Main

When flash droughts occur, the rapid depletion of soil water causes cascading impacts on agriculture, economy and society1. Severe events have occurred around the globe in recent years, such as in the USA in 2012 and 20172, Russia in 20103, southern Africa in 20154, eastern Australia in 20185 and southern China in 20226. Flash droughts are described as the outcome of either a primarily precipitation-driven or primarily temperature-driven pathway, with the latter involving heatwaves that enhance evapotranspiration and thereby foster soil desiccation7. Conversely, unusually wet (dry) soils can dampen (amplify) the development of extreme heat by modulating the surface energy partitioning8. In other words, soil water deficit can be seen as both the consequence of and a contributor to the occurrence of extreme heat9. Given the underlying tug of war between temperature-driven increases in atmospheric evaporative demand and soil dryness as flash droughts develop, a central question is: how are flash droughts accompanied—yet not necessarily driven—by extreme heat development across the globe?

There is extensive evidence that long-lasting droughts concurring with heat cause a widespread reduction of carbon uptake and, hence, an increase in atmospheric CO2 growth rate10,11. Flash droughts are likewise expected to exert adverse impacts on vegetation growth, with spatial heterogeneity governed by background climate conditions12,13. Global terrestrial ecosystems mitigate climate change by assimilating around one-third of anthropogenic carbon emissions every year14. The carbon sink capacity of land vegetation is highly regulated by water availability and temperature variability15,16. Dry conditions and heat stressors are harmful to plant photosynthesis and growth, causing depletion of carbon reserves across multiple ecosystems, including croplands, pasturelands and forests17,18. However, the impact of hot flash droughts on vegetation growth and carbon uptake remains unclear. Furthermore, the global ecological response to compound heat flash droughts (CHFDs) and non-heat flash droughts (NHFDs), based on the presence or absence of extreme heat during the flash drought, has not yet been explored.

Here, we assess how the frequency, onset speed, duration, severity and driving mechanisms differ for CHFDs and NHFDs around the globe, using long-term (1950–2022) atmospheric observation-derived datasets. We then apply supervised machine learning techniques to a combination of satellite-retrieved gross primary productivity (GPP)19, a solar-induced chlorophyll fluorescence (SIF) product20 and global biodiversity coverage maps21, to evaluate whether the impacts of CHFDs versus NHFDs differ on terrestrial ecosystems for different background environments.

Asymmetric characteristics of CHFDs and NHFDs

We first compare the frequency, severity and onset rate of CHFDs and NHFDs over the 1950–2022 period using the ERA5-Land and ERA5 datasets, respectively (Fig. 1 and Supplementary Fig. 1). The low latitudes, including India and eastern Asia, are hotspots of flash droughts, in particular for NHFDs (Fig. 1a,b and Supplementary Fig. 1a,b). These regions are generally covered by forests, savannas and croplands with humid climates (Extended Data Fig. 1), where abundant vegetation contributes to the occurrence of flash droughts through the consumption of soil water availability22. Nonetheless, these tropical to subtropical areas tend to have high rooting-zone water storage23. Soil moisture depletion enables evaporative cooling over these areas, where evapotranspiration remains high. This limits surface sensible heating and counteracts temperature escalations; hence, there is a higher proportion of NHFDs than CHFDs (Extended Data Figs. 2 and 3). By contrast, CHFD-prone regions (defined as having a ratio of the number of CHFD to total flash droughts higher than 50%; Methods) are located mainly in the extratropics (Extended Data Fig. 3). The high fraction of CHFDs in the Southern Hemisphere prevails in dry regions, partially related to a limited cooling effect of latent heat flux constrained by low soil moisture availability9,24. The high fraction of CHFDs in the Northern Hemisphere, usually located in extratropical humid regions, is related to stronger seasonality of shortwave radiation anomalies than in tropical regions. In addition, large-scale circulation patterns such as persistent high-pressure systems, the increasing frequency of split midlatitudes jets, and reduced cloud cover, which frequently emerge in CHFD-prone regions of the Northern Hemisphere25,26, largely favour the occurrence of hot flash droughts.

Fig. 1: Frequency, onset rate and severity of CHFDs and NHFDs during 1950–2022.
figure 1

a,b, The frequency of CHFDs (a) and NHFDs (b). c,d, The onset rate of CHFDs (c) and NHFDs (d). e,f, The severity of CHFDs (e) and NHFDs (f). g, Box plots of the frequency, onset rate and severity of CHFDs (n = 6,057 from ERA5-Land and n = 6,116 from ERA5) and NHFDs (n = 5,815 from ERA5-Land and n = 6,306 from ERA5) across CHFD-prone and NHFD-prone regions, respectively. The CHFD-prone (NHFD-prone) regions are defined as the regions where the number of CHFDs (NHFDs) divided by the total flash droughts exceeds 50% (see the Methods for detailed definitions). The centre line of the box plots (g) indicates the median value, and the box bounds (whiskers) indicate the 25th and 75th percentile (minimum and maximum) values. Outliers are not displayed. Asterisks indicate that the frequency, onset rate and severity of CHFDs are significantly (two-sided t-test, P = 0.05) different from these of NHFDs. The stippling in c and e denotes regions where the onset speed and severity of CHFDs is significantly (two-sided t-test, P = 0.05) different from that of NHFDs between 1950 and 2022.

Full size image

We find significant differences between CHFDs and NHFDs, with CHFDs exhibiting (1) larger severity, (2) longer duration, (3) longer recovery time and (4) a slower onset rate over the tropics (Fig. 1 and Supplementary Figs. 1 and 2). Globally, the severity is 6.7–90.8% (range between 25th and 75th percentile) and the recovery duration is 8.3–114.3% higher for CHFDs compared with NHFDs. These findings indicate that the processes and interactions giving rise to extreme heat during flash droughts significantly lengthen and exacerbate both the event duration and the magnitude. Heatwaves typically occur in conjunction with pronounced and persistent anticyclones, leading to subsidence, adiabatic warming and high atmospheric stability27,28,29, hence hindering the occurrence of precipitation and promoting soil desiccation during flash droughts. For comparison, we also incorporate the Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2) reanalysis30 and the physics-informed dataset Global Land Evaporation Amsterdam Model (GLEAM)31. Even though these products are available only from 1980 onwards, both datasets yield broadly consistent results with ERA5-Land and ERA5 (Supplementary Figs. 3–7).

Underlying mechanisms behind CHFD and NHFD onset

To uncover the physical mechanisms of CHFDs and NHFDs, we examine various hydrometeorological drivers in terms of surface flux patterns (Fig. 2 and Extended Data Figs. 4 and 5). Flash droughts are typically accompanied by a lack of precipitation, and we find no clear differences in the contribution of precipitation deficits between CHFDs and NHFDs. In the case of hot flash droughts, strongly positive anomalies of downward solar radiation enhance both surface latent and sensible heat fluxes. The development of these flash droughts is thus characterized by stronger soil moisture drawdown due to above-average evapotranspiration, which in turn is further promoted by increasing atmospheric moisture demand and, hence, surface sensible heating (Fig. 2a and Extended Data Figs. 4 and 5). In particular, the latent heat (soil moisture) should increase (decrease) rapidly in the initial stage, during which the soil moisture is still abundant. As soil desiccation continues, evapotranspiration becomes increasingly water-limited, manifesting as negative evaporative fraction anomalies (median −1.7% in CHFDs). However, this water limitation is too weak to compensate for the rapid radiation- and temperature-fuelled soil desiccation, such that the total soil moisture depletion during CHFDs is comparatively severe. For NHFDs, by contrast, anomalies of incoming radiation are only weakly positive. This energy surplus is partitioned into minor sensible and latent heating anomalies of similar magnitudes (Fig. 2b and Extended Data Fig. 4). Compared with CHFDs, evapotranspiration during NHFDs remains energy- rather than water-limited throughout the entire onset stage, as indicated by the near-zero change in evaporative fraction (median −0.4% in NHFDs). Less positive latent heat anomalies in NHFDs than in CHFDs and near-zero evaporative fraction anomalies indicate weaker soil moisture stress, and thus flash droughts without extreme heat are less severe than hot flash droughts. We obtain consistent results using the ERA5 dataset (Supplementary Figs. 8 and 9).

Fig. 2: Onset mechanisms of CHFDs and NHFDs.
figure 2

a,b, Median anomalies of precipitation (mm per day), surface net solar radiation (W m−2), surface sensible heat flux (W m−2), surface latent heat flux (W m−2), evaporative fraction (%), relative humidity (%) and daily maximum temperature (°C) during CHFDs over CHFD-prone regions (a) and during NHFDs over NHFD-prone regions (b) based on the ERA5-Land dataset.

Full size image

Previous studies have indicated that heatwaves accelerate flash drought onset through boosting evapotranspiration32. Our study shows that, globally, hot flash droughts do not always develop faster than NHFDs when gauged by declining soil moisture percentiles, the standard approach in flash drought research33. This is caused primarily by the fact that, from a seasonal perspective, CHFDs tend to occur in the peak of the warm season over most areas—a pattern consistently observed across different datasets, when the potential for severe soil desiccation is the highest (Supplementary Figs. 10–12). Although these events are associated with a stronger absolute soil moisture drawdown compared with NHFDs (Extended Data Fig. 5h), this does not result in a quicker relative soil moisture depletion (Fig. 1c,d), considering that onset speed as defined here gauges the soil moisture decline relative to what is expected for a given location and season. In fact, CHFDs have much faster soil moisture drawdown—or absolute onset speed—than NHFDs over the mid-to-high latitudes, where strong incoming radiation anomalies promote evapotranspiration (Extended Data Fig. 6). In the tropics, flash drought onset is accelerated by the warm environment, which provides ample energy for evapotranspiration and rapid soil drying under precipitation shortages. As such, even though NHFDs involve less excess radiation and lower evapotranspiration, the resulting absolute onset speed is comparable to CHFDs (Supplementary Figs. 13 and 14). However, over subtropical transition zones (for example, most of India or Mexico), we note strong evapotranspiration throttling under water limitation for CHFDs, which increases the atmospheric demand for moisture and enables extreme heat, yet also slows further soil desiccation (Extended Data Fig. 6). In other words, the absolute onset speed of CHFDs over these regions can be slower than NHFDs owing to (positive) soil moisture–temperature feedback.

Hot flash droughts reduce carbon uptake in the recovery stage

In a next step, to reveal how hot and dry conditions drive vegetation stress, we quantify the impacts of flash droughts on carbon uptake by estimating the associated GPP and SIF anomalies for the onset and recovery stages of flash droughts with and without extreme heat (Fig. 3 and Supplementary Fig. 15). During the onset stage, GPP anomalies have an interquartile range of −0.57 to 0.16 gC m−2 per day during CHFDs whereas they range from −0.74 to −0.04 gC m−2 per day during NHFDs, suggesting that CHFDs are associated with slightly weaker vegetation constraints during the onset. Spatially, boreal forests produce neutral or even slightly positive GPP and SIF anomalies during the onset of CHFDs, while the Southern Hemisphere CHFD-prone regions, mostly covered by grasslands and shrublands, produce more negative GPP and SIF anomalies (Fig. 3a,b and Supplementary Fig. 15a,b). This demonstrates a clear gradient of drought resistance across different ecosystems. Southern Hemisphere grasslands and shrublands with fewer biotic attributes and low water retention are more vulnerable to such dry extremes34,35.

Fig. 3: Anomalies of carbon variables during the onset and recovery stages of CHFDs and NHFDs.
figure 3

a,b, GPP anomalies during the onset stage of CHFDs (a) and NHFDs (b). c,d, GPP anomalies during the recovery stage of CHFDs (c) and NHFDs (d). e, A box plot of GPP and SIF anomalies during the onset and recovery stages of CHFDs (n = 6,057 from ERA5-Land and n = 6,116 from ERA5) and NHFDs (n = 5,815 from ERA5-Land and n = 6,306 from ERA5) over CHFD-prone and NHFD-prone regions, respectively. The centre line of the box plots (e) indicates the median value, and the box bounds (whiskers) indicate the 25th and 75th percentile (minimum and maximum) values. Outliers are not displayed. Asterisks indicate that the GPP and SIF anomalies during CHFDs are significantly (two-sided t-test, P = 0.05) different from these during NHFDs.

Full size image

Moving on to the recovery stage (Fig. 3c,d and Supplementary Fig. 15c,d), we find greater-than-usual decreases of GPP for CHFDs, with anomalies ranging from −1.09 to −0.20 gC m−2 per day. More negative GPP and SIF anomalies in the recovery stage of flash droughts than in the onset stage suggest that there is a lagged response. The CHFDs show a greater reduction in carbon uptake compared with the NHFDs, implying that the co-occurrence of dry and heat extremes hampers vegetation growth more severely than dry conditions alone. This is expected because hot flash droughts last longer and are more intense (Fig. 1), potentially resulting in stomatal limitations (plants close their stomata to reduce transpiration) and biochemical limitations (reduced carboxylation and electron transport rates)10. Geographically, GPP and SIF anomalies are less negative in tropical forests than in other regions, especially for CHFDs (Fig. 3c). This is partly due to less severe CHFDs in tropical forests relative to mid-to-high latitudes, as indicated by less negative soil moisture anomalies (m3 m−3) during drought development (Extended Data Fig. 5h). The higher drought resistance of tropical forests could additionally be rooted in biophysical properties such as abundant biomass and tree species diversity36,37, which can mediate the impacts of flash droughts via transpiration (a buffer effect). In boreal forests and the Southern Hemisphere extratropics, CHFDs are marked by pronounced soil moisture depletion (Extended Data Fig. 5h), leading to substantial constraints on carbon uptake. Furthermore, different canopy structures and leaf-level water use efficiency in high latitudes from tropical forests may contribute to more reduced carbon uptake during CHFDs10.

To better understand how flash droughts regulate carbon uptake, we construct supervised machine learning models describing the hydrometeorological conditions and background environment driving GPP changes across the forests, shrublands, savannas, grasslands and croplands (Methods; Fig. 4 and Supplementary Figs. 16–20). Hydrometeorological conditions are the leading predictors of GPP anomalies (>42.2% importance for all ecosystems across the onset and recovery stages of flash droughts), followed by the local climatology and biophysical properties (Fig. 4 and Supplementary Figs. 17 and 18). This supports the notion that flash droughts mediate vegetation productivity across varying ecosystems.

Fig. 4: Response functions of hydrometeorological, climatological and biophysical indicators to GPP anomalies during the onset and recovery of CHFDs.
figure 4

a,b, The importance of biophysical indicators (BP), including species richness of mammals, birds, butterflies and plants; climatic indicators (CM), including AI, annual mean and annual variation of daily maximum temperature (Tmax-Mean, Tmax-Var) and precipitation (PRE-Mean, PRE-Var); and hydrometeorological anomalies (HM) during CHFDs, including anomalies of daily maximum temperature (T'max), relative humidity (RH'), latent heating (SLHF'), sensible heating (SSHF'), precipitation (PRE') and soil moisture (SM') to GPP anomalies across the forests (a) and croplands (b). IncMSE (%) denotes the percent increase in mean squared error. c–j, Partial dependence plots for eight key predictors of GPP anomalies, including SM' (c), SSHF' (d), SLHF' (e), RH' (f), AI (g), Tmax-mean (h), plants (i) and butterflies (j), during the onset of CHFDs over forests (Forests-O) and croplands (Croplands-O) as well as during the recovery of CHFDs over forests (Forests-R) and croplands (Croplands-R).

Full size image

By comparing the hydrometeorological indicators regulating GPP anomalies during the onset and the recovery stages of CHFDs, we find that, for forest, shrubland and savanna ecosystems, surface latent heating matters most during the onset. Soil moisture becomes increasingly important during the recovery, however, pointing to the emergence of water limitation (Fig. 4a and Supplementary Fig. 17a–c). The rising importance of soil moisture from onset to recovery stages of CHFDs also appears for grassland and cropland ecosystems (Fig. 4b and Supplementary Fig. 17d,e). Overall, this enhanced importance of soil moisture during flash drought recovery across various ecosystems is consistent with more severe GPP reductions compared with the onset (Fig. 3), partly linked to the fact that carbon fluxes are more controlled by soil moisture stress rather than atmospheric dryness38,39. We note that cropland ecosystem productivity is anomalously constrained by CHFDs relative to other ecosystems (Supplementary Fig. 19), posing a serious threat to food security. The cropland ecosystems tend to show less negative GPP anomalies with lower aridity index (AI) values (Fig. 4g). This negative correlation between AI and GPP anomalies might relate to different crop species cultivated across semi-arid to humid regions. Forests with increasing AI values (when AI <0.65), by contrast, mitigate the ecosystem productivity decline (Fig. 4g). These different responses between croplands and forests partly reveal the difference between managed and natural ecosystems. Regarding biophysical properties, we find that ecosystems with higher biodiversity, for example, gauged by the number of plant species and butterflies, are slightly more drought resistant, consistent with expectations (Fig. 4i,j and Supplementary Fig. 19).

Implications of increasing CHFDs

How extreme heat interacts with flash—rather than slow—droughts, and how this affects ecosystems, has remained unclear. Extreme heat has been considered only as a physical driver of flash drought7,40, but our study reveals that extreme heat also plays a critical role in shaping the characteristics of flash droughts, as its emergence is associated with longer and more severe flash droughts. Furthermore, we find notable increases in the frequency of CHFDs over the past 73 years (Supplementary Fig. 21). This transition from NHFDs to CHFDs might accelerate under additional global warming, posing challenges for impact mitigation.

Our analysis relies primarily on the ERA5-Land and ERA5 products owing to their proven ability to capture flash droughts41,42. However, to make our analysis more robust, we also incorporate the MERRA-2 reanalysis and GLEAM for comparison. Although cross-dataset analysis introduces some uncertainty in the frequency and ratio of CHFDs and NHFDs, all datasets consistently support our key findings: compared with NHFDs, hot flash droughts are more severe and last longer; CHFDs also develop more rapidly in terms of absolute soil moisture decline (except in transitional climate zones) and tend to occur in peak summer.

To further assess impacts of flash droughts within our percentile-based framework, we analyse the minimum soil moisture, the absolute soil moisture decline and the difference to critical soil moisture thresholds during CHFDs and NHFDs (Methods and Supplementary Figs. 22–25). CHFDs cause a more severe soil moisture drawdown compared with NHFDs in mid-to-high latitudes, while showing minor differences in tropical humid areas (Extended Data Fig. 7). In subtropical transitional regions, CHFDs exhibit smaller absolute soil moisture decline but reach substantially lower minimum soil moisture than NHFDs. This occurs because CHFDs typically initiate when local soil moisture is already low in these regions. As CHFDs intensify, soil moisture is further depleted below critical thresholds, which may trigger catastrophic ecosystem losses in these subtropical transition zones (Supplementary Figs. 24 and 25).

Various other variables characterizing the surface water and energy budget (such as evapotranspiration) and combined drought metrics (including the flash drought intensity index) have increasingly been used to quantify flash drought impacts43. Comparing flash drought hotspots identified in this study with existing literature reveals some regional differences depending on the choice of drought metric (Extended Data Fig. 8). For instance, our results indicate that the boreal forest zone is a flash drought hotspot, consistent with findings based on an alternative soil moisture index44, yet not captured by the evaporative stress ratio indicator1. Furthermore, soil moisture-based indicators tend to detect more frequent flash droughts across the central and eastern USA, whereas precipitation-based indicators capture more events in the western USA45. It is difficult to identify a single index that is universally optimal across all regions. Future research should further explore the sensitivity of different flash drought indicators for global and regional monitoring and impact assessment.

Flash droughts threaten people’s livelihoods and reduce agricultural production46. Such extreme events tend to have disproportionate impacts on the vulnerable communities of the Global South47. Here, we quantify the global risk of CHFDs during 1950–2022, 1970–2022 and 1990–2022, by combining the hazard (occurrence probability of CHFDs), exposure (measured by population, cropland and pastureland) and vulnerability (country-level governance indicator (GI)48; Methods). We find that population and agriculture risks have increased globally in recent decades (Supplementary Fig. 26), driven mainly by the rising likelihood of CHFD occurrence. The Global South (Supplementary Fig. 27) has had a substantially high proportion of the risk to population and agriculture over the past three decades (Supplementary Fig. 26). Although our methodology focuses on the frequency of CHFDs and does not account for the severity of individual events, our results suggest that vulnerable countries in the Global South, including China, India and Indonesia, may suffer from continued increasing risks as the climate warms. Mitigation and adaptation actions need to be carefully considered to reduce the adverse impacts of CHFDs in these areas.

This study reveals the asymmetric characteristics and impacts of CHFDs and NHFDs on terrestrial ecosystems. We also note that our impact assessment does not incorporate parameterizations specific to biomes modulated by anthropogenic activities. In tropical savannas, such as northern Australia and southern Africa—among the world’s most fire-prone landscapes49,50—flash drought impacts can be more severe due to their potential to increase wildfire occurrence while also interacting with recent fire management initiatives. Future assessments should account for the complex interactions and feedbacks between flash droughts, associated hazards such as wildfires, and human adaptation strategies. While it remains challenging to fully understand how flash droughts impact ecosystem productivity across different climates, we call for better preparation of society to hot flash droughts in the future and particularly in the most vulnerable cropland ecosystems.

Methods

ERA5 and ERA5-Land reanalysis datasets

ERA5 is the fifth-generation atmospheric reanalysis of the European Centre for Medium-Range Weather Forecasts (ECMWF), providing estimates of a large number of atmospheric, land and oceanic variables from January 1940 to the present51. The atmospheric model is produced using 4D-Var data assimilation and model forecasts in CY41R2 of the ECMWF Integrated Forecast System and then coupled to a land-surface model (HTESSEL). ERA5-Land, by comparison, was produced by replaying the land component of the ERA5 climate reanalysis52. Specifically, it combines model data with observations across the world into a globally complete and consistent dataset using the laws of physics, and the climate forcing is corrected to account for the altitude difference between grid cells (lapse rate correction). Considering that ERA5-Land is available from 1950 to the present, we extract hydrometeorological variables from 1950 to 2022, that is, 2-m temperature (K), total precipitation (m), 2-m dew-point temperature (K), surface solar radiation flux (W m−2), surface latent heat flux (W m−2), surface sensible heat flux (W m−2) and root-zone soil moisture (m3 m−3). We aggregate the first three soil moisture layers in ERA5 and ERA5-Land outputs (for example, 0–7 cm, 7–28 cm and 28–100 cm) using weights determined by the distance of two adjacent depths to construct the root-zone (0–100 cm) averaged soil moisture: SM0–7cm × 0.07 + SM7–28cm × 0.21 + SM28–100cm × 0.72.

GLEAM and MERRA-2 datasets

The GLEAM output is a hybrid satellite-based modelling dataset dedicated to estimating global evapotranspiration and root-zone soil moisture53. This dataset has been widely used in a variety of applications, especially characterizing drought evolution54. MERRA-2 is the latest atmospheric reanalysis of the modern satellite era produced by NASA30. The land-surface model used in MERRA-2 is the Catchment Land Surface Model, which has the strength of addressing sub-grid-scale soil moisture variability and its effect on river run-off and evapotranspiration. We extract the water root-zone variable (0–100 cm) for this study.

Gridded ecosystem variables

To estimate the impacts of flash droughts on GPP and SIF, we used a Moderate Resolution Imaging Spectroradiometer (MODIS)-retrieved GPP dataset and a machine-learning-constrained SIF dataset19,20. The GPP dataset is based on an improved light use efficiency theory and is driven by satellite data from MODIS and climate data from National Centers for Environmental Prediction (NCEP) Reanalysis II, with a temporal resolution of 8 days for the period 2000–201919. The SIF dataset was generated by training a machine learning algorithm of Orbiting Carbon Observatory-2 (OCO-2) SIF observations and surface reflectance from the MODIS20. This dataset captures the temporal and spatial variability of raw OCO-2 SIF at the far-red band and is strongly related to spatiotemporal variations of GPP. Specifically, we used the clear-sky daily SIF for the period 2001–2022 with a 4-day temporal resolution. To understand the flash drought dynamics in different climate zones and land cover types, we also downloaded the AI from the Global Aridity Index and Potential Evapo-Transpiration Climate Database v355. The AI is defined as the average ratio of precipitation to potential evapotranspiration and has been widely used to measure precipitation availability over atmospheric water demand and to partition climate zones55. The land cover type product is obtained from the MODIS Terra and Aqua reflectance data using supervised classifications56.

Background global biodiversity patterns

To quantify how background biophysical conditions affect carbon uptake under flash droughts, we use a terrestrial species dataset including vascular plants, butterflies, birds and mammals, reflective of biodiversity patterns21. This dataset is derived from a large global database of around 1.9 billion records of plants, butterflies, birds and mammals. It contains more precise biodiversity information than voucher specimens in a natural history collection or direct field observations and has been widely applied in ecology research57.

Deriving evaporative fraction and relative humidity

Because relative humidity (RH) and evaporative fraction (EF) are not available in the ERA5 and ERA5-Land datasets, we estimate these by using 2-m dry bulb (Td) and dew point (Tdew) temperatures, latent heat flux and sensible heat flux. The atmospheric saturation vapour pressure (esat) is governed by the Clausius–Clapeyron relationship58

$${e}_{{{\mathrm{sat}}}}\left({T}_{{\mathrm{d}}}\right)={e}_{{\mathrm{s}}0}\exp \left[\frac{{L}_{{\mathrm{v}}}}{{R}_{{\mathrm{v}}}}\left(\frac{1}{{T}_{0}}-\frac{1}{{T}_{{\mathrm{d}}}}\right)\right]$$
(1)
$${e}_{{{\mathrm{sat}}}}\left({T}_{{{\mathrm{dew}}}}\right)={e}_{{\mathrm{s}}0}\exp \left[\frac{{L}_{{\mathrm{v}}}}{{R}_{{\mathrm{v}}}}\left(\frac{1}{{T}_{0}}-\frac{1}{{T}_{{{\mathrm{dew}}}}}\right)\right]$$
(2)
$${{\mathrm{RH}}}=\frac{{e}_{{{\mathrm{sat}}}}\left({T}_{{{\mathrm{dew}}}}\right)}{{e}_{{{\mathrm{sat}}}}\left({T}_{{\mathrm{d}}}\right)},$$
(3)

where \({L}_{{\mathrm{v}}}\) and \({R}_{{\mathrm{v}}}\) refer to the latent heat of vaporization (2.5 × 106 J kg−1) and vapour gas constant (461 J kg−1 K−1), respectively, and T0 = 273.15 K and es0 = 611 Pa are integration constants.

The evaporative fraction is defined as the ratio between the latent heat flux and the available energy and is controlled by environmental factors including air temperature, surface solar incoming radiation, wind speed, soil water content and leaf area index59

$${{\mathrm{EF}}}={{\mathrm{SLHF}}}/({{\mathrm{SLHF}}}+{{\mathrm{SSHF}}}).$$
(4)

EF represents the evaporative fraction, SLHF denotes surface latent heat flux and SSHF means surface sensible heat flux.

Identification of CHFDs and NHFDs

Flash droughts are identified on the basis of pentad-mean (consecutive 5-day period) soil moisture60 during the extended warm season (May to October for North Hemisphere and November to April for South Hemisphere). First, we estimate the percentiles for each calendar pentad during the 1950–2022 period. If the pentad-mean soil moisture decreases from above the 40th percentile to the 20th percentile, with an average decrease rate of more than 5% per pentad, a flash drought is considered to have occurred. The flash drought terminates when the soil moisture percentile rises back above the 20th percentile. Specifically, each flash drought is divided into the onset stage (soil moisture decrease from above the 40th percentile to the first time below the 20th percentile) and the recovery stage (Supplementary Fig. 28). We consider that a flash drought event must last for at least three pentads60. A CHFD is then defined as the simultaneous occurrence of a flash drought and extreme heat, that is, droughts in which the pentad-mean temperature percentile (based on the pentad maximum temperature, Tmax) is above a determined threshold (90th percentile). A NHFD, by contrast, is a flash drought during which the temperature percentile remains below the threshold (Supplementary Fig. 28). The onset speed (% per pentad) is calculated as the difference between the start (the first pentad in which the soil moisture exceeds the 40th percentile) and the end of onset (the first time the soil moisture dips below the 20th percentile) divided by the length of onset. The duration denotes the total length (pentads) of a CHFD or NHFD, whereas the severity denotes the cumulative soil moisture deficit (expressed as percentiles, %) to the threshold (40%) across a CHFD or NHFD episode. A CHFD-prone region is defined as a region where the ratio of the number of CHFDs over the number of total flash droughts is above 50%, while a NHFD-prone region is defined as a region where the ratio of the number of NHFDs to the total number of flash droughts is above 50%. CHFD-prone regions are typically located in eastern and northern North America, southern South America, southern Africa, and central and northern Asia. NHFD-prone regions are typically found in southern North America, tropical forests, central Africa, southern Asia and Indonesia. All the box plots characterizing CHFDs and NHFDs in the Article are based on CHFD-prone and NHFD-prone regions, respectively (Extended Data Fig. 3).

The absolute soil moisture decline of CHFDs and NHFDs is defined as the change in soil moisture (m3 m−3) from the onset to its lowest during an event. The minimum soil moisture of CHFDs and NHFDs denotes the lowest soil moisture (m3 m−3) during an event (Supplementary Figs. 22 and 23). The difference to critical soil moisture thresholds denotes the difference between the minimum soil moisture during an event and the critical soil moisture thresholds (Supplementary Figs. 24 and 25).

Impacts of CHFDs and NHFDs on terrestrial carbon uptake

To assess how extreme events affect terrestrial ecosystems, we extract the minimum GPP and SIF anomalies during the onset and recovery stage of each CHFD and NHFD episode and then average all anomalies over the available time period (2000–2019 for GPP and 2001–2022 for SIF). Here, the minimum GPP and SIF anomalies are used rather than accumulated and averaged anomalies during the extreme events because the former better represent a constrained situation whereas the latter can include a buffer effect by the background ecosystems. To reduce the impacts from seasonality, the anomalies of GPP and SIF are based on each month rather than the whole warm season.

Existing studies have reported that the vegetation response to flash droughts is also dependent on the background environment12,61. To provide a detailed picture of how carbon uptake is mediated by flash droughts and background conditions, we consider both hydrometeorological anomalies during flash droughts, long-term climates and ecosystem characteristics when building models. We apply supervised machine learning models—random forest algorithms—with 15 independent variables to predict the impacts of the onset and recovery stage of CHFDs and NHFDs on GPP and SIF dynamics over the forest, shrubland, savannas, grassland and cropland ecosystems. The 15 variables can be categorized into three groups: (1) hydrometeorological indicators, including anomalies of soil moisture, precipitation, surface sensible heat flux, surface latent heat flux, relative humidity and maximum temperature over flash droughts; (2) background climatic conditions, including AI (precipitation over potential evapotranspiration), annual mean precipitation, interannual variance of precipitation, annual mean of maximum temperature and interannual variance of maximum temperature; and (3) background biophysical indicators (species richness), including the number of native species of plants, butterflies, birds and mammals. The background climatic and biophysical indicators are constant across CHFDs and NHFDs for each pixel.

We test each random forest model with regression trees ranging from 200 to 300, 400 and 500 and achieve similar results, with the determination coefficients higher than 0.76. To assess the influence of different predictors on the target GPP and SIF anomalies, we estimate the partial dependence function (\(\widehat{{f}_{{p}_{{\mathrm{s}}}}}\)), which calculates the predicted mean response of the target GPP and SIF anomalies to one selected predictor (ps) while the other predictors (\({p}_{{\mathrm{c}}}^{i}\)) are allowed to change in their domain62:

$$\widehat{{f}_{{p}_{{\mathrm{s}}}}}=\frac{1}{n}\mathop{\sum }\limits_{i=1}^{n}f\left(\;{p}_{{\mathrm{s}}},{p}_{{\mathrm{c}}}^{i}\right),$$
(5)

where the summand is the predicted regression function and superscript i indicates one episode. The R package randomForest was used to implement the analysis63.

Population, cropland and pastureland risk

To assess the implications of CHFDs for socioeconomic systems, we estimate the population and agriculture (expressed by croplands and pasturelands) risks to CHFDs by aggregating exposure, hazards and vulnerability. The exposure is represented by local population (using the Gridded Population of the World dataset-Version 4, GPWv464), pastureland area (based on the Global Pastures dataset from the Socioeconomic Data and Applications Center (SEDAC)65) and cropland area (based on the Global Croplands dataset from SEDAC66) at 1° resolution, respectively. The CHFD hazard is estimated by the probability of occurrence. Vulnerability is measured by the worldwide GI, available from 1996 to 2020. The GI captures household, business and citizen perceptions of the quality of governance in more than 200 countries and territories and provides a composite governance index based on six categories: voice and accountability, political stability, government effectiveness, regulatory quality, rule of law and control of corruption. After standardizing the indicator from its original −2.5 to 2.5 range to the range 0–1, we use the arithmetic average of the six components as our main response variable and refer to it as the GI. As lower GI values denote higher vulnerability, we estimate the risk using (1 − GI). Overall, the risk of CHFDs is determined as

$${R}_{{{\mathrm{cp}}}}=(1-{{\mathrm{GI}}})\times \frac{1}{{{\mathrm{AC}}}}\times \frac{1}{A}\mathop{\sum }\limits_{i=1}^{A}{p}_{i}\times {{{\mathrm{land}}}}_{i}$$
(6)
$${R}_{{\mathrm{p}}}=(1-{{\mathrm{GI}}})\times \frac{1}{{{\mathrm{AC}}}}\times \frac{1}{A}\mathop{\sum }\limits_{i=1}^{A}{p}_{i}\times \frac{1}{N}\mathop{\sum }\limits_{j,i}^{N}{{{\mathrm{po}}}}_{j,i},$$
(7)

where Rcp (Rp) represents the risk of cropland and pastureland (population); landi denotes the cropland or pastureland area inside a grid cell; pi indicates the probability of occurrence of CHFDs (calculated as the ratio of the number of CHFDs to the total number of years) during a given time period; A is the number of grid cells inside a country; poj,i indicates the population of the ith grid cell in the year j; and N is the year numbers of population datasets (that is, 2000, 2005, 2010, 2015 and 2020). To avoid the effect of the country size, we also use AC (km2), using the area of a country to obtain averaged population as well as cropland and pastureland risks.