bias towards well-visited places. At most five climate measures are kept, chosen in an order based on the organism’s ecology, and a measure is dropped when it is strongly correlated (above 0.7) with one already kept. A second, simpler model (a smooth additive model) is run as a check; when the two disagree strongly about the area change the organism is marked low confidence.

“Suitable” means the model score is above a threshold set so that 95% of held-out records fall inside. The model is projected under each of 14 climate models and the median is shown; the share of models that agree can be shaded on the map (dotted where fewer than 3 in 4 agree).

Migratory birds use the breeding-season fit: only records from the breeding months (May to July in the north) and a breeding-season native range. Their pages say so. Resident birds, and animals such as the monarch, use all records.

Crops are shown where the climate is like where the crop is recorded, in the exact wording on each page. Unless the records are verified as cultivated (for the wine grape they are not), this is not a statement about where the crop can be grown.

Dispersal: three bounds

An organism cannot occupy all newly suitable land. Each future map can be read with three bounds: unlimited (any suitable area counts), limited (new area counts only within a reach of today’s range) and none (only the part of today’s range that stays suitable). The limited reach comes from a rule of thumb by group: about 15 km per decade for trees, 10 for crops and reptiles, 40 for insects and ticks, 50 for mammals and 100 for birds. These distances are uncalibrated and probably uncertain by a factor of about 10. The truth for a given organism usually lies between the limited and unlimited bounds, or below none if the range retreats faster than it can shift.

How much to trust a map: two checks

Every organism has two plain checks, because “is today’s map right?” and “can the future map be trusted?” are different questions. The organism’s page sums them up in one grade, with both checks one tap away (“Why?”): good when today’s map was checked against an independent survey and the future map passed the range-change test; limited when either check raised a concern; fair otherwise, which is the normal case. When an organism’s future numbers are withheld, the grade says so instead.

Today’s map: checked

The model’s map of today’s suitable climate was checked against an independent survey it had never seen (so far, the Breeding Bird Survey for birds and US Forest Inventory plots for trees: discrimination AUC at least 0.7 and best-threshold TSS at least 0.4). Internally this is Tier A.

Today’s map: fits the records

Checked by spatial-block cross-validation: the model is trained on some regions and tested on others. This is a fair test of skill in new places, but no independent survey exists for most organisms, so it is not checked against one. Internally this is Tier B. It is the normal case, not a warning.

Today’s map: looser fit

As above, but the fit is weaker than we would like (for example, discrimination below 0.7 or much of the suitable area far from any record). The reason is on the organism’s page.

Not shown

Organisms below the skill or range-check gates are not published (Tier C).

The page also states the cross-validation kind: “presence against background” compares records with a sample of background places, which is a weaker test than “presence and absence”.

The second check, Future map, is explained next. The future map always shows where climate may suit the organism, never a forecast of where it will live.

Range shifts: pass, fail, untested

The Future map check records whether we could show that models like this one can predict real range changes. We fitted the model to records from an early period, asked it where the organism should be found in a later period, and compared that with surveys of what actually happened.

• Fail (all birds tested): the model locates today’s ranges well, but for the real movements between the two periods the direction of change was no better than chance. In plain words: the model describes where the climate suits a bird today, but we could not show it knows where the bird will go. Treat the future map as a map of climate, and the Today’s map check as a statement about today only.
• Untested (trees, insects and others): there was not enough survey data on real change to run the test (for example, only 5 trees against a minimum of 20). Adult trees also lag the climate, so absence of change is expected even when the climate has moved. This is not a pass.
• Pass: the model reproduced observed change in the test. No organism has passed yet.

Withheld numbers

When more than 15% of the area projected to be suitable lies in a climate that nothing in the training data resembles (a “novel climate”, found with the MESS rule), the model would be extrapolating beyond what it has seen. The page then shows the numbers as withheld, with the reason and the share of novel climate. The map is still drawn, with those places marked by diagonal lines, so you can see where the problem lies.

Limitations

Climate twins (places)

• No extremes. Heat waves, heat index, cold snaps, hurricanes, wildfire and snowfall are not included. Two places can match on these fourteen measures and still differ in ways that matter to people.
• Estimated features. Hardiness zones and growing seasons are statistical estimates from monthly data, calibrated in North America only.
• Coarse model changes. Changes come from model grid cells roughly 50–300 km across and are added to local climate, so changes in fine-scale patterns, along mountains and coasts for example, are not captured.
• Grid resolution. The comparison grids (15 km and 0.5°) smooth mountain and coastal climates. A mountain or coastal town's best match can be somewhat worse than it would be at finer resolution.
• One simulation per model. Natural variability within each model is not sampled, so part of the spread between models is internal variability rather than model difference.
• Present-day pool. Matches are found among 1991–2020 climates, which already include some warming. Present climates elsewhere will also change, which the map does not show.
• Mixed variability sources. Places outside the contiguous United States use TerraClimate for year-to-year variability, which is coarser than PRISM.
• Scenarios are not forecasts. SSP5-8.5 in particular is now widely considered implausible for this century. Each arrow is a way to picture a scenario, not a prediction for a specific town.

Organism maps

• Dispersal. See above: the three bounds are rules of thumb, not predictions. Barriers, habitat loss and human movement are not modelled.
• Non-climatic limits. Soils, land use, fire, water bodies, competitors, predators, pathogens, pollinators and host species are not in the model. A place can be climatically suitable and have none of what the organism needs. The native-range check catches only some of this.
• Migratory handling. Breeding-season fits describe the breeding range only. Winter ranges, stopover places and migration routes are not modelled. For the barn swallow the breeding-range mask has no South America, so its wintering range there is absent. Month information is missing for some records and they are left out of breeding-season fits. The background still reflects year-round observer effort.
• Precipitation sensitivity. The change in rainfall in the future climate data is less well validated than the change in temperature. In a test on two trees, setting the rainfall change to zero changed the 2081–2100 area change by up to 17 percentage points, and the centre shift by only 10 to 50 km. Read area changes for organisms that depend on rainfall or water deficit with that in mind; the direction and size of the centre shift are more robust.
• Agreement with other models. For trees, the size and direction of the centre shift are similar to the US Forest Service’s projections, but not identical. For quaking aspen our shift is about half the size and the area change is smaller (about +5% against +35 to +59%). These are two models agreeing or disagreeing, not tests against what happened.
• Resolution. Cells are about 4.6 km. Mountains, coasts and small ranges are coarsely represented, and narrow-range organisms are less reliable.
• Records. CC0, CC-BY and CC BY-NC records are used. CC BY-NC records (a licence many iNaturalist observers choose) are used because this site is educational and non-commercial; if that ever changed they would have to be removed. Records reflect where people look: well-visited places have far more than remote ones.
• Thresholds. Several limits (novelty 15%, range-check and agreement limits, the dispersal distances) are first values that have not been calibrated.

Data and attribution

Climate twins (places)

Acknowledgements. We acknowledge the World Climate Research Programme, which, through its Working Group on Coupled Modelling, coordinated and promoted CMIP6. 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 CMIP6 and ESGF. Dataset citations for individual CMIP6 models are available from the CMIP6 Citation Service. We also thank the Pangeo project and Google Cloud Public Datasets for the analysis-ready CMIP6 archive, the AdaptWest Project, the Climatology Lab at the University of California, Merced, and the PRISM Climate Group at Oregon State University.

Organism maps

• GBIF.org occurrence downloads (CC0, CC-BY and CC BY-NC records; the CC BY-NC records are used for this non-commercial educational site only), each cited by DOI in the table below. Target-group sampling grids: birds 10.15468/dl.spdj4n, mammals 10.15468/dl.7xq6mb, insects and arachnids 10.15468/dl.at3mx7, plants 10.15468/dl.qqzta8 (GBIF.org, 10 October 2026). Thanks to the many data publishers and observers behind these records.
• TerraClimate (Abatzoglou et al. 2018, Scientific Data 5:170191, doi:10.1038/sdata.2017.191).
• NEX-GDDP-CMIP6 (NASA; Thrasher et al. 2022, Scientific Data 9:262, doi:10.1038/s41597-022-01393-4), and the modelling groups that produced the CMIP6 simulations.
• World Checklist of Vascular Plants (WCVP), Royal Botanic Gardens, Kew, native ranges of plants (via Plants of the World Online; CC BY 3.0). No DOI is cited because the one we first had did not resolve.
• North American Breeding Bird Survey, USGS, 2025 release (doi:10.5066/P14SNUV4), used to check the bird models, with thanks to the volunteers who run the routes.
• US Forest Service tree projections used for comparison: doi:10.2737/RDS-2024-0020.
• Maps are drawn with MapLibre GL; the base map is from OpenFreeMap and OpenStreetMap contributors.

Organisms are shown as an outline of the climate model, with no photographs. Code and data tables are in the project repository.

GBIF download DOI for each organism

Software

Data processing uses Python with NumPy, SciPy, pandas, xarray (reading the CMIP6 Zarr stores), rasterio, pyproj and Shapely. The interactive map uses MapLibre GL JS 5.24 with OpenFreeMap vector tiles; the heat layer, distances and arrows are computed in the browser. The North American grid uses an ellipsoidal Lambert azimuthal equal-area projection implemented in JavaScript and checked against pyproj to better than 10⁻⁷ m.

Updates and versions

The data are rebuilt from the original sources each year by a public, automated pipeline (source code in the project repository). Each rebuild is checked before release: counts of places, grid cells and models; a self-check that today's climate at each place finds itself (current median 0.00σ); and a comparison with the previous release, flagging for review if more than 10% of best matches move over 500 km or the median change in σ exceeds 0.25. A browser test loads the page with the new data. Nothing is published until the rebuild is reviewed. A monthly check watches for a new TerraClimate year, corrected or withdrawn CMIP6 datasets and the arrival of CMIP7 projections. This page describes method version 10.0, data version 2026-10-01, using 24 models.

Related tools

Other maps answer neighbouring questions, and are worth exploring alongside this one: the IPCC WGI Interactive Atlas (regional and global climate projections), Probable Futures (heat, drought and precipitation at +0.5 to +3 °C), the Climate Impact Lab map (heat, mortality and economic effects), ProPublica’s climate migration map and Future Urban Climates by Fitzpatrick & Dunn, the work this Places map is most closely modelled on. These are independent projects with their own methods, and none of them is a source for this site.

References

1. Alduchov, O. A., & Eskridge, R. E. (1996). Improved Magnus form approximation of saturation vapor pressure. Journal of Applied Meteorology, 35(4), 601–609. https://doi.org/10.1175/1520-0450(1996)035<0601:IMFAOS>2.0.CO;2
2. Abatzoglou, J. T., Dobrowski, S. Z., Parks, S. A., & Hegewisch, K. C. (2018). TerraClimate, a high-resolution global dataset of monthly climate and climatic water balance from 1958–2015. Scientific Data, 5, 170191. https://doi.org/10.1038/sdata.2017.191
3. Abernathey, R. P., Augspurger, T., Banihirwe, A., Blackmon-Luca, C. C., Crone, T. J., Gentemann, C. L., Hamman, J. J., Henderson, N., Lepore, C., McCaie, T. A., Robinson, N. H., & Signell, R. P. (2021). Cloud-native repositories for big scientific data. Computing in Science & Engineering, 23(2), 26–35. https://doi.org/10.1109/MCSE.2021.3059437
4. AdaptWest Project. (2022). Gridded current and projected climate data for North America at 1km resolution, generated using the ClimateNA v7.30 software (T. Wang et al., 2022). adaptwest.databasin.org
5. Beck, H. E., Zimmermann, N. E., McVicar, T. R., Vergopolan, N., Berg, A., & Wood, E. F. (2018). Present and future Köppen-Geiger climate classification maps at 1-km resolution. Scientific Data, 5, 180214. https://doi.org/10.1038/sdata.2018.214
6. Carver, R. W., & Merose, A. (2023). ARCO-ERA5: an analysis-ready cloud-optimized reanalysis dataset. 22nd Conference on Artificial Intelligence for Environmental Science, American Meteorological Society.
7. Daly, C., Halbleib, M., Smith, J. I., Gibson, W. P., Doggett, M. K., Taylor, G. H., Curtis, J., & Pasteris, P. P. (2008). Physiographically sensitive mapping of climatological temperature and precipitation across the conterminous United States. International Journal of Climatology, 28(15), 2031–2064. https://doi.org/10.1002/joc.1688
8. Eyring, V., Bony, S., Meehl, G. A., Senior, C. A., Stevens, B., Stouffer, R. J., & Taylor, K. E. (2016). Overview of the Coupled Model Intercomparison Project Phase 6 (CMIP6) experimental design and organization. Geoscientific Model Development, 9(5), 1937–1958. https://doi.org/10.5194/gmd-9-1937-2016
9. Fick, S. E., & 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. https://doi.org/10.1002/joc.5086
10. Fitzpatrick, M. C., & Dunn, R. R. (2019). Contemporary climatic analogs for 540 North American urban areas in the late 21st century. Nature Communications, 10, 614. https://doi.org/10.1038/s41467-019-08540-3
11. Forster, P., Storelvmo, T., Armour, K., Collins, W., Dufresne, J.-L., Frame, D., Lunt, D. J., Mauritsen, T., Palmer, M. D., Watanabe, M., Wild, M., & Zhang, H. (2021). The Earth's energy budget, climate feedbacks, and climate sensitivity. In V. Masson-Delmotte et al. (Eds.), Climate Change 2021: The Physical Science Basis. Contribution of Working Group I to the Sixth Assessment Report of the Intergovernmental Panel on Climate Change (pp. 923–1054). Cambridge University Press. https://doi.org/10.1017/9781009157896.009
12. Fox-Kemper, B., Hewitt, H. T., Xiao, C., et al. (2021). Ocean, cryosphere and sea level change. In V. Masson-Delmotte et al. (Eds.), Climate Change 2021: The Physical Science Basis (pp. 1211–1362). Cambridge University Press. https://doi.org/10.1017/9781009157896.011
13. Funk, C., Peterson, P., Landsfeld, M., Pedreros, D., Verdin, J., Shukla, S., Husak, G., Rowland, J., Harrison, L., Hoell, A., & Michaelsen, J. (2015). The climate hazards infrared precipitation with stations—a new environmental record for monitoring extremes. Scientific Data, 2, 150066. https://doi.org/10.1038/sdata.2015.66
14. Garner, G. G., Hermans, T., Kopp, R. E., Slangen, A. B. A., Edwards, T. L., Levermann, A., et al. (2021). IPCC AR6 sea level projections (Version 20210809) [Data set]. Zenodo. https://doi.org/10.5281/zenodo.5914709
15. Hausfather, Z., Marvel, K., Schmidt, G. A., Nielsen-Gammon, J. W., & Zelinka, M. (2022). Climate simulations: recognize the ‘hot model’ problem. Nature, 605, 26–29. https://doi.org/10.1038/d41586-022-01192-2
16. Hersbach, H., Bell, B., Berrisford, P., et al. (2020). The ERA5 global reanalysis. Quarterly Journal of the Royal Meteorological Society, 146(730), 1999–2049. https://doi.org/10.1002/qj.3803
17. Karger, D. N., Conrad, O., Böhner, J., Kawohl, T., Kreft, H., Soria-Auza, R. W., Zimmermann, N. E., Linder, H. P., & Kessler, M. (2017). Climatologies at high resolution for the earth's land surface areas. Scientific Data, 4, 170122. https://doi.org/10.1038/sdata.2017.122
18. Lee, J.-Y., Marotzke, J., Bala, G., Cao, L., Corti, S., Dunne, J. P., et al. (2021). Future global climate: scenario-based projections and near-term information. In V. Masson-Delmotte et al. (Eds.), Climate Change 2021: The Physical Science Basis (pp. 553–672). Cambridge University Press. doi:10.1017/9781009157896.006
19. Ledoit, O., & Wolf, M. (2004). A well-conditioned estimator for large-dimensional covariance matrices. Journal of Multivariate Analysis, 88(2), 365–411. https://doi.org/10.1016/S0047-259X(03)00096-4
20. King, A. D., & Harrington, L. J. (2018). The inequality of climate change from 1.5 to 2°C of global warming. Geophysical Research Letters, 45(10), 5030–5033. https://doi.org/10.1029/2018GL078430
21. Mahony, C. R., Cannon, A. J., Wang, T., & Aitken, S. N. (2017). A closer look at novel climates: new methods and insights at continental to landscape scales. Global Change Biology, 23(9), 3934–3955. https://doi.org/10.1111/gcb.13645
22. Maraun, D., & Widmann, M. (2018). Statistical Downscaling and Bias Correction for Climate Research. Cambridge University Press. https://doi.org/10.1017/9781107588783
23. Mora, C., Frazier, A. G., Longman, R. J., Dacks, R. S., Walton, M. M., Tong, E. J., et al. (2013). The projected timing of climate departure from recent variability. Nature, 502(7470), 183–187. https://doi.org/10.1038/nature12540
24. O'Neill, B. C., Tebaldi, C., van Vuuren, D. P., Eyring, V., Friedlingstein, P., Hurtt, G., Knutti, R., Kriegler, E., Lamarque, J.-F., Lowe, J., Meehl, G. A., Moss, R., Riahi, K., & Sanderson, B. M. (2016). The Scenario Model Intercomparison Project (ScenarioMIP) for CMIP6. Geoscientific Model Development, 9(9), 3461–3482. https://doi.org/10.5194/gmd-9-3461-2016
25. Peel, M. C., Finlayson, B. L., & McMahon, T. A. (2007). Updated world map of the Köppen-Geiger climate classification. Hydrology and Earth System Sciences, 11(5), 1633–1644. https://doi.org/10.5194/hess-11-1633-2007
26. PRISM Climate Group & USDA Agricultural Research Service. (2023). 2023 USDA Plant Hardiness Zone Map GIS data, 1991–2020 mean annual extreme minimum temperature. Oregon State University. https://prism.oregonstate.edu/phzm
27. PRISM Climate Group. (2026). PRISM monthly time series, 4 km, 1991–2020. Oregon State University. https://prism.oregonstate.edu
28. Rasp, S., Hoyer, S., Merose, A., et al. (2024). WeatherBench 2: a benchmark for the next generation of data-driven global weather models. Journal of Advances in Modeling Earth Systems, 16, e2023MS004019. https://doi.org/10.1029/2023MS004019
29. Riahi, K., van Vuuren, D. P., Kriegler, E., et al. (2017). The Shared Socioeconomic Pathways and their energy, land use, and greenhouse gas emissions implications: An overview. Global Environmental Change, 42, 153–168. https://doi.org/10.1016/j.gloenvcha.2016.05.009
30. Thrasher, B., Wang, W., Michaelis, A., Melton, F., Lee, T., & Nemani, R. (2022). NASA Global Daily Downscaled Projections, CMIP6. Scientific Data, 9, 262. https://doi.org/10.1038/s41597-022-01393-4
31. Wang, T., Hamann, A., Spittlehouse, D., & Carroll, C. (2016). Locally downscaled and spatially customizable climate data for historical and future periods for North America. PLoS ONE, 11(6), e0156720. https://doi.org/10.1371/journal.pone.0156720
32. Williams, J. W., & Jackson, S. T. (2007). Novel climates, no-analog communities, and ecological surprises. Frontiers in Ecology and the Environment, 5(9), 475–482. https://doi.org/10.1890/070037
33. Williams, J. W., Jackson, S. T., & Kutzbach, J. E. (2007). Projected distributions of novel and disappearing climates by 2100 AD. Proceedings of the National Academy of Sciences, 104(14), 5738–5742. https://doi.org/10.1073/pnas.0606292104