Vegetation resistance to compound drought and heatwave events buffers the spatial shift velocities of vegetation vulnerability

Abstract

Global warming is increasing compound drought and heatwave events. This elevates vegetation loss probability. Despite spatial shifts in vegetation loss probability being crucial for predicting spatial redistribution patterns of vegetation vulnerability across terrestrial ecosystems, they remain poorly understood under compound drought and heatwave events. In this study, using a vine copula model, vegetation loss probability was quantified under compound drought and heatwave events. Spatial shift velocities of vegetation loss probability were examined using the concept of velocity change. Spatial shift velocities of vegetation loss probability would undergo substantial increase based on the satellite observations and future simulations. However, vegetation resistance to droughts buffers spatial shift velocities of vegetation loss probability (p < 0.01). These findings provide evidence that vegetation vulnerability patterns will undergo substantial spatial changes under compound drought and heatwave events, leading to a spatial redistribution of vegetation in disturbance-prone areas.

Introduction

Global warming has caused a marked increase in compound drought and heatwave event (CDHWs) frequencies^1,2,3. These events substantially increase vegetation loss probability (VLP) across terrestrial ecosystems, thus highlighting vegetation’s vulnerability under the influence of CDHWs^4,5,6,7.

Droughts and heatwaves can cause climatic niche shifts^8,9,10, thereby forcing spatial shifts in patterns of vegetation distribution^10,11,12,13. This process is driven by alterations in physiological traits, community structures, genetic compositions of species, and distribution ranges^14,15. For example, rising temperatures may cause the distribution patterns of vegetation types to shift poleward or upslope^16,17. Such spatial redistribution of vegetation becomes uneven when it is driven by factors other than temperature¹⁸. For example, when warming and precipitation patterns change, it can result in multidirectional spatial shifts for drought conditions^13,19,20, which in turn can cause multidirectional shifts in vegetation movement patterns. This transfer process may cause a spatial redistribution of VLP, thereby impacting how vegetation responds to environmental stressors. However, the spatial shifts of VLP, which are related to the redistribution of vegetation, remain unclear since most previous studies have focused on the spatial distribution and trends of VLP under CDHWs^7,21,22.

Previous studies have used the concept of velocity⁹, which measures the rates of change in environmental conditions¹⁷, to project carbon gains from geographical vegetation shifts^13,16,17,23. However, these studies commonly disregard the disturbance and mortality processes driving these shifts²⁴. General shift velocities¹⁶ are often concerned with shifts in species composition and distribution. However, unlike such general shift velocities, the spatial shift velocity of VLP is directly tied to changes in vegetation vulnerability²⁵, which includes factors such as productivity loss, mortality risk, and the redistribution of environmental stressors. This metric not only captures the direction and velocity of VLP shifts, but also provides insights into how these shifts may exacerbate or mitigate the risks of vegetation loss under CDHWs.

Vegetation adapts to climate change in two ways: through fast mechanisms such as changes in physiology and phenology^18,26, and through slow mechanisms such as genetic adaptations and shifts in community composition¹⁸. Thus, vegetation has sufficient time to adapt when climate change is slow. This enables the structure and function of vegetation to track climatic niche shifts over a sufficient amount of time¹⁷, thus improving the carbon absorption capacities of terrestrial ecosystems²⁷. However, the acclimation and adaptation of vegetation to warming, along with limited resource availability, such as nitrogen^28,29, may prevent shifts in vegetation from keeping up with climatic niches³⁰. If vegetation shifts are slower than climatic changes^31,32,33, carbon losses could occur in southern boreal forests in response to a northward shift in low-carbon ecosystems²⁴. This shift may counterbalance carbon gains from expanding northern boreal forests²⁴. Therefore, it is crucial to understand how VLP shifts occur spatially under CDHWs to predict disturbance and mortality processes that could reduce carbon benefits obtained from vegetation expansion patterns.

This study examined the characteristics of spatial shifts in VLP under CDHWs as well as the driving factors underlying these shifts. It was conducted across global terrestrial ecosystems. The aim was to answer the following three questions: (1) How does VLP shift spatially under CDHWs? (2) What factors drive VLP spatial shifts under CDHWs? (3) How will VLP spatial shift velocities under CDHWs change in the future? First, using the vine copula model, the joint distribution of the kernel normalized difference vegetation index (KNDVI) and long-term continuous solar-induced fluorescence dataset (LCSIF) was used to quantify VLP under CDHWs. Furthermore, VLP spatial shift velocities were analyzed across different biomes based on the concept of velocity change. Next, a boosted regression tree (BRT) model was employed to assess the contribution rates of climate and biological factors to VLP spatial shift velocities. Finally, using the sixth coupled model intercomparison project (CMIP6) dataset, future spatial shifts in VLP were quantified. These analyses reveal spatial shifts in vegetation vulnerability patterns under CDHWs, and provide valuable insights for adaptive management strategies in the face of climate change.

Results

VLP Spatial patterns and trends triggered by CDHWs

The robustness of the vine copula model (Supplementary Fig. 1)^21,22 is demonstrated by how similar observations and simulated values are at two random locations (101.28°W, 44.37°N, and 61.77°E, 52.75°N). Moreover, the vine copula model results highlight the similarity of VLP based on standardized KNDVI (VLP_KNDVI) and standardized LCSIF (VLP_LCSIF). For example, when VLP ranged from the <10% quantile to the <40% quantile (where lower quantiles (e.g., <10%) represent higher vegetation loss probabilities under CDHWs), both VLP_KNDVI and VLP_LCSIF had high correlation coefficients of 0.69 (p < 0.01), 0.68 (p < 0.01), 0.56 (p < 0.01), and 0.55 (p < 0.01), respectively (Supplementary Fig. 2).

High VLP levels were mainly observed in southwestern North America, southern South America, central Eurasia, southern Africa, and eastern Australia (Fig. 1a–c and Supplementary Fig. 3). These areas were significantly and positively correlated between KDNI/LCSIF and SPEI (r > 0, p < 0.01), and significantly and negatively correlated with STI (r < 0, p < 0.01). These areas are thus heavily influenced by combined climatic stressors, such as prolonged droughts and heatwaves (Supplementary Fig. 4). Moreover, similar results were obtained when CDHWs thresholds were adjusted (a, SPEI < −0.5 and STI > 0.5; b, SPEI < −0.8 and STI > 0.8; c, SPEI < −1.3 and STI > 1.3) (Supplementary Fig. 5).

Fig. 1: Spatial patterns and trends of vegetation loss probability (VLP) under compound drought and heatwave events (CDHWs).

Spatial patterns of (a) VLP_KNDVI and b VLP_LCSIF, both at the <10% quantile. c Mean VLP_KNDVI and VLP_LCSIF values at each latitude. The shaded area in (c) of the latitude line represents ±0.5 standard error from mean a VLP_KNDV and b VLP_LCSIF values. Spatial trends of (d) VLP_KNDVI and e VLP_LCSIF, both at the <10% quantile. The black dots in (d, e) represent trends that were significant at p < 0.05. f Mean VLP_KNDVI and VLP_LCSIF trends at each latitude. The shaded area in (f) of the latitude line represents ±0.5 standard error from mean d VLP_KNDV and e VLP_LCSIF trends. The weighted average of (g) VLP_KNDVI and h VLP_LCSIF from the <10% quantile to <40% was extracted using a 15-year sliding window. VLP_KNDVI and VLP_LCSIF values in (g, h) are plotted in each sliding window.

Both VLP_KNDVI and VLP_LCSIF values increased within a 15-year sliding window, suggesting a widespread increase in vegetation vulnerability under CDHWs. For example, VLP_KNDVI and VLP_LCSIF values increased in more than 60% of the analyzed regions (Fig. 1d–e and Supplementary Fig. 6). Moreover, significant trends were observed in more than 32% and 37% of these areas (p < 0.05), respectively.

For VLP_KNDVI ranging from the <10% to the <40% quantile (Fig. 1g), trend values were 1.38 × 10⁻³·yr⁻¹ (p < 0.01), 1.23 × 10⁻³·yr⁻¹ (p < 0.01), 0.98 × 10⁻³·yr⁻¹ (p < 0.01), and 0.88 × 10⁻³·yr⁻¹ (p < 0.01), respectively. For VLP_LCSIF within the same quantile range (Fig. 1h), trend values were 4.94 × 10⁻³·yr⁻¹ (p < 0.01), 5.19 × 10⁻³·yr⁻¹ (p < 0.01), 5.39 × 10⁻³·yr⁻¹ (p < 0.01), and 5.40 × 10⁻³·yr⁻¹ (p < 0.01), respectively. Additionally, similar results were also obtained when the sliding windows of VLP_KNDVI and VLP_LCSIF were modified (13-year, 17-year, 19-year, 21-year, and 23-year) (Supplementary Fig. 7). This confirms the robustness of the observed spatial trends triggered by CDHWs.

VLP spatial shift velocities triggered by CDHWs

Based on the spatial trend-to-gradient ratio (See, Materials and methods, Fig. 1d, e and Supplementary Fig. 6 (spatial trend), Supplementary Fig. 8 (spatial gradient)), VLP_KNDVI and VLP_LCSIF undergo spatial shifts under CDHWs (Fig. 2c, d). These VLP spatial shifts under CDHWs were faster compared to those under independent drought (SPEI < −2) or heatwave (STI > 2) conditions (Supplementary Notes 1 and Supplementary Figs. 9, 10). This indicates that vegetation vulnerability patterns undergo strong spatial shifts under the influence of CDHWs, thereby exhibiting more substantial shifts in east- or westward directions compared to north- or southward directions (Fig. 2f–h). This is similar to the findings of Pieter et al.³⁴.

Fig. 2: VLP spatial shift velocities and directions under CDHWs.

Mean velocities of a VLP_KNDV and b VLP_LCSIF across different biomes. Error bars in (a, b) indicate ±0.5 standard error from the mean velocities of a VLP_KNDV and b VLP_LCSIF across different biomes. Letters on the horizontal axis in (a, b) represent various biome name abbreviations, with their full names provided in Supplementary Table 1 (1, TSMBF = Tropical and Subtropical Moist Broadleaf Forests; 2, TSDBF = Tropical and Subtropical Dry Broadleaf Forests; 3, TSCF = Tropical and Subtropical Coniferous Forests; 4, TBMF = Temperate Broadleaf and Mixed Forests; 5, TCF = Temperate Coniferous Forests; 6, BF = Boreal Forests/Taiga; 7, TSGSS = Tropical and Subtropical Grasslands, Savannas, and Shrublands; 8, TGSS = Temperate Grasslands, Savannas, and Shrubland; 9, FGS = Flooded Grasslands and Savannas; 10, MGS = Montane Grasslands and Shrublands; 11, TUN = Tundra; 12, MFWS = Mediterranean Forests, Woodlands, and Scrub; 13, DXS = Deserts and Xeric Shrublands; and 14, MF = Mangroves Forests). Spatial shift velocities of (c) VLP_KNDVI and d VLP_LCSIF, both at the <10% quantile. e Mean VLP_KNDVI and VLP_LCSIF velocities at each latitude. The shaded area in (e) of the latitude line represents the mean VLP ± 0.5 standard error from mean c VLP_KNDV and d VLP_LCSIF velocities. Spatial shift velocity directions for (f) VLP_KNDVI and g VLP_LCSIF, both at the <10% quantile. The values of “N”, “NE”, “E”, “SE”, “S”, “SW”, “W”, and “NW” for the color bar in (f, g) represent the directions of north, northeast, east, southeast, south, southwest, west, and northwest, respectively. h The proportion of area in VLP_KNDVI and VLP_LCSIF velocity directions.

For VLP_KNDVI, as it ranged from the <10% quantile to the <40% quantile (Fig. 2c) and Supplementary Fig. 11a, c, e, spatial shift velocities were 8.03 km·yr⁻¹, 7.41 km·yr⁻¹, 6.29 km·yr⁻¹, and 5.31 km·yr⁻¹, respectively. For VLP_LCSIF, within the same quantile range (Fig. 2d) and Supplementary Fig. 11b, d, f, spatial shift velocities were 7.13 km·yr⁻¹, 6.49 km·yr⁻¹, 5.94 km·yr⁻¹, and 5.49 km·yr⁻¹, respectively.

When comparing VLP_KNDVI and VLP_LCSIF spatial shift velocities at different resolutions (0.1° × 0.1°, 0.25° × 0.25°, and 0.5° × 0.5°) (Supplementary Fig. 12), we found their spatial distributions to be similar. However, VLP_KNDVI and VLP_LCSIF spatial shift velocities at higher resolutions exceeded those at lower resolutions, which may be attributed to the greater spatial span at coarser resolutions within the same 3 × 3 moving window, thus resulting in higher spatial heterogeneity.

To assess the robustness of these findings, VLP_KNDVI and VLP_LCSIF spatial shift velocities were also tested using 13-year, 17-year, 19-year, 21-year, and 23-year sliding windows (Supplementary Fig. 13). This yielded similar results. High VLP_KNDVI and VLP_LCSIF spatial shift velocities were primarily observed in central and southern North America, the tropical rainforest regions of northern South America, most parts of Africa, central and western Australia, India, and the Indochinese Peninsula (Fig. 2c, d). Among these biomes, tropical and subtropical dry broadleaf forests (TSDBF) exhibited the highest spatial shift velocities for both VLP_KNDVI (Fig. 2a) and VLP_LCSIF (Fig. 2b). This may be related to the northward shift of low-carbon ecosystems, which could result in carbon loss in the northern edge forests of the tropical and subtropical regions²⁴.

Additionally, using the “trends in net land-atmosphere carbon exchange” (TRENDY-v11) project, 16 state-of-the-art terrestrial ecosystem models were tested under CDHWs to reproduce the spatial shift velocities of standardized leaf area index (LAI) (VLP_LAI) and standardized gross primary productivity (GPP) (VLP_GPP). The models generally reproduced vegetation vulnerability spatial patterns; however, 14 VLP_LAI models underestimated spatial shift velocities (Supplementary Fig. 14), while nine VLP_GPP models overestimated them (Supplementary Fig. 15).

Drivers of VLP spatial shift velocities triggered by CDHWs

Across the board and based on the BRT model, driver contribution rates to VLP spatial shift velocities from the <10% to <40% quantiles consistently responded to the considered variables (Fig. 3a, b), (Supplementary Fig. 18). Therefore, the following analysis focuses on the VLP < 10% quantile.

Fig. 3: Drivers of VLP spatial shift velocities at the < 10% quantile under CDHWs across global terrestrial ecosystems.

The longitudinal axis in (a, b) provides abbreviations across different biomes, with their full names provided in Fig. 2 and Supplementary Table 1. The horizontal axis in (a, b) provides abbreviations of drivers. See Supplementary Table 2 for their full names. c–e show partial dependence plots of the top three variables contributing the most to VLP_KNDVI. f–h show partial dependence plots of the top three variables contributing the most to VLP_LCSIF. The horizontal axis of (c–f) represents standardized driving variables using z-score methods. The horizontal axis in (c–g) of “1-α”, “1-β”, and “1-γ” represents resilience and resistance to drought, and resistance to temperature, respectively. Higher values of “1−α” indicate greater resilience in terrestrial ecosystems. Higher value of “1−β” and “1−γ” indicate greater resistance to drought and resistance, respectively, to temperature in terrestrial ecosystems (See “Materials and methods”).

Overall, at a VLP < 10% quantile, vegetation’s resistance to drought (represented by the 1-β values, see “Materials and methods”) buffer VLP_KNDVI and VLP_LCSIF spatial shift velocities, as determined by BRT partial dependence plots (Fig. 3c, f). At a VLP < 10% quantile, the contribution rates of vegetation resistance responses to drought (1-β values) toward VLP_KNDVI and VLP_LCSIF spatial shift velocities were 52.5% and 44.9% (Fig. 3a, b), respectively.

Despite differences in the most important factors contributing to VLP_KNDVI and VLP_LCSIF spatial shift velocities among different biomes, vegetation resistance response to drought (1-β values) was consistently the most important contributing factors across most biomes (Fig. 3a, b) and (Supplementary Fig. 18). This suggests that the primary factors buffering the spatial shifts of vegetation vulnerability under CDHWs across all global terrestrial ecosystems is vegetation resistance response to drought (1-β values) factors.

We compared partial dependence plots of vegetation resilience (represented by the 1−α values), vegetation resistance to drought (1−β values), and vegetation resistance to heat (represented by the 1−γ values) in relation to VLP_KNDVI and VLP_LCSIF across various global biomes. We specifically found that the vegetation resistance to drought (1−β values) exerts a significant buffering effect on both VLP_KNDVI and VLP_LCSIF (r < 0, p < 0.01) in various biomes, including TSMBF (Tropical and Subtropical Moist Broadleaf Forests), TCF (Temperate Coniferous Forests), BF (Boreal Forests/Taiga), TSGSS (Tropical and Subtropical Grasslands, Savannas, and Shrublands), FGS (Flooded Grasslands and Savannas), MGS (Montane Grasslands and Shrublands), MFWS (Mediterranean Forests, Woodlands, and Scrub), and DXS (Deserts and Xeric Shrublands) (Supplementary Figs. 19–S20).

Future VLP spatial shift velocities triggered by CDHWs

Based on future average datasets regarding GPP, LAI, STI, and SPEI from 10 institutions (for dataset details, see Supplementary Table 3), global-scale VLP_LAI and VLP_GPP values increase significantly (p < 0.01) (Supplementary Figs. 21, 22) throughout the 21st century. These increases indicate a substantial upward trend in vegetation vulnerability spatial patterns under CDHWs across global terrestrial ecosystems.

Furthermore, this analysis projected that VLP_GPP and VLP_LAI will experience faster spatial shift velocities in the future. Moreover, these estimates were higher than the observed change velocities of VLP_KNDVI (8.03 km·yr⁻¹) (Fig. 2c) and VLP_LCSIF (7.13 km·yr⁻¹) (Fig. 2d), as well as CMIP6 historical simulation velocities of VLP_LAI (8.41 km·yr⁻¹) (Fig. 4) and VLP_GPP (4.16 km·yr⁻¹) (Supplementary Fig. 23) from the CMIP6 dataset under most climate scenarios. This pattern likely reflects differences in emission trajectories and climate forcing among these scenarios. Specifically, SSP5-8.5, a high-emission pathway, induces more rapid warming and earlier intensification of extreme climatic conditions, which can accelerate vegetation responses during the mid-term period. In contrast, warming under the lower-emission SSP1-2.6 and SSP2-4.5 scenarios is more gradual, with cumulative impacts becoming more pronounced only in the late period. This may be associated with a higher trend in VLP under these climate scenarios (Supplementary Figs. 21, 22).

Fig. 4: VLP_LAI spatial shift velocities at <10% quantile under CDHWs in the 21^st century.

The future period in (a–i) was divided into an early period (2015–2047) (a–c), a mid-period (2047–2073) (d–f), and a late period (2068–2100) (g–i) (See, “Materials and methods”). The black dots in (a–i) represent projections of future VLP_LAI spatial shift velocities exceeding 1.5 times that of historical values (1982–2014). Numbers in (a–i) represent weighted average velocities, while percentages in (a–i) in parentheses indicate the proportion of future VLP_LAI spatial shift velocities that exceed 1.5 times those of the historical period (1982–2014).

Vegetation vulnerability is expected to undergo substantial spatial redistributions in the future (2015–2100) compared to the historical period (1982–2014). Specifically, the late period exhibits the highest proportion of area where vegetation vulnerability is expected to undergo substantial spatial redistributions. For example, VLP_LAI spatial shift velocities of the late period were projected to exceed 1.5 times the standard deviation of their historical spatial shift velocities over 53%, 51%, and 83% of the area under the SSP1-2.6, SSP2-4.5, and SSP5-8.5 climate scenarios, respectively (Fig. 4g–i). VLP_GPP spatial shift velocities of late period were projected to exceed 1.5 times the standard deviation of their historical spatial shift velocities over 54%, 51%, and 80% of the area under the SSP1-2.6, SSP2-4.5, and SSP5-8.5 climate scenarios, respectively (Supplementary Fig. 23g–i).

Moreover, VLP_LAI and VLP_GPP spatial shift velocities of the late period are primarily concentrated in southwestern North America, southern South America, central Eurasia, southern Africa, and eastern Australia as compared to the historical period (Fig. 4g–i) and (Supplementary Fig. 23g–i). This pattern is similar to historically observed regions with higher VLP spatial distributions, suggesting that vegetation vulnerability patterns in these regions will undergo more pronounced changes in the late period.

Discussion

This study quantified VLP spatial patterns under CDHWs using the vine copula model. VLP is lower in cold high-latitude regions, indicating that temperature rather than moisture has a larger influence on vegetation growth in these environments³⁵. Conversely, VLP is higher in tropical regions, such as southern Africa and eastern Australia, driven by factors such as frequent droughts, extreme weather events, overuse of land, poor soil conditions, frequent fires, invasive alien species, and water shortages^36,37.

Furthermore, a significant increasing trend in VLP (p < 0.01) under CDHWs across terrestrial ecosystems was also observed, suggesting that these ecosystems have become more sensitive to such events⁴. Under CDHWs, vegetation experiences enormous physiological stress: higher temperatures increase transpiration rates^38,39, while drought conditions limit water supply from the soil, thereby preventing plants from obtaining sufficient water for photosynthesis and other vital functions⁴⁰. This compound stress causes growth stagnation, leaf wilting, and even plant mortality, which substantially increases vegetation loss risk²².

By using velocity change, we observed that VLP undergoes spatial shifts, which were quantified as distances over time (km·yr⁻¹)¹⁷. This metric helps to understand how quickly vegetation loss moves across landscapes, thus contributing to a better understanding of how vegetation vulnerability spatial patterns are redistributed²⁴. These shifts can offset potential carbon benefits of warming-induced vegetation expansion, causing substantial changes in terrestrial ecosystems. Many studies have demonstrated substantial correlations between vegetation shifts and temperature velocities^41,42,43, highlighting the impact of warming on vegetation shift spatial dynamics. Given that terrestrial ecosystems are controlled by both temperature and water availability, vegetation shifts are also affected by drought velocity^10,13. For example, Shi et al.¹³. identified a substantial correlation between herbaceous vegetation and aridity velocity in the Sahel region and southern Africa, emphasizing the role of water stress in driving spatial changes in vegetation.

Arid regions are known to experience high VLP due to limited precipitation and soil moisture⁷. As insufficient water availability hampers normal growth and survival, the dry conditions in these areas increase vegetation susceptibility to CDHWs. This study shows that VLP spatial shift velocities are linked to vegetation resistance responses toward SPEI anomalies. On the one hand, global warming is expected to cause a spatial redistribution of climatic niches, thereby prompting vegetation migration in order to maintain current growth conditions^9,32. On the other hand, vegetation resistance to climatic changes causes a lag in shift velocities compared to climatic niche velocities^44,45,46,47. This asynchrony in range shifts could cause changes in species distributions or the formation of new biomes^48,49. Dry conditions primarily drive high VLP velocities, while vegetation resistance responses to droughts slow these shifts.

Vegetation vulnerability patterns undergo strong spatial shifts, thus exhibiting substantial changes, which are highly correlated with the carbon budget. For example, the upper part of the conceptual diagram indicates that changes in warm and wet climatic conditions due to CDHWs cause carbon loss (Fig. 5a); thus, landscapes transition from vegetated to non-vegetated states, indicating high vegetation vulnerability patterns in these areas. The lower part of the conceptual diagram indicates that the transition from CDHWs to warmer and wetter climatic condition causes vegetation greening; thus, landscapes transition from non-vegetated state to vegetated states, indicating low vegetation vulnerability patterns in these areas, which may cause increased carbon uptake.

Fig. 5: Schematic diagram of VLP spatial shift velocities under CDHWs.

a Conceptual diagram of vegetation spatial shift velocities. This figure was created by utilizing resources from the Integration and Application Network (IAN) of the University of Maryland Center for Environmental Science. These resources are made available under a Creative Commons CC BY-SA 4.0 license. b Schematic diagram of a gradient-based method to calculate migration velocity. c Future climatic conditions are inclined towards warming (T_max) and drying (SPEI) under SSP1−2.6, SSP2-4.5, and SSP5-8.5 scenarios.

VLP velocity change can also be understood as the spatial shift produced by changes in plant individuals (species) or community composition in response to climatic conditions¹⁷. On one hand, asynchronous range shifts could cause the dispersal of original species distribution ranges or new forms of contact due to varying velocities of change among plant species, thereby generating new biomes^48,49. On the other hand, changes in biotic interactions hinder or facilitate further range shifts in biomes⁵⁰. For example, alterations in predation dynamics, herbivory, host-plant associations, competition, and mutualistic symbioses of species have substantial biome level impacts⁴⁹. Changes in species’ ranges that occur among foundational or habitat-forming species can cause species redistribution across entire biomes, which in turn alters ecosystem productivity and carbon storage¹⁸.

Adaptive strategies informed by VLP velocities can help to mitigate climate change impacts. Assisted migration involves relocating species to habitats where climatic conditions are favorable⁵¹, particularly in areas with high VLP velocities where species migration is too slow⁵². Fire management can be optimized by adjusting strategies in fire-prone areas with rapid vegetation shifts by using early detection and controlled burns. Restoration and conservation efforts can prioritize areas with high VLP velocities to preserve ecosystem functioning⁵³.

Our findings on VLP spatial shift velocities are directly relevant to several key international frameworks. An example includes the UN Decade on Ecosystem Restoration (2021–2030), which emphasizes the importance of restoring degraded ecosystems to prevent biodiversity loss and mitigate climate change⁵⁴. Our study underscores the need for targeted restoration efforts in areas that are most vulnerable to climate-induced vegetation shifts. Additionally, the Paris Agreement aims to limit global warming to well below 2 °C, with efforts to restrict the temperature rise to a maximum of 1.5 °C⁵⁵. Our results, which assessed vegetation vulnerability shifts in response to CDHWs, can inform climate mitigation and adaptation strategies outlined in the Paris Agreement. Furthermore, our study aligns with the Sustainable Development Goals (SDGs), particularly SDG 13 (Climate Action) and SDG 15 (Life on Land)⁵⁶. By assessing how vegetation loss and shifts impact ecosystems, our findings contribute to understanding the broader environmental implications of climate change and support ongoing efforts to manage and restore ecosystems as part of global climate strategies.

This study provides evidence for spatial shifts of vegetation vulnerability under CDHWs. When VLP ranges from the <10% quantile to <40% quantile, VLP_KNDVI velocities range from 8.03 km·yr⁻¹ to 5.31 km·yr⁻¹, while VLP_LCSIF velocities range from 7.13 km·yr⁻¹ to 5.49 km·yr⁻¹ during 1982–2021. Furthermore, CMIP6 dataset simulations suggest that VLP will increase significantly in the future (p < 0.01). Compared to the early and mid periods, this projected increase will cause spatial VLP_LAI and VLP_GPP shift velocities to exceed 1.5 times the standard deviation of their historical values of 53%, 51%, and 83% of the area under the SSP1-2.6, SSP2-4.5, and SSP5-8.5 climate scenarios, respectively. However, vegetation resistance buffers VLP_KNDVI and VLP_LCSIF spatial shift velocities, as identified by the BRT model. Based on the widespread evidence of increasing VLP under CDHWs across terrestrial ecosystems, this study provides strong evidence that VLP will undergo large changes in their spatial patterns. This finding offers a perspective for projecting the spatial redistribution of vegetation vulnerability across terrestrial ecosystems under CDHWs.

Materials and methods

Climate datasets

The monthly CRU-TS 4.08 from the climatic research unit time series (CRU-TS), with a spatial resolution of 0.5° during 1982–2021, was used in this study. This time series includes temperature (T, °C), precipitation (Pre, mm), and potential evapotranspiration (PET, mm). Based on the datasets of Pre and PET, SPEI at the 1–12-month timescale (SPEI-01–SPEI-12) were calculated using the “SPEI” function in R 4.31.

Temperature data were standardized into the Standard Temperature Index (STI) across timescales of 1–12 months (STI-01–STI-12) based on the cumulative distribution function. For all subsequent analyses, climate data were averaged for the warm season, which was determined in each grid cell. Specifically, for each grid cell, the warm season was defined as the three consecutive months with the highest average temperature. To further accurately identify warm seasons, we used the CRU-TS 4.08 monthly temperature dataset covering the years 1982–2021. This approach ensured that the warm season reflected local climatic conditions rather than a fixed global period (Supplementary Fig. 26).

For example, if a pixel’s warm season was January–March, then the warm season STI for that pixel was calculated as the average of the STI values for January, February, and March. Similarly, if a pixel’s warm season was February–April, then the warm season STI was computed as the average of the STI values for February, March, and April; …; if a pixel’s warm season was November–January, then the warm season STI was derived by averaging the STI values for November, December, and January (of the following year); finally, if a pixel’s warm season was December–February, then the warm season STI was obtained by averaging the STI values for December, January, and February (with the appropriate alignment of years).

Vegetation datasets

The vegetation observation datasets included the normalized difference vegetation index (NDVI) and LCSIF during 1982–2021, which were used as proxies of canopy greenness and vegetation productivity, respectively.

From 1982 to 2021, NDVI data were drawn from the GIMMS-3G+ (third-generation V1.2) dataset, which offers global coverage at a spatial resolution of 0.0833° × 0.0833°. These NDVI measurements are provided twice per month. The dataset integrates information from multiple AVHRR sensors and mitigates challenges such as calibration degradation, orbital drift, and the effects of volcanic eruptions. Furthermore, by using the maximum composite method, the biweekly NDVI data were aggregated into monthly NDVI. We also excluded regions where the multi-year average NDVI < 0.1, in order to avoid interference from non-vegetated areas.

LCSIF was generated using neural networks to map the BRDF-adjusted red and near-infrared reflectance’s from MODIS MCD43C4 to clear-sky Nadir Mode SIF observations from the Orbiting Carbon Observatory-2 (OCO-2). LCSIF exhibits a relatively extended temporal coverage (1982–2021). Moreover, compared with empirical vegetation indices, the reconstructed SIF data predicts more reliably, indicating that SIF captures more unique information about photosynthetic dynamics. We selected clear-sky daily averaged SIF, while excluding values <0.

Formula (1) below was used to convert NDVI to KNDVI. The advantages of KNDVI are that it reduces sensitivity to soil background noise in areas with sparse vegetation and minimizes canopy saturation effects in areas with a high vegetation density⁵⁷. LCSIF can be used as a proxy for vegetation productivity and offers considerable advantages regarding the directness, accuracy, and sensitivity of obtaining photosynthesis information. This helps to better understand and monitor the dynamics of ecosystem photosynthesis⁵⁸.

$${kNDVI}=tanh ({NDV}{I}^{2})$$

(1)

To verify vegetation observation datasets, auxiliary analyses were conducted using GPP and LAI generated by the offline dynamic global vegetation model, representing vegetation productivity and canopy greenness, respectively. These GPP and LAI datasets were sourced from the “trends in net land-atmosphere carbon exchange” (TRENDY-v11) project⁵⁹. Simulations included four scenarios: S0 (CO₂-only), S1 (CO₂ + climate), S2 (CO₂ + climate + land use), and S3 (CO₂ + climate + land use + nitrogen). S3 simulation results were used, which account for CO₂ increase, climate change, and land use changes (Supplementary Table 4). The selected GPP and LAI datasets cover the period from 1982 to 2021.

Earth system models dataset

Monthly datasets on LAI, GPP, maximum and minimum temperatures, precipitation, relative humidity, total soil moisture, and radiation were obtained from CMIP6 for both historical (1982–2014) and future (2015–2100) periods under the three radiative forcing pathways of SSP1-2.6, SSP2-4.5, and SSP5-8.5. We only selected earth system models that included all variables. A total of 10 earth system models produced these data at the time of writing this paper, which were downloaded from the Department of Energy Lawrence Livermore National Laboratory server. To reduce uncertainty, the ensemble mean was calculated for each variable across the 10 institutions (for details of the CMIP6 dataset, see Supplementary Table 3).

The physiological response of vegetation to rising CO₂ concentrations was incorporated by using monthly, zonally resolved CO₂ data from 1982 to 2100 under both scenarios. PET, considering CO₂ concentration ([CO₂]), was calculated using the reference crop Penman-Monteith equation. It was specifically combined with precipitation to calculate SPEI_PET [CO₂] (SPEI_CO2) using the “SPEI” package in R at the 1–12-month timescale. Additionally, CMIP6 average temperature data (calculated using average maximum and minimum temperatures) were used to simulate the STI at the same timescale under both future scenarios.

Auxiliary datasets

Auxiliary datasets included global land use datasets (Supplementary Fig. 27) and terrestrial ecosystems (Supplementary Fig. 28). Based on the MCD12C1 v006 dataset, we excluded any pixels that changed from 2001 to 2021 to eliminate the influence of human activities. Moreover, according to the International Geosphere-Biosphere Programme’s land classification scheme, we excluded land use types such as barren terrain, ice-covered areas, farmland, and urban areas to eliminate the influence of farmland, human activities, and non-vegetated areas. Before conducting any analyses, we employed this land-use dataset to mask climate, vegetation, and earth system model data, thereby excluding the impact of non-vegetated regions. Finally, we divided the earth into 14 biomes based on the terrestrial ecoregions of the world dataset⁶⁰.

Quantifying VLP spatial patterns triggered by CDHWs

First, the seasonal and trend decomposition LOESS function in R 4.31 was used to decompose the monthly vegetation indices into trend, seasonal, and residual components. A time series of the residual components of these indices were used for further analysis. Hereafter, vegetation indices were standardized (SVI) based on the cumulative distribution function. The purpose of using standardized indices (including SPEI, STI, and SVI) was to enable a comparison of VLP under CDHWs across different pixels and regions⁶¹. All vegetation indices were averaged for the warm season.

The absolute values of the correlation coefficients between SVI, SPEI, and STI for the preseason 1–12 months during 1982–2021 were calculated (Supplementary Fig. 29) to determine optimal response months of the standardized vegetation indices (i.e., standardized KNDVI and LCSIF) for SPEI and STI. Hereafter, the joint distribution probability of SVI, SPEI, and STI in the optimal response months (the months corresponding to the maximum absolute value of the correlation coefficient) was established on a pixel-by-pixel basis to identify the spatial patterns of VLP under CDHWs. This was based on the vine copula model.

The vine copula model is advantageous because it can capture the nonlinear relationship between VLP and CDHWs without assuming a normal distribution, while also identifying different VLP levels under CDHWs⁷. The vine copula model also offers a flexible and robust framework for capturing complex, nonlinear dependencies among multiple variables, particularly in cases where tail dependencies are significant^62,63. Compared to Bayesian hierarchical models—which can be computationally intensive and require strong assumptions about the data structure—and many machine learning approaches—which often act as “black boxes” with less interpretability—the vine copula approach provides a clear, interpretable structure for modeling interdependencies while still maintaining computational efficiency⁶².

The range of VLP under CDHWs was [0,1]. Higher VLP values indicated higher vegetation vulnerability patterns under CDHWs, as described in Formula (2):

$$\begin{array}{c}{F}_{{SVI}{|}{SPEI},{STI}}({VI}\le {svi}{|}{SPEI}\le {spei},{STI}\ge {sti})=\frac{F({SVI}\le {svi},{SPEI}\le {spei},{STI}\ge {sti})}{F({SPEI}\le {spei},{STI}\ge {sti})}\\ =\frac{F({spei},{svi})-F({spei},{sti},{svi})}{F({spei})-F({spei},{sti})}\end{array}$$

(2)

where ${F}_{{SVI|SPEI},{STI}}({VP}\le {svi|SPEI}\le {spei},{STI}\ge {sti})$ represents the cumulative distribution function of VLP under drought (SPEI ≤ spei) and heatwave (STI ≥ sti) conditions. SPEI ≤ spei are representative of values below a certain drought threshold, while STI ≥ sti represent values exceeding a certain heatwave threshold. Furthermore, the joint probability density function can be expressed by Formula (3):

$$\begin{array}{c}f\left(s{{\mbox{vi}}},{Spri},{Sti}\right)={f}_{{SVI}}\left({svi}\right){{\cdot }}{f}_{{SPEI}}({spei}){f}_{{STI}}({sti})\to (m{\arg }i{nal\; densities})\\ {{\cdot }}{c}_{{STI},{SVI}}({u}_{{STI}},{u}_{{SVI}}){{\cdot }}{C}_{{STI},{SPEI}}({u}_{{STI}},{u}_{{SVI}})\to ({Edges}\,{{\mbox{in}}}\,{{\rm{Tree}}}1)\\ {{\cdot }}{c}_{{SVI},{SPEI}{|}{STI}}({F}_{{SVI},{STI}}(s{{\mbox{vi}}}{|}{sti}),{F}_{{spei},{sti}}({spei}{|}{sti}))\,\to ({Edges}\,{{\mbox{in}}} \, {{\rm{Tree}}}2)\end{array}$$

(3)

where f_i(x_i) represents the probability density distribution function of SVI, SPEI, and STI; F_i(x_i) represents the cumulative distribution functions of SVI, SPEI, and STI; C_{SVI, STI}, C_{SPEI STI}, and C_{SVI, SPEI} represents the joint densities of SVI, SPEI, and STI, respectively. The VLP under the same thresholds of CDHWs derived from Formula (4) was as follows:

$$F({SVI}{|}{SPEI},{STI})=\frac{\partial {C}_{{SVI}{|}{SPEI},{STI}}(u({SVI}{|}{SPEI}),u({SPEI}{|}{STI}))}{\partial u({SPEI}{|}{STI})}$$

(4)

Under the conditions of joint probability of CDHWs (STI ≥ 2 and SPEI ≤ -2), and to describe VLP under CDHWs, VLP was quantified for four scenarios (below the 10th percentile (SVI < 10% quantile), below the 20th percentile (SVI < 20% quantile), below the 30th percentile (SVI < 30% quantile), and below the 40th percentile (SVI < 40% quantile). Smaller quantile values of VLP indicate greater vegetation loss under the same CDHWs thresholds. By ranking these standardized residuals of SVI, lower quantiles (e.g., <10%, <20%, <30%, and <40%) effectively capture the relative extent of vegetation loss, with lower values indicating more pronounced deviations. Unless otherwise specified, CDHWs refer to the joint occurrence of standard precipitation-evapotranspiration index (SPEI) < −2 and standardized temperature index (STI) > 2.

This percentile method is a non-parametric approach that has been widely used in previous studies to quantify vegetation anomalies. For example, Konings et al.⁶⁴ employed NDVI deficit percentiles to examine vegetation sensitivity to atmospheric conditions, and similar approaches have been applied to assess extremes in vegetation carbon sequestration and monitor crop growth⁶⁵. While these quantile thresholds are not directly tied to specific vegetation mortality thresholds, they are ecologically meaningful since they represent reductions in vegetation functioning relative to a multi-year baseline.

Five ternary copula families were used (Gaussian, Student’s t, Frank, Gumbel, and Joe). The best-fitting copula was selected for each grid cell to represent the joint probability distribution of ternary copulas based on the Bayesian information criterion. Ternary copula modeling was implemented using the “rvinecopulib” function in R 4.31.

Quantifying VLP trends triggered by CDHWs

VLP was measured using a sliding window with a length of 15 years, thereby generating a time series of VLP coefficients for each location. Moreover, to test the robustness of VLP coefficients, 13-year, 17-year, 19-year, 21-year, and 23-year sliding windows were also used to generate a time series of VLP coefficients for each location.

To assess VLP coefficient trends, the Mann-Kendall trend test, a non-parametric statistical method used to detect trends in time series data, was employed with the confidence level set to 0.95 to measure VLP coefficient tendency.

Quantifying VLP spatial shift velocities triggered by CDHWs

Velocity (km·yr⁻¹) is defined as the rate at which VLP moves spatially in a certain direction under CDHWs. This velocity can be interpreted as the distance traversed by the VLP in space per unit of time. The advantage of using velocity is that it converts variables of different units into the same spatial metric, enabling a comparison of changes in their spatial patterns. There are two methods to test spatial shift velocities, namely gradient-based methods and simulation-based methods. In this study, a gradient-based method was used to calculate VLP velocity (g_vi) under CDHWs. The “VoCC” package in R version 4.2.0 was used to implement this function⁶⁶.

Although this method measures VLP velocity, it does not directly capture the rates of change for species composition or productivity loss. Instead, it quantifies spatial shifts in vegetation vulnerability as influenced by CDHWs. Velocity represents the rate at which vulnerability, in terms of mortality risk and the potential for loss in vegetation cover, changes across space. As such, while vegetation migration rates often refer to the spatial movement of species, this study’s focus on VLP emphasized the shifting risk of vegetation loss rather than shifts in species per se.

gV_i is defined as the ratio of the time trend (t_i) to the spatial gradient (g_i) (Fig. 5b), where t_i is the slope at grid point i, estimated using the ordinary least squares regression model; the local spatial gradient (g_i) is defined as the vector sum of both longitudinal and latitudinal pairwise differences for each focal cell. A 3×3 neighborhood was used to define the local spatial gradient. The vector direction of spatial shift velocity (gV_i) depends on the sign of the time trend (g_i) and the direction of the spatial gradient (g_i). The vector direction of shift velocity (θ) ranges from 0 to 360°, where the direction of true north is 0°, rotation is clockwise, and true south is 180°.

Because of the positive sign of the spatial gradient, the sign of the spatial velocity depends on the spatial trend¹³. Thus, a positive VLP velocity represents an increasing spatial shift in vegetation vulnerability, whereas a negative VLP velocity represents a decreasing spatial shift in vegetation vulnerability.

$$g{V}_{i}=\frac{{t}_{i}}{{g}_{i}}$$

(5)

Quantifying the drivers of VLP spatial shift velocities triggered by CDHWs

Evidence suggests that spatial shifts in climate and biological factors cause changes in the shifts of vegetation^16,17,24, which could influence the redistribution of vegetation spatial patterns across entire terrestrial ecosystems. However, the resilience and resistance of vegetation to climate change often result in a lagging of spatial shift velocities. Specifically, vegetation with a high climatic resistance exhibits slower spatial shift velocities under exposure to similar climates¹⁶.

VLP_KNDVI/VLP_LCSIF spatial shift velocities are considered to be drivers, while predictor drivers are mainly divided into two categories: biological factors and climate forts. For climate factors, we primarily selected SPEI, STI, CDHWs, and short-wave radiation from the warm season of 1982–2021, along with CO₂ from the warm season of 2001–2018. From here we calculated the spatial shift velocities of these factors. Some studies have shown that the spatial shift velocities of SOS and EOS are related to changes in the spatial patterns of vegetation productivity¹⁶. Therefore, we also calculated their spatial shift velocities.

For the resilience and resistance factors, according to Supplementary Notes 2, we calculated the characterization factors α, β, and γ for the resilience and resistance of KNDVI and LCSIF. Their range is [0,1]. The α values represent KNDVI and LCSIF resilience (α_KNDVI, α_LCSIF). Larger α values represent lower KNDVI or LCSIF resilience. The β values represent the resistance responses of KNDVI and LCSIF (β_KNDVI, β_LCSIF) to SPEI. The γ values represent the resistances to temperature for KNDVI and LCSIF (γ_KNDVI, γ_LCSIF). Larger values of β and γ represent smaller resistance values for KNDVI or LCSIF (Supplementary Table 6). To avoid confusion, we transformed these parameters by calculating (1 – α), (1 – β), and (1 – γ) so that larger values correspond to higher resilience and resistance values.

Before training the model, higher correlated drivers were removed. For example, if the spatial shift velocities of warm-season precipitation, warm-season soil moisture, and warm-season temperature exhibited strong correlations (|r| > 0.7), the drivers with lower correlations to the spatial shift velocities of VLP_KNDVI/VLP_LCSIF were eliminated. Through initial variable screening, we selected 10 potential explanatory variables to explain VLP_KNDVI/VLP_LCSIF spatial shift velocities.

The contribution rates of these variables to spatial shifts were assessed using the BRT model. The “gbm”⁶⁷ package in R was used to implement this function. In the BRT model, 10 variables were used as drivers. Z-scores were used to standardize the data, ensuring that all data input into the model adhered to a normal distribution. The BRT model was set to 1000 trees, a learning rate of 0.001, a tree complexity of 5, and a bag fraction of 0.5⁶⁸.

Quantifying future VLP spatial shift velocities triggered by CDHWs

First, based on a 15-year sliding window, VLP trends for GPP and LAI (VLP_GPP and VLP_LAI) were quantified under different climate scenarios (Historical, SSP1-2.6, SSP2-4.5, and SSP5-8.5) for the period of 1982–2100. Hereafter, VLP_GPP and VLP_LAI spatial shift velocities were quantified for the period of 1982–2100. If future VLP_GPP and VLP_LAI velocities exceed 1.5 times that of the historical period (1982–2014), then the spatial patterns of vegetation vulnerability in these areas may experience substantial changes under future CDHW conditions.

To more precisely study the future VLP spatial shift velocities triggered by CDHWs, we divided the future period into three phases: early, mid, and late. However, because the future period spans only 86 years while each phase requires 33 years of data (based on a 15-year moving average yielding 19 values), a non-overlapping division would require a total of 99−13 years more than what is currently available. Ideally, these extra 13 years would be distributed evenly as ~6.5 years of overlap between adjacent phases. Since fractional years are impractical, we introduced integer overlaps instead. Thus, the future period was divided into an early period (2015–2047), a mid-period (2047–2073), and a late period (2068–2100).

Additionally, to assess uncertainties related to model selection, we computed VLP spatial shift velocities for each earth system models under each climate scenario and quantified the biases across models (Supplementary Figs. 24–S25).

Reporting summary

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