Multilayer soil moisture depletion intensifies drought impacts on global ecosystems

Abstract

When soil moisture across all layers of the entire profile is depleted, ecosystems face the highest risk because they lose all vertical buffering capacity against drought. However, these vertically compound soil moisture droughts, in which every soil layer suffers simultaneous water deficits, have been largely ignored because previous monitoring and risk assessments relied on entire-profile mean soil moisture. Here we identify vertically (0–100 cm) compound droughts worldwide during warm seasons of 1981–2020 using multilayer soil moisture products. Their annual duration and ratio have increased across more than half of global lands over the past four decades, with pronounced hotspots in southwestern North America, South America, Africa, mid-latitude Eurasia, southeastern Asia and Australia. Vertically compound droughts cause substantial declines in ecosystem productivity, particularly in forests and croplands, threatening carbon sink stability and food security. By contrast, diagnosing drought solely from entire-profile averages causes an overestimation of vertically compound drought duration by about 128.3% but an underestimation of associated ecosystem losses by about 27.8%. These projections and systematic biases highlight the urgency of developing multilayer monitoring, mitigation and adaptation strategies for vertically compound droughts.

Vertically compound soil moisture droughts—when all soil layers are depleted—have intensified worldwide, driving sharper ecosystem losses than single-layer droughts, according to global analyses of multilayer soil moisture.

Main

Soil moisture drought undermines ecosystem structure and functioning^1,2, and its impacts are vertically heterogeneous^3,4. Ecosystem components such as plant roots and soil microbial communities exhibit vertically layered distribution within the soil profile^5,6, giving ecosystems buffering capacity when moisture deficits are confined to a single soil layer⁷. For instance, plants with both shallow and deep roots can still access water from other depths when one layer dries⁸, and soil respiration can also be partly sustained by other layers⁹. By contrast, when moisture deficits extend throughout the entire root zone, vegetation loses access to usable water at all depths and soil microbes lose their moist refuges. Water stress within the profile then shifts from being distributed to being superimposed, meaning that the vertical buffering and refuge space within the soil profile is largely lost. This loss of buffering can trigger nonlinear ecosystem responses, such as hydraulic failure in plants and loss of microbial refuge space, making large-scale tree mortality¹⁰, crop failures¹¹ and soil degradation¹² more likely and potentially irreversible. These concurrent moisture deficits across all soil layers (that is, vertically compound drought) therefore represent a high-risk form of drought, yet current knowledge about them remains insufficient to support the development of type-specific monitoring, mitigation and adaptation strategies.

Although vertically compound droughts are qualitatively recognized as posing higher risks to ecosystems, our quantitative understanding of their long-term evolution and ecological impacts remains limited. Most existing studies diagnose drought using entire-profile mean soil moisture^13,14, which conflates vertically mixed wet–dry conditions with truly profile-wide desiccation. This profile averaging can underestimate drought severity and impacts, while overstating the extent of entire-profile drought (Fig. 1). For instance, observations show that in some regions, such as the northern part of eastern Asia, droughts in the entire soil profile could be the net consequence between drying in deep soils and wetting in surface soils¹⁵. As a result, there is still no systematic quantification of how much ecosystem productivity is reduced and how vulnerable key ecosystems such as forests and croplands are under vertically compound droughts. Moreover, observations and simulations both show that anthropogenic forcings (including climate change and land-use change) have caused widespread entire-profile soil drying and intensified both surface and deep droughts over the globe^16,17. However, the extent to which anthropogenic forcings have reshaped the spatio-temporal evolution of vertically compound droughts remains unclear. Under this knowledge gap, traditional entire-profile-mean perspectives are likely to systematically underestimate the ecological losses associated with vertically compound droughts and to provide insufficient guidance for mitigation and adaptation. It is therefore urgent to adopt a multilayer perspective to identify and characterize vertically compound droughts, to evaluate their ecological impacts and to clarify their response to anthropogenic forcings.

Fig. 1. Illustration for identification of multilayer soil moisture drought.

a, Heatmap shows the classical detection of all drought types across the entire soil profile. b, Global map of vegetation-type classification from United States Geological Survey (USGS) (Methods). Antarctica, Greenland and deserts were excluded from this study (grey areas). c, Heatmap show the temporal evolution of all drought types across the entire soil profile for a grid cell in East Asia (purple triangle in b) during the 2020 warm season based on ERA5-Land. Specifically, the top-down and bottom-up evolutions of vertically compound drought occurred in June and August 2020, respectively. Drought intensity is defined as the difference between drought threshold and soil moisture. d, Histogram presents temporal changes in daily precipitation from ERA5-Land during the same period. Shading area represents land–atmosphere coupling intensity from ERA5-Land, whereas red and yellow lines show temporal changes in daily maximum air temperature from ERA5-Land and solar-induced fluorescence proxy from CSIF (Methods), respectively. Basemap data in b from Natural Earth (https://www.naturalearthdata.com).

In this study, we identified vertically compound droughts across the globe and revealed their spatio-temporal evolution during warm seasons from 1981 to 2020 using multilayer soil moisture from observations, reanalyses and land-surface model outputs (Supplementary Tables 1 and 2). We investigated the physical mechanisms of these events and their dominant hydrometeorological drivers. Next we quantified the associated ecological losses and used an interpretable machine learning framework to disentangle how hydrometeorological, climatic, vegetation and edaphic factors shape their geographic variability. Finally, we assessed the anthropogenic influence on vertically compound droughts and projected their future changes under the high-emissions scenario based on Earth system model simulations (Supplementary Table 3).

Spatio-temporal evolutions of vertically compound drought

We identified local soil moisture droughts in three layers (surface layer: 0–10 cm; middle layer: 10–40 cm; and deep layer: 40–100 cm). On the basis of all combinations of local drought occurrence across the three layers, warm-season drought days were classified into distinct drought types from a vertical perspective (Fig. 1a). Vertically compound drought is defined as the concurrent occurrence of drought in all three soil layers. During the warm seasons of 1981–2020, across three reanalyses and five land-surface model outputs, the global average annual duration of vertically compound drought is 6.0 $\pm$ 1.7 days (ensemble mean ± standard deviation), accounting for 28.5% $\pm$ 9.0% of all drought days over the global vegetated regions (Fig. 1b). Their occurrence ratios to all drought types (hereafter ratios) are about 1.5 times those of the two single-layer drought types, that is, surface-layer drought (annual duration: 4.5 $\pm$ 0.7 days; ratio: 19.0% $\pm$ 2.4%) and deep-layer drought (4.4 $\pm$ 1.0 days; 18.4% $\pm$ 3.0%), indicating that vertically compound drought is not rare.

Spatially, vertically compound droughts exhibit relatively longer durations and higher ratios in the tropics, the Middle East and East Asia (Fig. 2a,b), regions that are particularly prone to high-risk, profile-wide desiccation during the warm seasons of 1981–2020 based on the state-of-the-art reanalysis dataset (ERA5-Land). Surface-layer drought tends to be more prevalent over western North America, central Asia, western Europe, northern Africa and western Australia, whereas deep-layer drought is more prominent over southern Africa, southern Australia and northern high latitudes (Extended Data Fig. 1a–d). Temporally, more than half of global land regions exhibit upward trends in both duration and ratio for all three drought types based on ERA5-Land (duration: 60.6%–69.3% of land regions; ratio: 53.6%–61.3%). Southwestern North America, South America, Africa, mid-latitude Eurasia, southeastern Asia and Australia emerge as six major hot spots where all three drought types have increased (Fig. 2c,d and Extended Data Fig. 1e–h), consistent with previous findings for surface and root-zone soil moisture droughts^18,19. Seven additional datasets consistently support the increases in vertically compound drought in these hotspots (hatched areas in Fig. 2c,d; except central Africa). Vertically compound drought also exhibits a clear biome dependence (Supplementary Note 1). Although the spatial coverage is limited, 105 in situ stations with multilayer soil moisture observations support the widespread occurrence of vertically compound drought (Extended Data Fig. 2).

Fig. 2. Climatological means and trends in vertically compound drought during the warm seasons from 1981 to 2020.

a, Spatial distribution of climatological mean annual duration of vertically compound drought based on ERA5-Land. The accompanying histogram shows the area-weighted mean duration for different vegetation types based on ensemble mean across three reanalysis datasets and five land-surface model outputs (n = 8). White diamonds denote individual dataset results. b, Spatial distribution of the annual ratio of vertically compound drought to all drought days. c,d, Same as a and b but for the trends in annual duration (c) and ratio (d). Hatching shows regions where more than five additional datasets agree with ERA5-Land on the sign of change, and stippling indicates statistically significant trends at the 0.05 significance level (one-sided Mann–Kendall test). Results in regions without hatching should be interpreted with greater caution because the inter-dataset spread is larger. The accompanying histograms show the percentage of each vegetation type within regions showing positive trends (n = 8). Basemap data from Natural Earth (https://www.naturalearthdata.com).

Extended Data Fig. 1. Climatological means and trends in other drought types during the warm seasons from 1981 to 2020.

a–d Spatial distributions of climatological mean of annual duration (a, b) and ratio (c, d) of surface-layer drought and deep-layer drought from ERA5-Land. e–h Same as a–d, but for the trends in annual duration (e, f) and ratio (g, h) of surface-layer drought and deep-layer drought. Hatching shows regions where more than five additional datasets agree with ERA5-Land on the sign of change, and stippling indicates statistically significant trends at the 0.05 significance level (one-sided Mann-Kendall test). i Same as e, but for profile-average drought. j Relative difference in annual duration between profile-average drought and vertically compound drought. Stippling indicates statistically significant difference in annual duration between two drought types at the 0.05 significance level (two-sided Student’s t test). Basemap data from Natural Earth (https://www.naturalearthdata.com).

Extended Data Fig. 2. Observed vertically compound drought.

a The same as Fig. 1b. Each dot represents a multilayer soil moisture station from ISMN and OzNet. b, c Spatial distributions of multi-year mean of annual duration (b) and ratio (c) of observed vertically compound drought. d Heatmap show the temporal evolutions of all drought types across the entire soil profile for a station over the boreal forest (label “s1” in b) during the 2014 warm season based on ERA5-Land. Drought intensity is defined as the difference between drought threshold and soil moisture. e–i Same as d, but for stations with labels “s2” over the tropical forest, “s3” over the austral forest, “s4” over the cropland, “s5” over the grassland, and “s6” over the shrubland. Basemap data in a–c from Natural Earth (https://www.naturalearthdata.com).

We also identified root-zone soil moisture drought^18,20,21, treated here as profile-average drought, based on soil moisture over 0–100 cm during the warm seasons of 1981–2020. Regions with increasing profile-average drought broadly coincide with those where vertically compound drought increases (Extended Data Fig. 1i). At the global scale, however, the annual duration of profile-average drought (13.7 $\pm$ 1.2 days) is 2.3 times that of vertically compound drought (Extended Data Fig. 1j and Supplementary Fig. 1). A regional case indicates that during the 2020 warm-season profile-average drought in eastern Asian forests, more than half of the event duration featured moist surface layer while middle and/or deep layers were already dry (Fig. 1c,d). Extending this analysis to the global scale, 67.8% $\pm$ 2.9% of profile-average drought days not classified as vertically compound drought occurred when the surface layer remained moist while the middle and/or deep layers exhibited moisture deficits, indicating that this asymmetry contributes substantially to the discrepancy between profile-average and vertically compound droughts. Together, these findings indicate that diagnosing drought solely from entire-profile mean soil moisture overestimates the duration of true soil moisture drought throughout the entire profile by about 128.3%, because a substantial fraction of these events represent mixed wet–dry vertical states rather than genuine vertically compound drought.

Physical mechanisms behind vertically compound drought

To reveal the physical mechanisms driving vertically compound droughts, we compared the hydrometeorological conditions and land-surface energy budgets among different drought types over identified hotspot regions during the warm seasons of 1981–2020 (Fig. 3 and Extended Data Fig. 3). Vertically compound droughts typically arise through two archetypal evolutionary pathways: ‘top-down’ and ‘bottom-up’ (Fig. 1c,d). In the top-down pathway, an initial surface-layer drought gradually propagates into deeper soils (June 2020 in Fig. 1c,d). Surface-layer drought develops when low precipitation (composite anomaly: −1.6 mm d⁻¹) and high atmospheric evaporative demand (that is, vapour pressure deficit, +4.5 hPa) first deplete surface-layer soil moisture^22,23. Reduced cloud cover (−0.2%) and enhanced solar radiation (+31.9 W m⁻²) increase the net surface energy input, which is partitioned into sensible and latent heat, thereby warming the surface-layer soil (+2.2 °C) and near-surface air (+2.6 °C). Because deeper layers remain moist at this stage, evapotranspiration is still sustained, and part of the excess energy is released as latent heat, so that soil moisture deficits (-0.05 m³ m⁻³) remain concentrated in the surface layer. With continued enhanced solar radiation (+24.2 W m⁻²), heat accumulates within the soil and propagates downward, warming the entire soil profile (+2.7 °C). This maintains soil evaporation and plant transpiration, accelerating moisture depletion throughout the profile. Meanwhile, dry surface-layer soil (-0.07 m³ m⁻³) and elevated air temperature (+2.5 °C) strengthen land–atmosphere coupling (+1.9). Under this enhanced land–atmosphere interaction²⁴, atmospheric evaporative demand is further amplified (+4.7 hPa), drawing more water from deeper soils into the near-surface air²⁵. Together, these processes drive the downward penetration of surface-layer drought and eventually produce vertically compound drought, characterized by extremely low latent heat flux (−15.0 W m⁻²) and evaporative fraction (−20.0%), indicating severe water stress across all soil layers (middle and deep layers: −0.06 and −0.05 m³ m⁻³).

Fig. 3. Physical mechanisms behind vertically compound drought during the warm seasons from 1981 to 2020.

Area-weighted mean maximum near-surface 2-m air temperature anomaly, cloud fraction anomaly, precipitation anomaly, solar net shortwave radiation anomaly, vapour pressure deficit anomaly, surface sensible heat flux anomaly, surface latent heat flux anomaly, evaporative fraction anomaly, soil temperature anomaly across three layers, soil moisture anomaly across three layers and land–atmosphere coupling intensity during surface-layer droughts, vertically compound droughts and deep-layer droughts over identified hotspots of soil moisture droughts (grey regions in Extended Data Fig. 4), respectively, based on the ERA5-Land and ERA5 datasets. Brown colour, dry soil; blue colour, wet soil.

Extended Data Fig. 3. Hydrometeorological drivers of soil moisture drought during the warm seasons from 1981 to 2020.

a–q Boxplots show land-atmosphere coupling intensity, evaporative fraction anomaly, maximum near-surface 2-m air temperature anomaly, soil temperature anomaly (three soil layers and entire-profile mean), soil moisture anomaly (three soil layers and entire-profile mean), solar net shortwave radiation anomaly, land sensible heat flux anomaly, land latent heat flux anomaly, precipitation anomaly, cloud fraction anomaly, and vapor pressure deficit anomaly during four drought types over hotspots of soil moisture droughts, respectively (n = 8,090). In the boxplots, the upper whisker represents Q3 + 1.5 × IQR, followed by the third quartile (Q3), the median (horizontal line), the first quartile (Q1), and the lower whisker representing Q1 – 1.5 × IQR, where IQR is the interquartile range (Q3 – Q1). r Grey shading areas show hotspots of soil moisture droughts. Basemap data in r from Natural Earth (https://www.naturalearthdata.com).

In the bottom-up pathway, vertically compound drought emerges from pre-existing deep-layer drought (August 2020 in Fig. 1c,d). Deep-layer drought can persist or newly form when cloud cover ( + 0.02%) and precipitation ( + 0.2 mm d⁻¹) recover but mainly replenish the surface-layer soil (−0.003 m³ m⁻³). In this stage, surface-layer soil moisture deficits are interrupted and land–atmosphere coupling weakens ( + 0.4), whereas deep-layer soil recovers much more slowly and remains dry (−0.04 m³ m⁻³). When deep soils remain dry, the recharged surface layer becomes highly vulnerable to renewed atmospheric demand. As air temperature increases again, land–atmosphere coupling is re-intensified and evaporative demand rises so that evaporation and plant water uptake rapidly deplete the remaining surface soil moisture. Surface soil moisture deficits then reconnect with the persistent deep-layer drought, converting an initially deep-layer drought into a vertically compound drought. Globally, bottom-up evolution is the dominant pathway of vertically compound drought formation, although its dominance weakens in wet regions and forests, where top-down development becomes relatively more important (Supplementary Note 2). Moreover, antecedent vegetation growth can also precondition subsequent drought^26,27 (Extended Data Fig. 4 and Supplementary Note 3).

Extended Data Fig. 4. Effect of antecedent vegetation growth on soil moisture drought during the warm seasons from 1981 to 2020.

a Map shows the differences in entire-profile mean soil moisture predicted by the full model (including both antecedent leaf area index and soil moisture) relative to the baseline model (including antecedent soil moisture) during vegetation negative effects on vertically compound droughts. Areas with data unavailability are filled with white. Hatching indicates regions where the constructed models show sufficiently high predictive skill and where lagged dependency is confirmed. b, c Same as a, but for surface-layer drought and deep-layer drought. d histograms show the area-weighted mean vegetation lagged effect for different drought types across forests and croplands. Basemap data in a–c from Natural Earth (https://www.naturalearthdata.com).

Synthesizing the above analyses, vertically compound droughts are jointly modulated by multiple hydrometeorological and land-surface factors. We attributed their occurrence and development to five key drivers: reduced precipitation supply, enhanced solar radiative heating and its promotion of evaporation, increased atmospheric dryness and associated increases in evaporative demand, additional evapotranspiration induced by strengthened land–atmosphere coupling and water consumption from antecedent vegetation growth. Given the strong collinearity among these drivers, we employed ridge regression to quantitatively separate their relative contributions to vertically compound droughts (Methods). The results show that land–atmosphere coupling provides the dominant contribution over the globe, followed by antecedent vegetation growth (Extended Data Fig. 5). Land–atmosphere coupling accounts on average for 36.5% of entire-profile soil moisture deficits during vertically compound droughts across identified hotspots, while antecedent vegetation growth contributes 20.0%. The land–atmosphere coupling, reflecting the loop between soil moisture and air temperature, emerges as the key process sustaining vertically compound droughts, while antecedent vegetation growth further amplifies moisture deficits, particularly in deeper soils.

Extended Data Fig. 5. Relative contribution of hydrometeorological drivers to vertically compound droughts during the warm seasons from 1981 to 2020.

a–e Spatial distribution of relative contribution of precipitation deficit (a), solar radiation (b), atmospheric aridity (c), antecedent leaf area index (d), and land-atmosphere coupling (e) to entire-profile soil moisture deficits during vertically compound droughts of 1981–2020 based on ridge regression model. Precipitation deficit is defined as the difference between climatological mean and precipitation (that is, negative anomaly). Soil moisture deficits (that is, drought intensity) are defined as the difference between drought threshold and soil moisture. Atmospheric aridity is defined as vapor pressure deficit anomaly. Antecedent leaf area index is defined as the average leaf area index over the 30 days preceding vertically compound droughts. f Boxplots show relative contributions of different drivers during vertically compound droughts over hotspots of soil moisture droughts, respectively (n = 8,090). In the boxplots, the upper whisker represents Q3 + 1.5 × IQR, followed by the third quartile (Q3), the median (horizontal line), the first quartile (Q1), and the lower whisker representing Q1 – 1.5 × IQR, where IQR is the interquartile range (Q3 – Q1). White diamonds denote area-weighted average of relative contributions. Basemap data in a–e from Natural Earth (https://www.naturalearthdata.com).

Ecological impacts of vertically compound drought

We quantified the ecosystem impacts of vertically compound droughts by estimating the associated anomalies in carbon uptake (that is, gross primary productivity, GPP) (Fig. 4a–c and Extended Data Fig. 6), expressed as mean daily GPP loss during drought days. On the global average based on ERA5-Land, vertically compound droughts cause the largest GPP losses (0.23 g C m⁻² d⁻¹; ensemble mean of GPP products), followed by surface-layer droughts (0.12 g C m⁻² d⁻¹), whereas deep-layer droughts exert comparatively weaker effects (0.04 g C m⁻² d⁻¹). This pattern indicates that once surface-layer soil moisture deficits have developed, additional moisture deficits in deeper layers substantially amplify vegetation stress. By contrast, the mean GPP loss associated with profile-average droughts identified solely from entire-profile mean soil moisture is only 0.18 g C m⁻²d⁻¹, about 27.8% lower than that of vertically compound droughts. We considered potential selection effects arising from differences in atmospheric evaporative demand and energy forcing by comparing the two drought types under matched vapour pressure deficit and radiation anomalies (Supplementary Note 4).

Fig. 4. Ecological losses of vertically compound drought during the warm seasons from 1981 to 2020.

a, GPP anomaly with a 2-day lag under vertically compound droughts based on ensemble mean of GPP products. Areas with data unavailability are filled with white. The accompanying histogram shows the area-weighted sum of GPP anomalies across different vegetation types. b,c, Same as in a but for surface-layer droughts (b) and deep-layer droughts (c), respectively. d, Relative importance of hydrometeorological, climatic, vegetated and edaphic variables in controlling spatial variations of GPP anomalies across the forests (green) and croplands (yellow) during vertically compound droughts estimated using random forest model and Shapley (SHAP) value. e–j, Partial dependence plots of the three most important predictors (n = 5,641 for forest (e–g) and n = 2,544 for cropland (h–j)): latent heat (e,i), mean temperature (f), mean precipitation (h) and precipitation frequency (g,j). The y axes shows Shapley values of the corresponding predictors (x axes). Solid lines represent the best-fit relationships, and shaded areas represent the corresponding 95% confidence interval. The partial dependence plots indicate the marginal effects of individual variables on the response, independent of other variables. Basemap data in a–c from Natural Earth (https://www.naturalearthdata.com).

Extended Data Fig. 6. Ecological losses of different drought types during the warm seasons from 1981 to 2020.

a–d GPP anomalies associated with vertically compound droughts from 1981 to 2020 based on the ensemble mean of GPP products. Areas with data unavailability are filled with white. The accompanying histograms show the area-weighted sum GPP anomalies across different vegetation types (n = 2), with red diamond representing results from individual product. e–x Same as a–d but for GPP anomalies with a 1-day delay (e–h), 2-day delay (i–l), 3-day delay (m–p), 4-day delay (q–t), and 5-day delay (u–x) under four drought types, respectively. Basemap data from Natural Earth (https://www.naturalearthdata.com).

Spatially, southeastern North America, southern and eastern South America, southern Africa, northern Australia and mid- to low-latitude Eurasia have experienced pronounced GPP losses associated with vertically compound droughts (Fig. 4a). Forests and croplands are the two ecosystem types most strongly affected, with total GPP losses reaching 9.8 $\times$ 10¹²g C d⁻¹ and 8.0 $\times$ 10¹² g C d⁻¹, respectively. In forests, these GPP losses directly threaten the stability of the terrestrial carbon sink, whereas in croplands, these losses translate into risks for regional food security. To further reveal the geographic controls on GPP losses associated with vertically compound droughts, we applied random forest models and Shapley analysis to characterize the hydrometeorological, climatic, vegetated and edaphic conditions underlying the spatial variations in GPP losses over forests and croplands (Methods). Hydrometeorological factors represent environmental conditions during vertically compound droughts, climatic factors represent the warm-season climate background and vegetated and edaphic factors represent long-term land-surface properties. The fitted random forest models explain 86% and 78% of the spatial variance in GPP anomalies for forests and croplands, respectively.

Overall, for forest GPP losses, latent heat flux is identified as the leading factor, accounting for 33.0% of the total importance, followed by warm-season mean temperature and precipitation frequency (Fig. 4d). Shapley value analysis shows that forests with smaller latent heat flux exhibit larger GPP losses (Fig. 4e). Reduced latent heat flux reflects stronger water limitation and more severe soil moisture deficits during vertically compound droughts, rather than acting as an independent driver of GPP decline. We noted that forest GPP losses increase with the local warm-season mean temperature up to about 15 °C, beyond which this sensitivity levels off (Fig. 4f). This pattern suggests a transition from colder, more energy-limited boreal and temperate forests to warmer forests in which drought impacts on GPP remain high but no longer strengthen as clearly with further increases in local warm-season mean temperature²⁸. Warm-season precipitation frequency is negatively correlated with GPP losses (Fig. 4g), implying that more frequent precipitation events can interrupt the development of vertically compound droughts and thereby mitigate damage to vegetation structure and function²⁹.

In cropland ecosystems, warm-season mean precipitation is the most important control on GPP losses (contribution 26.8%), followed by latent heat flux and precipitation frequency (Fig. 4d). Despite the similar partial dependence on latent heat flux and precipitation frequency between forests and croplands, warm-season mean precipitation reveals contrasting responses of rain-fed croplands to vertically compound drought across precipitation regimes (Fig. 4h–j). In water-limited croplands (mean precipitation <2 mm d⁻¹), slightly higher precipitation supports a higher GPP baseline and stronger water use (Supplementary Fig. 2), so once vertically compound droughts occur, the absolute GPP loss becomes larger³⁰. In contrast, in humid croplands (mean precipitation >2 mm d⁻¹), higher mean precipitation implies a stronger drought buffering capacity³¹, thereby weakening impacts of vertically compound droughts on vegetation. These different responses between forests and croplands partly reveal fundamental differences between natural forest ecosystems and managed agroecosystems in water use strategies and drought vulnerability.

Anthropogenic influences on vertically compound drought

We further used 118 ensemble members from the Coupled Model Intercomparison Project Phase 6 (CMIP6) to assess the response of long-term changes in vertically compound drought to anthropogenic forcings and to project its future evolution. We further quantified long-term trends in duration and ratio of vertically compound drought over 1981–2020 across these hotspots (Fig. 5a,b). ERA5-Land shows statistically significant (p < 0.01) increases in both duration and ratio, with trends of about 1.5 days per decade and 1.4% per decade, respectively. These trends lie outside the 95th percentile of those derived from 10,000 overlapping 40-year segments based on CMIP6 pre-industrial control simulations (Fig. 5c,d), indicating that the observed increase in vertically compound drought cannot be explained by internal climate variability alone. Both methods suggest that anthropogenic forcings are very likely (>90%) responsible for the increase in vertically compound droughts during the warm seasons of 1981 to 2020 (Supplementary Note 5).

Fig. 5. Detection and attribution of the evolution of vertically compound drought during the warm seasons from 1981 to 2020.

a,b, Temporal changes in anomalies of duration (a) and ratio (b) of vertically compound drought within identified hotspots. The black line represents the ERA5-Land, while the red line shows the ensemble mean of CMIP6 historical simulations (HIST). Shaded areas indicate the ±1 ensemble standard deviation range. c,d, The histograms and probability density functions of the trends in these anomalies of duration (c) and ratio (d) based on 10,000 overlapping 40-year segments from the pre-industrial control (PiControl) simulations (n = 10,000). Vertical black and red lines represent the trends from ERA5-Land and CMIP6 HIST ensemble mean, respectively. The left inset box includes the optimized scaling factors and their 5–95% confidence intervals from the regularized optimal fingerprint method where the anomalies from ERA5-Land are regressed by the anomalies from CMIP6 HIST ensemble mean. The grey error bands in the right inset box represents 5–95% and 20–80% confidence intervals of the distribution of Spearman correlation coefficients for temporal changes in these anomalies between the CMIP6 HIST ensemble mean and 10,000 overlapping 40-year segments from PiControl simulations (n = 10,000). The red point represents the Spearman correlation coefficient between the ERA5-Land and CMIP6 HIST ensemble mean. e–h, Same as Fig. 2c,d but based on the ensemble mean of CMIP6 HIST simulations during 1981–2020 (e,f) and SSP5–8.5 projections during 2061–2100 (g,h). Hatching shows regions where more than 70% of the models agree on the sign of change, and stippling indicates statistically significant trends at the 0.05 significance level (one-sided Mann–Kendall test). Basemap data in e–h from Natural Earth (https://www.naturalearthdata.com).

Continued anthropogenic forcings are projected to further intensify vertically compound droughts. Under the high-emissions scenario (SSP5–8.5, where SSP stands for Shared Socioeconomic Pathway) during 2061–2100, the globally averaged annual duration and ratio of vertically compound droughts are projected to increase from historical values of 5.8 $\pm$ 1.8 days and 23.8% $\pm$ 9.1% to 16.3 $\pm$ 5.7 days and 36.5 $\pm$ 10.6%, respectively. The ensemble mean shows that hotspots of soil moisture droughts identified in the historical period retain the same trend direction in future projections, but the magnitudes of the trends in duration (1.8 days per decade) and ratio (1.2% per decade) increase by about 125.0% and 9.1%, respectively, compared with the historical period (Fig. 5e–h). Moreover, these increases expand spatially, with regions exhibiting increasing duration and ratio extending to about 82.3% and 77.3% of global land regions from about 75.6% and 72.3% historically, respectively, particularly in South America, Africa and northern high latitudes. Similar intensification of vertically compound drought is also evident under the moderate-emissions scenario (SSP2–4.5; Supplementary Fig. 3), indicating that this pattern is robust across emission pathways. Together, these projections indicate not only that vertically compound droughts themselves will become more widespread but also that their proportion among all soil moisture droughts will rise substantially.

Discussion

Vertically compound drought highlights the importance of the vertical soil moisture structure and provides a complementary perspective to conventional surface- or root-zone drought metrics. Rather than simply refining existing drought definitions, it reveals that the co-occurrence of moisture deficits across multiple soil layers constitutes a distinct drought state with additional ecological risk. This state is often obscured by profile averaging, yet it reflects the loss of hydrological buffering and ecological refuge space within the soil profile. By retaining vertical structural information, vertically compound drought distinguishes states with partial interlayer compensation (that is, mixed wet–dry profiles) from those involving whole-profile moisture deficits. This distinction is critical for drought characterization and risk assessment because the latter state is more likely to involve persistent water limitation, intensified land–atmosphere coupling and amplified ecosystem impacts. Data uncertainties and future monitoring and adaptation needs for vertically compound droughts (Supplementary Note 6). Accordingly, identifying vertically compound drought deepens our understanding of the mechanisms and ecological consequences of soil moisture drought evolution, while also providing a stronger basis for assessing climate-extreme impacts on ecosystems and improving simulations of multilayer soil moisture and carbon-cycle dynamics.

Methods

Gridded datasets

The ERA5-Land reanalysis produced by the European Center for Medium-Range Weather Forecast (ECMWF) is a land-focused enhancement of the fifth-generation European Reanalysis (ERA5)³², which offers hourly multilayer soil moisture at depths of 0–7 cm, 7–28 cm, 28–100 cm and 100–289 cm, multilayer soil temperature, near-surface 2-m air temperature, near-surface 2-m dewpoint temperature, surface sensible heat flux, surface latent heat flux, solar net shortwave radiation, precipitation, actual evapotranspiration and potential evapotranspiration. ERA5-Land is generated by driving the state-of-the-art land-surface model, that is, Carbon Hydrology-Tiled ECMWF Scheme for Surface Exchanges over Land (CHTESSEL) with a downscaled version of the ERA5 dataset³³. ERA5-Land 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)¹³. Given that ERA5-Land assimilates a richer set of observations and employs a more advanced land-surface scheme, its soil moisture estimates are considered more reliable^17,34. We therefore base our main analyses in this study primarily on the ERA5-Land results. Given the variations in the quantity and quality of observations assimilated into ERA5-Land, particularly the incorporation of satellite data since 1979, we selected the period from 1981 to 2020 for our analysis.

To ensure the robustness of ERA5-Land results, we validated them against two reanalysis and five land-surface model outputs: (1) the China Meteorological Administration global land-surface reanalysis dataset (CRA-Land)³⁵, which provides 3-hourly soil moisture at depths of 0–10 cm, 10–40 cm, 40–100 cm and 100–200 cm during 1981–2020. (2) The Noah Land Surface Model of the National Aeronautics and Space Administration Global Land Data Assimilation System (GLDAS-Noah) version 2.0 (1981–2014) and 2.1 (2000–2020)³⁶, which provides 3-hourly soil moisture at depths of 0–10 cm, 10–40 cm, 40–100 cm and 100–200 cm. To ensure consistency, we integrated GLDAS v2.0 and v2.1 by employing the common period of 2000–2014. The cumulative distribution function matching method was used to correct grid-by-grid biases in soil moisture¹⁷. (3) The Inter-Sectoral Impact Model Intercomparison Project simulation round 3 A (ISIMIP3A)³⁷, which provides land-surface model outputs of global daily soil moisture. Two global land-surface models (MIROC-INTEG-LAND³⁸ and ORCHIDEE-MICT³⁹) under the framework of ISIMIP3A were driven by bias-corrected and downscaled atmospheric forcings, providing five combination outputs (Supplementary Table 2). All ISIMIP3A simulations considered water consumption sectors (for irrigation, domestic and industrial purposes), reservoir management and land-use change under 1,901 social and economic scenarios run.

We collected other variables: (1) hourly total cloud cover from ERA5 during 1981–2020³². (2) Daily gross primary production (GPP) from the FLUXCOM ensemble (1981–2020)⁴⁰, which was produced through three machine learning methods (artificial neural network, random forest and multivariate adaptive regression) driven by global eddy covariance carbon flux tower measurements and meteorological forcings⁴¹. Two FLUXCOM ensemble driven by different meteorological forcings, that is, ERA5 and the collection of the University of East Anglia Climatic Research Unit and Japanese Reanalysis (CRUJRA), were obtained. (3) Daily GPP from the FLUXSAT (2001–2020), which was derived from the MODerate-resolution Imaging Spectroradiometer (MODIS) instruments on National Aeronautics and Space Administration Terra and Aqua satellites⁴². GPP from FLUXSAT is estimated using the MODIS Nadir Bidirectional Reflectance Distribution Function-Adjusted Reflectances product as input to neural network models that globally upscale GPP estimated from selected collocated FLUXNET2015 and OneFlux eddy covariance tower sites used for model training. (4) Hourly GPP from the RTL-LUE (2001–2020), which was constructed using a modified radiation scalar two-leaf light-use efficiency model and integrates inputs including downward shortwave radiation, dewpoint temperature and air temperature from ERA5-Land, leaf area index from Global Land Surface Satellite, land cover from MODIS and atmospheric CO₂ concentration from National Oceanic and Atmospheric Administration⁴³. (5) Leaf area index (LAI) from GLOBMAP (1981–2000, half month; 2001–2020, 8 day) created by fusion of MODIS and historical advanced very high-resolution radiometer data⁴⁴. (6) Contiguous solar-induced fluorescence (CSIF) created by neural network with surface reflectance from the MODIS and SIF from the Orbiting Carbon Observatory-2⁴⁵. (7) Rooting depth estimated by microwave vegetation optical depth from the advanced microwave scanning radiometer⁴⁶. (8) Clay fraction, sand fraction and silt fraction from the global gridded soil information (SoilGrids) version 2.0⁷². (9) Land-cover type from the United States Geological Survey (USGS); (10) observed precipitation from the Global Soil Wetness Project Phase 3 (GSWP-3)⁴⁷ to identify the desert regions where climatological annual precipitation is below 100 mm (ref. ⁴⁸). (11) Observed precipitation and potential evapotranspiration from the Climatic Research Unit gridded time series version 4.08 (CRU TS4.08) to identify the dry and wet regions⁴⁹. Dry and wet regions were identified using the aridity index that was calculated based on the ratio between annual precipitation and potential evapotranspiration from CRU during 1981–2010. Regions were classified as dry regions when the aridity index $\le$0.65 and wet regions when the aridity index >0.65 (ref. ⁵⁰).

We also used the Coupled Intercomparison Project Phase 6 (CMIP6) simulations covering the historical climate⁵¹ (1981–2014; historical in CMIP, HIST) and future emissions scenarios⁵² (2015–2100; ssp245 and ssp585 in Scenario Model Intercomparison Project, SSP2–4.5 and SSP5–8.5). We extended the historical simulations to 2020 by combining the simulations from SSP5–8.5^53,54. The unforced pre-industrial control (PiControl) experiments were used to quantify the internal climate variability. For our analysis, we selected 13 models that run 118 ensemble members under all above forcings and scenarios and output daily multilayer soil moisture (Supplementary Table 3). For the above datasets, hourly variables were daily averaged and weekly or semi-monthly variables were linearly interpolated. All variables then were remapped to 1° × 1° horizontal resolution (bilinear for temperatures, second-order conservative for fluxes). Details of the above datasets refer to Supplementary Table 1.

In situ observations

We collected station-based observations of multilayer soil moisture. These station-based observations were obtained from two sources: (1) the Australia hydrological monitoring network (OzNet)⁵⁵, which provides multilayer (0–30 cm, 30–60 cm and 60–90 cm) soil moisture data at 20-min intervals during 1986–2023 and (2) the International Soil Moisture Network (ISMN)⁵⁶, which collects and harmonizes soil moisture datasets from global networks during 2002–2023. Given the differences in the number and depth of layers provided by different stations, we applied the following criteria for station selection: (1) availability of multilayer soil moisture data; (2) a maximum measured depth of at least 90 cm; (3) data timesteps ranging from minutes to daily; (4) with at least 10 years in total during which the missing rate in the warm seasons is less than 20% and (5) ‘good quality’ labels from ISMN quality flag. We eventually selected 105 stations, including 79 in North America, 22 in Australia, 1 in Europe and 3 in Africa (Extended Data Fig. 2 and Supplementary Table 4).

Identification of vertically compound droughts

To ensure consistency across different soil depths, multilayer soil moisture was interpolated onto a common profile (0–10 cm, 10–40 cm and 40–100 cm) based on total-water-conserving method^17,57. For each layer, a local drought day is identified when soil moisture falls below the calendar-day tenth percentile during warm seasons of 1981–2020 (climatological period)⁵⁸. The warm season is defined as May to September for the Northern Hemisphere and November to March for the Southern Hemisphere¹⁷. A 15-day moving average was applied to reduce high-frequency noise before threshold estimation⁵⁹. The tenth percentile threshold is recommended by the United States Drought Monitor as part of its ‘Drought classification-percentile range for most indicators’ and corresponds to its ‘D2-Severe Drought’ category⁶⁰.

On the basis of all combinations of local drought occurrence across the three layers (0–10 cm, 10–40 cm and 40–100 cm), warm-season drought days were classified into seven drought types (Fig. 1a). Vertically compound drought is identified when drought occurs simultaneously in all three layers; surface-layer drought is identified when drought occurs in the first layer (0–10 cm) and deep-layer drought is identified when drought occurs in the third layer (40–100 cm). For each drought type, we calculated the duration (unit: days) defined as the cumulative number of days on which that drought type occurs and the ratio (%) defined as the proportion of days with that drought type relative to the total number of days with any drought type. Moreover, we identified profile-average drought based on entire-profile mean soil moisture (0–100 cm). In this study, significance tests for trends in duration and ratio of droughts were performed using the Mann–Kendall test with pre-whitening, which accounts for temporal autocorrelation commonly present in soil-moisture-related time series⁶¹. Sensitivity tests for vertically compound drought identification are provided in Supplementary Note 7.

Composite analysis for vertically compound droughts

To compare the hydrometeorological conditions and land-surface energy budgets among different drought types, we analysed composite maps of land–atmosphere coupling intensity, evaporative fraction anomaly, maximum air temperature anomaly, soil temperature anomaly (three soil layers and entire-profile mean), soil moisture anomaly (three soil layers and entire-profile mean), solar net shortwave radiation anomaly, sensible heat flux anomaly, latent heat flux anomaly, precipitation anomaly, cloud fraction anomaly and vapour pressure deficit (VPD) anomaly for each drought types based on ERA5-Land (Fig. 3 and Extended Data Fig. 3). Because land–atmosphere coupling intensity, evaporative fraction and VPD are not available in the ERA5-Land dataset, we estimated these three variables: (1) the land–atmosphere coupling intensity ($\mathrm{\pi }$ metric; unitless) is used to quantitatively describe land–atmosphere interactions, defined as⁶²:

$$\mathrm{\pi }=\left[{\left(R_{\mathrm{n}}-\lambda E\right)}^{\prime }-{\left(R_{\mathrm{n}}-\lambda E_{\mathrm{p}}\right)}^{\prime }\right]\times T^{\prime }$$ 1

where apostrophe denotes standardized anomalies, $R_{\mathrm{n}}$ is solar net shortwave radiation, $T$ is near-surface 2-m temperature, $E$ and $E_{\mathrm{p}}$ denote actual and potential evapotranspiration, respectively, and $\lambda$ is the latent heat of vaporization. When the great potential of soil moisture to affect air temperature concurs with an anomalously high air temperature, the soil moisture associated energy balance is believed to be at play as expressed by a large land–atmosphere coupling intensity that thereby indicates stronger soil moisture-temperature coupling conditions; (2) the evaporative fraction is defined as the ratio between the latent heat flux and the available energy as follows¹³:

$$\mathrm{EF}=\mathrm{SLHF}/\left(\mathrm{SLHF}+\mathrm{SSHF}\right)$$ 2

where $\mathrm{EF}$ represents the evaporative fraction, $\mathrm{SLHF}$ denotes surface latent heat flux and $\mathrm{SSHF}$ means surface sensible heat flux and (3) the VPD (unit: hPa) is defined as the difference between saturated vapour pressure ($E_{\mathrm{s}}$; unit: hPa) and actual vapour pressure ($E_{\mathrm{a}}$; unit: hPa) as follows¹³:

$$E_{\mathrm{s}}=6.11\times \exp \left(\frac{17.67\times T}{243.5+T}\right)$$ 3

$$E_{\mathrm{a}}=6.11\times \exp \left(\frac{17.67\times T_{\mathrm{d}}}{243.5+T_{\mathrm{d}}}\right)$$ 4

where $T_{\mathit{d}}$ is near-surface 2-m dewpoint temperature, respectively. Moreover, we performed composite analyses of GPP anomalies associated with each drought type. Because ecosystem responses to drought are not instantaneous (Fig. 1d), we calculated GPP anomalies at lags of 0–5 days relative to the occurrence of each drought type (Extended Data Fig. 6). We found that GPP generally declines at a 2-day lag across the globe, that the spatial pattern of GPP anomalies changes little beyond a 2-day lag and that GPP anomalies at different lags exhibit high spatial correlation (r: 0.79–0.99; p < 0.01). We therefore used the 2-day lag as a representative measure of drought-induced GPP loss. The 2-day lag is used here as a representative value for consistent comparison, rather than as a universal response time, because biome-specific differences may exist. The overall consistency across FLUXCOM products driven by different forcings, alternative drought-identification datasets and additional GPP products lends strong support to the robustness of our main conclusions (Supplementary Note 8).

Lagged dependency for antecedent vegetation activity

To assess whether antecedent vegetation growth affects the occurrence of different drought types, we quantified the lagged effects of vegetation on soil moisture droughts within a lagged dependency framework⁶³. To better capture the vegetation–land–atmosphere interactions, we implemented this framework using nonlinear random forest regression rather than traditional linear models⁶⁴. Specifically, for each drought type, we built two random forest models to predict entire-profile mean soil moisture on drought days. The baseline model was driven solely by antecedent entire-profile mean soil moisture, thus representing the combined effects of antecedent atmospheric forcing and soil moisture memory. In this model, the predictand was the normalized entire-profile mean soil moisture on the drought day, and the predictors were the time series of normalized entire-profile mean soil moisture over the preceding 30 days. The full model additionally incorporated antecedent vegetation growth by including both normalized entire-profile mean soil moisture and normalized LAI over the preceding 30 days as predictors.

Model performance was evaluated using tenfold out-of-sample cross validation to avoid overfitting. For each grid cell and drought type, we considered antecedent vegetation growth to exert a lagged dependency if both models achieved a greater coefficient of determination than 0.1 and the full model outperformed the baseline model⁶⁵. The difference between the soil moisture predictions from the full and baseline models on drought days was then interpreted as the lagged effect of antecedent vegetation growth on soil moisture deficits.

Drivers for vertically compound droughts

We performed the ridge regression to estimate contributions of different factors to vertically compound droughts, especially when explanatory variables exhibit high interdependencies. To mitigate potential instability arising from interdependencies among predictors, we applied ridge regression by introducing a regularization term to the standard least-squares cost function. This regularization is controlled by a tuning parameter (λ). The objective function is expressed as:

$${\beta }^{\wedge }={\displaystyle \sum _{\mathrm{i}=1}^{\mathrm{n}}}{\left(\mathrm{y}-{\beta }_0-\sum {\beta }_{\mathrm{i}}{\mathrm{x}}_{\mathrm{i}}\right)}^2+\lambda \sum {\beta }^2$$ 5

where ${\beta }^{\wedge }$ denotes the estimated regression coefficients, y is the dependent variable, ${\beta }_0$ is the intercept and ${\beta }_{\mathrm{i}}$ is the coefficient corresponding to the independent variable x_i. The ridge tuning parameter (λ) was determined iteratively, starting from 0.01 and increasing in increments of 0.01 until the variance inflation factor of all predictors was reduced below 3 (ref. ⁶⁶).

In the ridge regression model, the soil moisture deficits (the difference between drought threshold and soil moisture) were normalized as the dependent variable, and precipitation anomaly, solar net shortwave radiation anomaly, VPD anomaly, land–atmosphere coupling intensity and the average LAI over the 30 days preceding the droughts were normalized as the independent variables. Finally, the relative contribution (${\eta }_{\mathrm{j}}$) of each independent variable $j$ to the dependent variable is estimated as follows⁶⁶:

$${\eta }_j=\frac{\left|{\beta }_j^{\wedge }\right|}{\sum _{i=1}^5\left|{\beta }_i^{\wedge }\right|}$$ 6

where ${\beta }^{\wedge }$ denotes the estimated regression coefficients. Our attribution results were robust across sensitivity tests and were further supported by alternative regression approaches (Supplementary Note 9).

Drivers for spatial variation in GPP losses

For each ecosystem (forests and croplands), we built a random forest model to identify the factors (hydrometeorological, climatic, vegetated and edaphic conditions; Supplementary Table 5) that contribute the most to the geographic variation in GPP losses associated with vertically compound droughts. Hydrometeorological factors represent environmental conditions during vertically compound droughts, including land–atmosphere coupling intensity, evaporative fraction anomaly, entire-profile soil temperature anomaly, solar radiation anomaly, sensible heat flux anomaly, latent heat flux anomaly, precipitation anomaly, cloud fraction anomaly and vapour pressure deficit anomaly; climatic factors represent the warm-season climate background, including climatological mean of warm-season mean 2-m air temperature, warm-season mean precipitation, warm-season precipitation frequency and aridity index defined as the ratio of climatological mean of warm-season precipitation to potential evapotranspiration; and vegetated and edaphic factors represent land-surface properties, including root depth, silt fraction, clay fraction, sand fraction and climatological mean of warm-season leaf area index.

GPP anomaly associated with vertically compound droughts for all grid cells across forests (or croplands) was used as the target variable, while the 19 factors served as predictor variables, forming the model dataset. This dataset is randomly divided into a 70% training set and a 30% validation set. Model hyperparameters are optimized using tenfold cross validation. After training, the final random forest model achieves a coefficient of determination of 0.86 for forests (0.78 for croplands). The final random forest models were applied to compute Shapley values⁶⁷, enabling an assessment of the sensitivity of the target variable to predictor variables and providing an improved interpretation of feature importance. Shapley values evaluate all possible combinations of predictors to quantify each variable’s marginal contribution to model performance⁶⁸. For instance, when examining the latent heat flux, the method first tests the model accuracy for all combinations excluding this factor and then evaluates how adding it improves predictive performance. A lower Shapley value indicates a stronger positive contribution to GPP losses (that is, negative anomaly) associated with vertically compound droughts.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41561-026-02083-1.

Author contributions

X.G. and L.W. conceived and designed the study. R.W., S.L. and Y.G. conducted the calculation and visualization. X.G. conducted the analyses and drafted the paper. X.G., R.W., D.C., C.M., S.L., Y.G., S.W., L.L., T.G., C.L., J.L., S.H., X.L., J.W., X.Z. and L.W. contributed to the discussion, editing and improving the paper.

Peer review

Peer review information

Nature Geoscience thanks Emily Black, Wei Shangguan and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Primary Handling Editor: Xujia Jiang, in collaboration with the Nature Geoscience team.

Data availability

All data used in this study are openly available online. ERA5-Land is available at the CDS website (https://cds.climate.copernicus.eu/datasets/reanalysis-era5-land?tab=overview)⁶⁹. ERA5 is available at the CDS website (https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels?tab=overview)⁷⁰. CRA-Land is publicly available at the China Meteorological Data Service Centre website (https://data.cma.cn/en/#/CRA-40/data-download-detail?title=NAFP_CRA40_FTM_3HOR_NC). GLDAS-Noah is publicly available at the GES DISC website (https://disc.gsfc.nasa.gov/datasets/GLDAS_NOAH025_3H_2.1/summary and https://disc.gsfc.nasa.gov/datasets/GLDAS_NOAH025_3H_2.0/summary). ISIMIP3A is publicly available at the ISIMIP website (https://data.isimip.org/). OzNet is publicly available at http://www.oznet.org.au/. ISMN is publicly available at https://ismn.earth/en/. GLOBMAP is publicly available via Zenodo at https://doi.org/10.5281/zenodo.4700264 (ref. ⁴⁴). CSIF is publicly available at the TPDC website (https://data.tpdc.ac.cn/en/data/d7cccf31-9bb5-4356-88a7-38c5458f052b/). Root depth is publicly available via GitHub at https://github.com/yalingliu-cu/plant-strategies. SoilGrids is publicly available at the ISRIC website (https://www.isric.org/explore/soilgrids). Land-use type is publicly available at the USGS website (https://lpdaac.usgs.gov/). FLUXCOM is available from the Data Portal of the Max Planck Institute for Biogeochemistry (https://www.fluxcom.org/CF-Download/). FLUXSAT is available from the NASA Earthdata catalogue (https://www.earthdata.nasa.gov/data/catalog/ornl-cloud-fluxsat-gpp-fpar-1835-2). RTL-LUE GPP is available from the Science Data Bank (https://www.scidb.cn/detail?dataSetId=d99507b2a60947128d602c29504d9ac3). CRU is available (https://crudata.uea.ac.uk/cru/data/hrg/cru_ts_4.07/cruts.2304141047.v4.07/). CMIP6 is publicly available at the ESGF website (https://aims2.llnl.gov/search). GSWP-3 is publicly available at the DIAS website (https://doi.org/10.20783/DIAS.501) (ref. ⁴⁷). The data necessary to reproduce the main results are provided via Zenodo at https://doi.org/10.5281/zenodo.20089130 (ref. ⁷¹).

Code availability

Analysis and figure generation were performed using R (version 4.2.0). The computer code necessary to reproduce the main results is provided via Zenodo at https://doi.org/10.5281/zenodo.20089130 (ref. ⁷¹).

Competing interests

The authors declare no competing interests.

Extended data

is available for this paper at https://doi.org/10.1038/s41561-026-02083-1.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41561-026-02083-1.