Humid tropical forests are global hotspots of terrestrial biodiversity1,2, playing critical roles in species conservation3, influencing climate regimes4 and terrestrial carbon cycles5. Yet the ecological integrity of global tropical forests is being diminished by clearing, selective logging and wildfires6, and by increasingly frequent extreme weather events, such as blowdowns and droughts, driven by climate change7. Moreover, novel climates—those with no recent historic analogues—are predicted to appear first in the tropics and subtropics8,9,10,11.

It is generally assumed that the impact of climate change on the forest subcanopy and understorey will be lower than elsewhere on Earth because temperature conditions below the canopy are buffered from temperature extremes, reducing the severity of warming impacts12,13,14. Beneath forest canopies, direct sunlight is strongly reduced and evapotranspirative cooling is increased, dampening temperature fluctuations compared with open habitats and resulting in cooler below-canopy maximum temperatures, warmer minimum temperatures, and lower seasonal and interannual variability12,15. However, the relative stability in temperature regimes through evolutionary history means that tropical forest organisms evolved under a narrower range of climate conditions than extratropical biota and can therefore tolerate a smaller margin of warming above their thermal optima16,17,18. Key questions, therefore, are whether the range of below-canopy temperatures currently experienced by tropical forest biota are novel in relation to historic climate, and how novelty varies between structurally intact and degraded tropical forests.

There is little understanding of how microclimates beneath the forest canopy—the conditions actually experienced by tropical forest organisms—are changing pantropically. Recent efforts to monitor within-forest temperatures14,19 have revealed that forests warm at a slower rate than non-forested areas, yet logging-induced canopy perforations increase understorey temperatures for up to 5 years relative to intact forest20. Although these are an important first step to quantifying below-canopy climate novelty, they provide only limited temporal coverage from a relatively small number of locations. Accordingly, mapping forest microclimates at a global scale has been identified as an important, yet unexplored, future research avenue13. Integrating a recently developed mechanistic microclimate model21 with empirical temperature measurements and satellite-derived land-cover data, we quantify hourly below-canopy temperature (5 cm above the ground) at 5 km grid resolution between 1990 and 2019 across forests in the humid tropics, including tropical rainforest and tropical moist deciduous forest (hereafter tropical forests6). We tackle two key objectives: (1) quantifying recent temperature novelty of forests pantropically compared with a historic baseline, to map those most at risk from warming and those that are currently providing climate refugia; and (2) comparing how overall temperature change across the last 30 years affects the degree of below-canopy temperature novelty in (i) undisturbed forest within ecologically unfragmented areas (defined as wilderness areas22), (ii) undisturbed forest in more fragmented landscapes and (iii) degraded forest.

Novelty is widespread across tropical forests

To accurately represent climate conditions experienced by the majority of forest-dwelling organisms, we modelled below-canopy, near-ground, hourly temperatures across the world’s tropical forest regions (approximately 9.3 million km2). Temperature is a primary constraint on species distributions and ecological function23,24,25, and so we derived estimates of seven temperature-based bioclimatic variables widely shown to affect species distributions26 (Methods). These variables represent annual trends, such as mean annual temperature and seasonality, and the incidence of extremes. Because tropical forests typically experience low temporal variability in temperature, small changes can result in climate conditions that lie entirely outside the normal range to which species are adapted. Consequently, we assume that temperature novelty is a better measure of climate vulnerability than the overall magnitude of temperature changes27. We derived an index of climate novelty from the fractional overlap in climate values between two time periods, representing the fraction of years in the most recent period (2005–2019) in which climate lies outside the range of conditions experienced in the historical baseline period (1990–2004); locations with high novelty are those with no historic climate analogue, making community disruption and species extinctions more likely28. We initially calculated novelty indexes for each temperature variable and a cumulative novelty index (the sum of novelty scores for all bioclimatic temperature variables, rescaled between 0 and 1) for undisturbed tropical forests6, that is, without any disturbances (degradation or deforestation) between 1982 and 2019.

Our results contradict the widely held belief that environments below forest canopies will be buffered from the worst impacts of warming12,14. Instead, our model suggests that, between 2005 and 2019, the majority of the world’s undisturbed tropical forests experienced climate conditions at least partially outside the range of baseline historic conditions (>0.25 fractional novelty in bioclimatic variables) and substantial portions of undisturbed tropical forests transitioned to almost entirely novel climatic averages (>0.80 fractional novelty in bioclimatic variables). For swaths of the Amazon and Congo basins, and Sundaland (insular Southeast Asia), annual climate conditions were almost entirely unprecedented relative to historic baselines (Fig. 1 and Extended Data Fig. 1). Highly novel temperature regimes have not only occurred in the lowlands, but also in tropical mountain systems including the tropical Andes and the Mentarang range in Northern Borneo, where there have been particularly high levels of warming over the last 30 years.

Fig. 1: Below-canopy temperature novelty across global, undisturbed-only tropical forests.
figure 1

ad, Novelty at 5 km gridded resolution (n = 317,809) mapped for mean annual temperature (a), mean diurnal temperature range (b), temperature seasonality (c) and cumulative temperature novelty (d) whereby novelty values for all temperature variables are summed and then rescaled between 0 and 1 to assess collective impact. The plots on the right show the distribution of novelty scores for each temperature variable and continental group (Central and South America, Africa, and Southeast Asia and Australia). Dotted lines indicate mean values. Ring plots, inset with maps, show the percentage of undisturbed forest for each continental group experiencing minimal (0.0–0.2), low (0.21–0.4), moderate (0.41–0.6), high (0.61–0.8) and extreme (0.81 to 1.0) novelty scores with colours scaled to match novelty map colours: see Supplementary Table 5 for a detailed breakdown of percentages.

Latin America experienced the highest overall cumulative temperature novelty, and the highest novelty in mean annual temperature, mean diurnal temperature range and temperature seasonality (Fig. 1). Here, 27% of undisturbed forest has recently experienced highly or entirely novel regimes in mean annual temperature (>0.61 fractional novelty) and 31% experienced highly or entirely novel mean diurnal temperature ranges. These mostly occurred in the northern tropical Andes and Pacific coast of South America—both of which are global biodiversity hotspots supporting high numbers of threatened and endemic species29. Additionally, 23% of undisturbed forests in Latin America, especially in the northern tropical Andes and the Brazilian Shield, have shifted to highly or entirely novel regimes in temperature seasonality; likely to be, in part, a consequence of increasing El Niño intensification30. In Africa, a high proportion of undisturbed forest locations (56%) transitioned to novel mean annual temperature regimes. These locations were primarily concentrated across the Congo Basin which also experienced strong shifts to novel mean diurnal temperature ranges and temperature seasonality. In Southeast Asia and Australia, cumulative temperature novelty was noticeably lower than in the rest of the global tropics. Nevertheless, high novelty in mean annual temperature was widespread, occurring across 24% of forest locations, predominantly across New Guinea, Sundaland and Wallacea biodiversity hotspots.

There are fragmented but substantial areas of tropical forest that have not recently transitioned to novel regimes in annual climate variables. Parts of the Guiana Shield region and much of the southwestern Amazon in Brazil and Peru have recently experienced annual climates similar to the historic baseline. In Africa, parts of the western Congo Basin, southwest Cameroon and the western Gulf of Guinea coastline have relatively stable climate regimes across multiple bioclimatic variables. Relative to the rest of the tropics, a high proportion (75%, compared with 46% in Latin America and 40% in Africa) of tropical forest locations in mainland Southeast Asia and Australia have not recently transitioned to highly novel temperature regimes; the mean fractional novelty in climate conditions relative to historic baselines is noticeably lower than elsewhere for most temperature variables. Papua New Guinea, coastal Indonesia and areas of continental Asia appear to be the locations most insulated from novel conditions as a result of climate change. A considerable proportion of these tropical forests are located along coastlines, where overall change in temperature across the 30 year period was less severe than within continental interiors.

Contiguous forest does not prevent temperature novelty

Although many tropical forests have experienced extensive degradation by a variety of anthropogenic disturbances, they also form a considerable part of Earth’s most ecologically unfragmented environments—wilderness areas22. Because these areas are, by definition, subjected to fewer forms of anthropogenic disturbance, any impacts caused by climatic changes are of particular concern. By contrast, degraded forest might be expected to undergo transitions to novel temperatures due to widespread perforations and reductions in canopy cover resulting from wildfires and selective logging. To investigate whether there were differences in temperature novelty between tropical forests with different degrees of human disturbance, we compared the novelty of: (1) undisturbed forest located in areas where habitat is largely unfragmented30; (2) all other undisturbed forested areas where there will be greater pressures from human activities; and (3) forested areas classified as degraded6 in 2019 (that is, where a visual disturbance or repeated visual disturbances less than 2.5 years was observed = between 1982 and 2019).

As expected, novelty in temperature regimes was more prevalent within tropical forests than outside of them (Supplementary Fig. 6). Moreover, we found no evidence that unfragmented, intact ecological areas provided additional mitigation against modelled novel temperatures. Indeed, the mean temperature novelty in unfragmented, ecologically intact forests was higher than in other undisturbed forests and degraded forests, albeit with greater spatial variance (Fig. 2, Supplementary Table 1 and Extended Data Fig. 2). This outcome can be attributed to lower interannual variability in thermal conditions within intact forests, whereby incremental changes in temperatures result in novel temperature regimes.

Fig. 2: The distribution of temperature novelty across undisturbed and degraded tropical forests.
figure 2

Box plots of the distribution of temperature novelty scores in tropical forest across Africa (AFR, n = 80,599), Central and South America (CSA, n = 208,002), and Southeast Asia and Australia (SEAA, n = 85,596) for mean annual temperature, mean diurnal temperature range and temperature seasonality. Climate novelty values are separated into three distinct forest classifications: undisturbed tropical forest outside ecologically unfragmented areas, undisturbed tropical forest within ecologically unfragmented areas (defined by wilderness areas22) and degraded tropical forest only. The horizontal line within the box plot displays the median of the data, the box limits refer to the interquartile range, and the whiskers extend to the minimum and maximum values. The data points falling outside the whiskers are outliers.

Tropical forests in Latin America’s ecologically unfragmented areas were worst affected overall, having experienced highly novel temperature regimes across almost all bioclimatic variables, while in Africa, Southeast Asia and Australia they experienced especially high novelty in mean annual temperatures (Supplementary Table 1). Pristine forests such as these, many of which form part of Indigenous peoples’ lands, are vital for biodiversity conservation30. They support higher levels of biodiversity, reducing the risk of extinction for highly threatened taxa—especially terrestrial mammal species which are often otherwise in conflict with urban populations31—and providing spatial connectivity across environmental gradients allowing for gene flow and genetic adaptation under climate change32. Our findings suggest that the climate regimes of degraded tropical forests can still offer some resistance to rising temperatures. Fortunately, these forests have been shown to retain some conservation value and are especially important in regions with little remaining undisturbed forest33. Although forest disturbances will alter the climate by removing parts of the canopy, some studies have shown that disturbed forests can attain canopy closure very quickly and so the impacts on climate can be temporary20,34.

Stable climates are more sensitive to long-term change

Our index of novelty represents the fraction of years in the recent period in which temperature lies outside the range of baseline historic conditions and is thus influenced by both absolute change and interannual variance. Because species are also sensitive to temperature thresholds35, we sought to establish the relative importance of absolute change in temperature, as opposed to interannual variability, on novelty. We calculated the mean change in bioclimatic variables below the canopy and then quantified the relationship between overall changes in each bioclimatic variable and recent fractional temperature novelty (Fig. 3 and Extended Data Fig. 3). We found statistically significant relationships between all mean changes in variables and their corresponding novelty indices (Supplementary Table 2). For instance, a <1 °C increase in mean annual temperatures over the last three decades was equivalent to almost entirely novel below-canopy mean annual temperatures across most of the tropics. Consequently, those forest locations highlighted as having recently transitioned to a highly novel temperature regime (Extended Data Figs. 4 and 5) also typically experienced high temperature change over the last 30 years (Supplementary Fig. 4).

Fig. 3: The relationship between the change in the mean of temperature variables and recent temperature novelty.
figure 3

ac, Scatterplots showing the correlation (as investigated using piecewise GLMs with a binomial logit) between the below-canopy novelty of each temperature variable and the change in the same variable (that is, the difference between the mean of 1990–2004 and the mean of 2005–2019) across undisturbed tropical forests (n = 317,809) for mean annual temperature (a), mean diurnal temperature range (b) and temperature seasonality (c). Each point represents one grid cell for Africa (n = 67,799), Central and South America (n = 185,883), and Asia and Australia (n = 64,127). See Supplementary Table 2 for model results for each group. Tests were conducted using two-sided Wald tests with a significance level set at P < 0.01. No adjustments were made for multiple comparisons because each temperature variable was analysed and presented separately.

The relationship between novelty and absolute change was strongest across mean annual temperature and mean diurnal temperature regimes. Within undisturbed tropical forests, the average long-term change in mean annual temperature over the last three decades was highest in Africa (0.50 °C), followed by Latin America (0.41 °C), and Southeast Asia and Australia (0.37 °C). Across Latin America, mean annual temperatures in four of the recent years were up to 0.25 °C higher than those that occurred during the El Niño drought of 1997–1998, and the mean diurnal temperature range was up to 0.32 °C higher in all 15 recent years (Supplementary Table 4). However, there are forest regions, such as west of the Albertine Rift, which did not recently experience novelty in certain climate variables despite long-term changes occurring, suggesting that interannual variability confers resilience to changing climate conditions in some areas.


Tropical forests are the world’s most diverse terrestrial ecosystems, hosting more than 62% of vertebrate species36 and over 75% of flowering plant species1. Mechanistic modelling of the below-canopy environment suggests there have been pronounced shifts in below-canopy climate regimes to novel conditions in a significant proportion of tropical forests, including globally important national parks, Indigenous reserves and large tracts of ecologically unfragmented areas. Novel temperature regimes were frequently a signature of low-lying continental interiors where there is limited opportunity for species to access elevational climate gradients if thermal limits are breached18.

Although we cannot confidently draw conclusions on implications for biota that do not occur near the ground, recent research in largely undisturbed and/or primary lowland tropical has found changes in species composition37,38,39 and significant declines in animal, insect and plant populations40,41,42,43. These changes are attributed to warming temperatures and are consistent with our findings. For instance, the abundance of terrestrial and near-ground insectivore bird species has declined in primary tropical forest in the Brazilian Amazon since the 1980s44, with evidence that these species respond to warming by closely tracking cooler microclimates16, while in Panama, most understorey bird species experienced large (>50%) proportional losses in estimated abundances since 1977, irrespective of ecology42. As a result of lower natural climate variability, many tropical forest species have narrow realized climate niches37 and are not pre-adapted to warmer conditions45. The ongoing transition of tropical forest environments to almost entirely novel temperature regimes can easily precipitate changes in niche availability, favouring species with higher temperature affinity46, and triggering changes in community composition through trophic cascades47,48.

We also identified areas pantropically where temperature novelty has been low despite ongoing climatic change. Although many are highly fragmented49 and dispersed across continents, especially in Africa, these tropical forests are the best candidates to act as climate refugia and are crucial to conservation efforts. Their usefulness as refugia will depend on their connectivity to areas with unfavourable climates, enabling species range-shifts50. Severe fragmentation will also influence temperature novelty as edge effects have been shown to reduce microclimatic temperature buffering up to 20 m into the interior51. It is paramount that distant wealth-related drivers of deforestation and degradation begin to be sufficiently addressed52 and that intact candidate refugia are urgently and vigorously protected, via legal protection53, carbon payments54 or empowering indigenous communities55. In turn, we urgently need to direct forest restoration programmes to improve the connectivity, overall size and interior (non-edge effected) area of fragmented refugia56,57. Notwithstanding the fundamental need for global carbon emission reductions, the prioritization and protection of refugia and the restoration of highly threatened forests is vital to mitigate further damage to global tropical forest ecosystems.


Climate data and modelling

Using a recently developed grid version21 of a previously published mechanistic microclimate model58, we quantified hourly below-canopy climate conditions across the global tropics (30° S to 30° N) between 1990 and 2019. The microclimate model was run in daily time increments and then hourly temperatures—at 0.05 m above the ground—were derived using the model’s interpolation methods, which infer hourly data from daily minima and maxima using the diurnal cycle in the ambient temperatures provided as inputs to the model. Full details of the model are provided58, but in summary the following workflow is implemented. First, the model downscales hourly input climate-forcing data to the desired spatial resolution (in this case 5 km gridded resolution) using spatial interpolation and the application of an elevation- and humidity-dependent lapse rate correction. Temperature and water vapour at the desired height are modelled mechanistically using principles of energy conservation, that is, by assuming that components of the energy budget remain in balance, and by solving the energy budget to derive differences between near-ground and ambient temperature using the Penman–Monteith equation. Radiative fluxes through the canopy are estimated using Seller’s two-stream approximation model59. Sensible and latent heat fluxes are assumed to depend on wind speed, which in turn is attenuated vertically by canopy foliage using the method described60. Wind speed is terrain-adjusted using the method described61. Latent heat fluxes are assumed additionally to depend on the stomatal conductance of leaves, which is quantified from the availability of photosynthetically active radiation using the method described62. Ground heat fluxes are quantified from canopy–soil temperature gradients, the latter contingent primarily on radiation absorbed by the ground using the method described63.

The hourly climate-forcing data required to drive the microclimate model were obtained from the ERA5 fifth-generation ECMWF atmospheric reanalysis of the global climate, using the single levels surface dataset64 at a 0.25° gridded spatial resolution for the 30 year time period. The ERA5 climate data assimilate past climate observations with climate model predictions to generate a series of climate variables for atmospheric, land-surface and sea parameters. The following climate variables were extracted for the extent of the study area: (1) air temperature at 2 m, (2) dewpoint temperature at 2 m, (3) pressure at surface, (4) precipitation rate, (5) U-wind speed at 10 m (west to east component), (6) V-wind speed at 10 m (south to north component), (7) total cloud cover, (8) downward long-wavelength radiation and (9) downward solar radiation, which was partitioned into direct and diffuse components using the method described65.

Additionally, the following environmental predictors were obtained to drive the microclimate model: (1) annual habitat type, sourced from the European Space Agency Climate Change Initiative66 at a gridded spatial resolution of 5 km; (2) annual vegetation height, sourced from ORNL DAAC67 at a gridded spatial resolution of 5 km; (3) monthly plant area index, calculated as the sum of monthly leaf area index (LAI) and 20% of the monthly maximum LAI. Monthly LAI values were sourced from the National Oceanic and Atmospheric Administration68 and spatially aggregated to a gridded spatial resolution of 5 km, with missing values estimated from the LAI at the same location in other months using a locally informed month effect accounting for seasonal cycles; (4) monthly canopy and ground reflectance at a spatial resolution of 5 km gridded resolution, calculated by first deriving the fractional canopy cover from surface albedo68 and monthly LAI values and then using the fractional canopy cover to partition surface albedo between ground and canopy; both steps used the microclima69 package for R v.4.270; (5) soil type, sourced at a gridded spatial resolution of 250 m from soilgrids.org71, which was then resampled to a gridded spatial resolution of 5 km using the nearest-neighbour method; (6) a digital elevation model, sourced from the US Geological Survey72 at a gridded spatial resolution of 7.5 arcsec and resampled to 5 km using a bilinear method; and (7) a topographic wetness index at a gridded spatial resolution of 5 km, calculated by using the digital elevation model to derive flow accumulation. Inevitably the model is sensitive to uncertainty and error in the data used to drive the model. Sensitivity analysis indicated it was most sensitive to assumed LAI, although mostly at low LAI values, which are typically derived with greater accuracy73.

To assess whether the microclimate model was more accurate than ERA5 climate data and adequately represented below-canopy conditions, the hourly modelled microclimate temperatures and the ERA5 hourly temperatures were both compared to in situ measurements of temperature obtained from 70 locations under tropical forest canopies across the Americas, Africa and the Sundaland and represented in the SoilTemp database74. Below-canopy temperatures were modelled independently of our global results at a higher gridded resolution of 500 m (reflecting the higher-resolution LAI datasets available for recent years). Additionally, to avoid duplicating results, we did not use temperature loggers that were located in the same 500 m grid cell (Extended Data Fig. 6). For each temperature logger, microclimate was modelled as above using ERA5 reanalysis climate variables and vegetation parameters (Supplementary Table 6) which matched the time and duration of the empirical temperature observations (Supplementary Data 1). The average for each recorded time series of empirical temperatures—taken separately for microclimate and ERA5 reanalysis—was used to derive a single root mean square error (r.m.s.e.) to quantify similarity to the logger observations (Extended Data Fig. 7). As indicated by the r.m.s.e. from empirical observations, the microclimate model (r.m.s.e., 2.73) was more accurate than ERA5 (r.m.s.e., 3.62). Moreover, we derived a r.m.s.e. score for mean temperature at each logger location, separately for both microclimate and ERA5 reanalysis temperatures (Supplementary Data 1).

Temperature novelty analysis

The hourly modelled below-canopy climate conditions were used to calculate the annual bioclimatic variables detailed75, namely: (1) mean annual temperature, (2) mean diurnal temperature range, (3) isothermality (diurnal range/annual range × 100), (4) seasonality, (5) maximum temperature of the warmest month, (6) minimum temperature of the coldest month and (7) annual temperature range. The annual bioclimatic variables were then split into a baseline historical time period (1990–2004) and the most recent time period (2005–2019). For each grid cell, we derived an index of novelty for each bioclimatic variable (n = 7) from the fractional overlap in each variable’s values between the two periods76. Specifically, for each bioclimatic variable, we measured the fractional overlap between two sets of 15 annual values. This was done by computing the frequency distribution curves of the annual values across historical and recent time periods separately, and then novelty was derived as 1 minus the proportion of overlap in annual values between the two periods, calculated as:

$${\mathrm{novelty}}=1-(2\times {\mathrm{intersection}}\,{\mathrm{area}}/{\mathrm{total}}\,{\mathrm{area}}\,{\mathrm{of}}\,{\mathrm{both}}\,{\mathrm{curves}})$$

This novelty index represents the fraction of years in the recent period (2005–2019) in which the climate lies outside the range of conditions that occurred in the baseline historical period (1990–2004). For example, if both mean annual temperatures and interannual variance in mean annual temperature were identical in both periods, the novelty index would be zero. If two-thirds of the mean annual temperatures in the latter period lay outside the range of temperatures in the historic period, then the novelty index would be 0.6667. Thus, the locations with novelty indexes closer to 1 are those with no recent climate analogue relative to the recent historical baseline.

To exclude forest in which climate change could be amplified by interacting human activities such as deforestation, novelty index values for each of the seven bioclimatic variables were extracted for the locations of tropical moist forest which were still undisturbed in 20196, defined as all closed forests in the humid tropics including the tropical rainforest and the tropical moist deciduous forest without any observed disturbances (degradation or deforestation) across the full observation period defined by the available Landsat data (1982–2019). The definition is not based on percentage of canopy cover and does not discriminate between primary- and secondary-growth tropical forest because there are no Landsat data available prior to 1982. However, it is probable that undisturbed tropical forest cover as estimated here is close to the true extent of primary tropical forest due to the amount of time that they have been undisturbed.

We investigated whether climate novelty differed between tropical forests with different degrees of human disturbance using three categories of forest: (1) undisturbed forested areas in ecologically unfragmented areas30, (2) undisturbed forested areas outside of ecologically unfragmented areas, and (3) tropical forests classified as degraded6 in 2019 (where a visual disturbance or repeated visual disturbances have been observed from space between 1982 and 2019, but each disturbance event lasted <2.5 years and was therefore not classified as deforestation). In this study, we considered deforested locations as no longer able to meet the biological requirements of a tropical forest and so these locations were not included in our analyses.

Finally, we calculated the overall change in each bioclimatic variable (n = 7) as the difference between the mean value for each time period (baseline period, 1990–2004; recent period, 2005–2019). We used a generalized linear model (GLM) with a binomial logit to fit the relationship between overall changes in each bioclimatic variable and recent fractional temperature novelty for each grid cell.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.