Biodiversity and Wetting of Climate Alleviate Vegetation Vulnerability Under Compound Drought‐Hot Extremes

Global warming has intensified the intensity of compound drought‐hot extremes (CDHEs), posing more severe impacts on human societies and ecosystems than individual extremes. The vulnerability of global terrestrial ecosystems under CDHEs, along with its key influencing factors, remains poorly understood. Based on multiple remote sensing data, we construct a Vine Copula model to appraise vegetation vulnerability under CDHEs, and attribute it to climatic and biotic factors for five different vegetation types. High vulnerability is detected in central and southern regions of North America, eastern and southern regions of South America, Southern Africa, northern and western Europe, and northern and eastern Australia. The drier the climate, the higher will be the vulnerability. Furthermore, biodiversity and biomass are key biotic factors influencing the vulnerability of various vegetation types, such that ecosystems with richer biodiversity and higher biomass have lower vulnerability to CDHEs. The findings deepen understanding of terrestrial ecosystem response to CDHEs.


Introduction
Vegetation, as a key component of terrestrial ecosystems, serves as a crucial link between the soil and atmosphere, and its dynamics have a profound impact on carbon balance, water cycles, and climatic conditions (Linscheid et al., 2020;Piao et al., 2019;B. Zhang, Tian, et al., 2022).With global warming, intensities of climate extremes have increased over the past decades in many regions across the world, including droughts, heatwaves, severe storms, floods, and so on (Z.Huang et al., 2023;Ma & Yuan, 2024;Tian & Zhu, 2024;Tripathy & Mishra, 2023;et al., 2023).Hence, investigating how vegetation reacts to CDHEs globally is essential under the effects of climate change, which will help us develop strategies for ecosystem management.
Various studies have investigated responses of vegetation dynamics to droughts, high temperatures, and CDHEs at regional and global scales based on field experiments, model simulations, and remote sensing data sets (J.Chen et al., 2021;Deng et al., 2021;Martinez-Sancho et al., 2022;Smith et al., 2022;Trugman, 2022;Xi & Yuan, 2022).Due to the lack of temporal and spatial continuity of field experiments and large biases of model simulations, remote sensing-based vegetation indices, with their long-term and large-scale monitoring capability, have been increasingly utilized to map ecosystem conditions (AghaKouchak et al., 2015;Q. Chen et al., 2023).Some researchers have focused on correlation analysis to analyze the impacts of droughts on vegetation by calculating Pearson correlation coefficients between drought/hot indices and vegetation indices (Jin et al., 2023;X. Zhang & Zhang, 2019).However, correlation analysis struggles to fully capture the interdependence between drought/hot extremes and vegetation dynamics, known as tail dependence (Aas et al., 2009;Hobaek Haff et al., 2010).The absence of a significant overall correlation between two variables does not preclude the possibility of tail dependence, which presents a significant challenge in assessing the impacts of climate extremes on ecosystems.
The copula-based conditional probability is an efficient method for measuring the impacts of drought/hot extremes on vegetation (Fang et al., 2023;X. Wu et al., 2022;G. Zhang et al., 2024).However, under highdimensional conditions (e.g., CDHEs effects on vegetation), the conventional single-or multi-parameter copula models lack effectiveness in quantifying dependent structures due to the heterogeneous relationships between any two variables (H.Wu et al., 2022;B. Zhang, Wang, et al., 2022).To overcome this issue, H. Li et al. (2021) developed a vine copula model to investigate the vulnerability of vegetation dynamics, represented by Normalized Difference Vegetation Index (NDVI), to CDHEs in the Xinjiang region.Nevertheless, previous studies have yet revealed the key factors influencing global vegetation vulnerability.Additionally, the NDVI index is plagued by its saturation issue, which limits its performance in dense vegetation-covered regions (Camps-Valls et al., 2021).Therefore, we employ various vegetation indices to characterize vegetation productivity and evaluate vegetation vulnerability to CDHEs at a global scale to ensure the robustness of the results.Subsequently, a machine learning model is utilized to attribute the vulnerability to different climatic and biotic variables.The results can broaden understanding of terrestrial vegetation responses to CDHEs and advance ecosystem models for better evaluation and prediction of CDHEs impacts.

Vegetation Indices Preprocessing
To achieve reliable results, several vegetation indices, including kernel NDVI (kNDVI), Enhanced Vegetation Index (EVI), Near-Infrared Reflectance of Vegetation (NIRv), and Solar-induced Chlorophyll Fluorescence (SIF), were used to monitor vegetation dynamics (Badgley et al., 2017;Camps-Valls et al., 2021;Frankenberg et al., 2011;Huete et al., 2002).kNDVI, proposed by Camps-Valls et al. (2021) to overcome the saturation issue of NDVI, performs better in representing vegetation biomes, especially in dense vegetation-covered regions.NIRv is a recently proposed vegetation index designed to mitigate the effect of mixed-pixel issues and influences of background brightness and soil contamination problems (Badgley et al., 2017).SIF can be used to monitor plant photosynthesis, and it is closely related to gross primary production (Li & Xiao, 2019;Yang et al., 2017).EVI, more sensitive to high biomass vegetation than NDVI, represents a comprehensive composite property of leaf chlorophyll content, canopy structure, and green leaf area (Huete et al., 2002;J. Zhang et al., 2022).More detailed descriptions can be found in Text S1 in Supporting Information S1.Due to the existence of seasonality and trends of vegetation indices, they cannot be used directly to analyze the responses of vegetation dynamics to CDHEs.Therefore, the Seasonal and Trend decomposition using Loess (STL) method was used to decompose the vegetation indices, of which the residual component (detrended and de-seasonalized) is used for further analysis (Cleveland et al., 1990).The STL method and an example of kNDVI decomposition are presented Text S2 and Figure S1 in Supporting Information S1.

Identification of Compound Drought-Hot Extremes
In this paper, Standardized Precipitation and Evaporation Index (SPEI) was selected to represent droughts (Vicente-Serrano et al., 2010, 2013).The details of the SPEI calculation process are presented in Text S3 in Supporting Information S1.Standardized Temperature Index (STI) was used to monitor hot conditions in this Songbai Song, Kai Feng, Tianliang Jiang, Jinbai Huang, Pengcheng Xu, Xiaolei Fu paper, and the computation method is given in Text S4 in Supporting Information S1.Considering the cumulative effects of CDHEs on vegetation growth, we first evaluated the cumulative effect durations of SPEI and STI on vegetation dynamics using the Pearson correlation coefficient (Text S5 in Supporting Information S1 for estimating cumulative effects of droughts).The correlation coefficients between detrended and de-seasonalized vegetation indices (VIs_de) and multi-timescale SPEI/STI (1-12 months) were calculated for each pixel, the time scale corresponding to the maximum correlation coefficient was identified at each pixel, and the SPEI and STI corresponding to maximum correlation time scale were selected (SPEI_m and STI_m).The extremely dry and hot conditions were identified with SPEI_m < 2 and STI_m > 2 (H.Wang, Zhang, et al., 2023).Leveraging the statistical relationship between SPEI_m/STI_m and VIs_de, we conducted 2,000 simulations using Vine Copula.Based on these simulations, we calculated the probability of VIs_de falling below a certain threshold when SPEI_m < 2 and STI_m > 2. This method ensures the robustness of the results.

Estimating Vegetation Vulnerability Under Compound Drought-Hot Extremes
The vegetation loss probability (VLP) given CDHEs is calculated to assess the vegetation vulnerability under CDHEs.
For several random variables X = (X 1 , ⋯, X n ), the joint cumulative probability function (CDF) can be written as and its probability density function (PDF) can be obtained using the chain rules as: Based on the vine copula decomposition, the joint probability density function can be expressed as (Aas et al., 2009): where f(x i ) and u i = F(x i ) are probability density and cumulative probability functions of the variable x i , respectively.c(u i , u j ) represents the joint density function of F(x i ) and F(x j ).The conditional probability of F(x|v) can be calculated as: Constructing vine copula models involves two key steps: (a) selecting and fitting marginal distributions and (b) selecting vine structures and determining parameters.We fitted SPEI and STI series to a Gaussian normal distribution, because SPEI and STI were calculated through Gaussian inverse transformation for cumulative probability of water deficit and temperature.We also employed the Kolmogorov-Smirnov (KS) method to assess the feasibility of SPEI and STI sequences adhering to a normal distribution.In terms of VIs_de, five distributions were selected as candidates (Table S1 in Supporting Information S1), and KS-test and AIC metric were utilized to ascertain the most suitable distribution for each pixel during each month.The maximum likelihood method was used to estimate the parameters.Consequently.A vine copula model was developed based on the marginal probabilities of SPEI_m, STI_m, and VIs_de, where six copula functions were selected to join variables (Table S2 in Supporting Information S1), and the optimal copula function was determined with AIC matric (Dißmann et al., 2013).

Factors Affecting Vegetation Vulnerability
Effects of multiple factors, including biotic, climatic, and soil structure variables, on VLP under CDHEs across diverse vegetation categories were modeled using a machine-learning method (Random Forest; RF).RF is a simple machine-learning method for attribution analysis (Breiman, 2001).RF provides valuable advantages for attribution analysis, including its ability to assess feature importance, model nonlinear relationships, handle high-dimensional data, maintain robustness against noise and outliers, and offer interpretability through feature analysis.These attributes enable RF to effectively identify significant drivers within data sets and elucidate their impact on outcome changes.In terms of biotic variables, biodiversity, above-ground biomass (AGB), and maximum root depth were selected as factors influencing VLP.For climatic variables, precipitation, temperature, radiation, aridity index, and vapor pressure deficit (VPD) were selected (Y.Li et al., 2023;Yao et al., 2023).The method to estimate VPD is given in Text S6 in Supporting Information S1.In addition, soil clay content for soil structure was also selected.The RF regression was used to quantify the effect of each variable on VLP.The partial dependence of VLP to each variable was also investigated.The RF regression was trained by a bagging algorithm, for building a robust model.One can find more details of the RF method in Text S7 in Supporting Information S1.The framework for evaluating VLP to CDHEs and the respective effect of the factors considered is shown in Figure 1.

Cumulative Effects of Drought and Hot Events on Vegetation
Drought and hot events have cumulative effects on vegetation dynamics.Based on the correlations between VIs_de and SPEI/STI at different time scales, the maximum correlations and the corresponding time scales at

Distribution of Vegetation Vulnerability With Respect to CDHEs
The overall correlations are challenging to evaluate the impact of extreme events on vegetation dynamics, given that vegetation health is influenced by various factors, including root zone depth, biomass, and so on.Motivated by this challenge, the vine copula was used to quantify the VLP with respect to specific CDHEs.Figure 2 shows the distribution of VLP (VIs_de < 10%) given SPEI_m < 2 and STI_m > 2 for kNDVI, EVI, NIRv, and SIF, respectively.The VLP for different thresholds of vegetation loss are also presented in Figure S7 in Supporting Information S1.High VLP is mainly located in central and southern regions of North America, eastern and southern regions of South America, Southern Africa, northern and western Europe, and northern and eastern Australia.In the northern hemisphere (NH), from 60°N to Equator, VLP initially increases, then decreases, and then alternates between increase and decrease, with the two peaks occurring around 40°N and 20°N, respectively.In cold regions, the correlations between SPEI and VIs are always negative (Figure S4 in Supporting Information S1), while they are positive between STI and VIs (Figure S5 in Supporting Information S1), which indicates that in a cold environment, temperature rather than water availability dominates plant growth, meaning that global warming would promote vegetation growth and increase the vegetation productivity (Keenan & Riley, 2018;Zhao & Running, 2010).In contrast, in equatorial regions, vegetation with high production and biodiversity exhibits lower vulnerability due to the abundance of water resources in tropical rainforest areas.In Eastern Africa, characterized by a dry climate and limited precipitation, grasses and shrubs dominate as the primary vegetation cover.The ecosystem in this region is more vulnerable and more prone to disturbances from extreme events, resulting in regions with high vulnerability.In the southern hemisphere (SH), moving southward from the Equator, VLP undergoes both increasing and decreasing trends, reaching peak values in regions located at latitudes around 30°S (SIF) and 50°S (kNDVI, EVI, and NIRv).
The VLP estimated using SIF and other VIs show obvious divergence in some regions of SH, such as southern and western Australia (Region 1 enclosed in Figure S6 in Supporting Information S1), which is characterized by arid or semi-arid climate and covered by grasses, shrubs, and savannas, where VLP for SIF is higher than that of other vegetation types.However, VLP for SIF is lower than that of other vegetation types in humid and semi-humid regions of the southernmost part of South America (Region 2 shown in Figure S6 in Supporting Information S1). Figure S8 in Supporting Information S1 compares the variations of SIF and other VIs in the two selected regions.In Region 1, SIF responds more quickly to dry and hot events than other VIs, likely because SIF more directly represents plant physiological and bio-chemical conditions than other types of VIs (Frankenberg et al., 2011).
There are significant divergences in VLP among different vegetation types (Figure 2 and Figure S9 in Supporting Information S1).The VLP of grasses is markedly higher than that of other vegetation types, followed by shrubs and savannas, with evergreen forests exhibiting the lowest VLP.Under extremely dry and hot conditions (SPEI_m < 2 and STI_m > 2), the average VLP for grasses is approximately 0.51, while for evergreen forests, it is only about 0.2.For evergreen forests, VIs are negatively correlated with SPEI_m, and positively correlated with STI_m.In rainforests with abundant rainfall, higher vapor pressure deficit (VPD) caused by higher temperatures can promote stomatal opening and transpiration processes, which improve plant growth (Green et al., 2020).Compared to forests, grasses obtain water from shallow soil layers, making them more sensitive to CDHEs, which can deplete surface water and accelerate evapotranspiration, resulting in vegetation wilting or even dying (Hao et al., 2021;X. Zhang & Zhang, 2019).We also find an obvious difference in the VLP distributions of different vegetation types represented by kNDVI and other VIs (Figure 2).This is primarily due to differences in the coverage areas of VPL for kNDVI and other VIs.

Attribution of Vegetation Vulnerability
For each vegetation type, the RF regression model was constructed to relate VLP to climatic, biotic, and soil structure factors.The RF regression performed well in modeling VLP (R 2 ranging from 0.53 to 0.73 across all vegetation types globally).For evergreen forests, deciduous forests, and savannas, aridity is the most important climatic factor influencing VLP (Figure 3).With an increase in the aridity index, VLP shows a noticeable decreasing trend, indicating that the drier the climate, the higher the vulnerability of vegetation to CDHEs, which is consistent with previous research (Hao et al., 2021).Arid regions typically receive less precipitation, resulting in limited soil moisture availability.This makes vegetation more susceptible to the impacts of CDHEs as they cannot access sufficient water to sustain normal growth and survival.The climatic conditions in these regions can lead to rapid evaporation of soil moisture, exacerbating the severity of water stress experienced by vegetation (Zhu et al., 2024).Additionally, arid regions often have higher evapotranspiration rates and lower relative humidity, further intensifying the loss of vegetation moisture and making it more susceptible to CDHEs.Overall, the increased vulnerability of vegetation to CDHEs in arid regions is primarily due to the combined effects of limited water supply, rapid soil moisture evaporation, and high evapotranspiration rates.This relationship is also displayed for other vegetation types.Radiation and temperature are also important factors influencing VLP in evergreen forests, where solar energy is a dominant limiting factor in vegetation growth(K.Huang & Xia, 2019;Nemani et al., 2003;Tang & Dubayah, 2017).For shrubs and grasses, precipitation is the most important climatic factor, demonstrating that water availability dominates vegetation growth, VLP shows a decreasing trend with increasing precipitation.It can be found that there appears a threshold-based manner in the relationships between VLP and precipitation, AI, temperature, and biodiversity.This could be attributed to the fact that extremely arid environments often host vegetation species that have evolved specialized survival strategies to thrive and grow under dry conditions, thereby being particularly adapted to drought conditions (Ackerly, 2004;Hõrak, 2017;Y. Liu et al., 2023).Vegetation in these regions may have adapted to the prolonged periods of drought by developing deep root systems or other physiological mechanisms to obtain water and may enter dormancy during drought periods to protect themselves.
However, some climatic variables have a more complex relationship with VLP, such as temperature and VPD.Generally, VLP initially increases with rising temperatures before decreasing again, peaking around 10°C-20°C, primarily distributed in temperate regions, with higher potential evapotranspiration rates, meaning that vegetation and soil moisture evaporate more quickly.When CDHEs occur, the impact of water limitation on vegetation may be more pronounced, as vegetation requires more water to sustain its growth and survival (Gampe et al., 2021).VLP initially increases with VPD before stabilizing.An increase in VPD signifies accelerated water evaporation from the air, leading to higher rates of soil and vegetation moisture loss, making it more susceptible to the impacts of CDHEs.In addition, elevated VPD can prompt plants to close their stomata to reduce water loss through transpiration, consequently lowering photosynthetic rates and growth rates (Liu et al., 2020;Zhao et al., 2023).Consequently, regions with high VPD typically exhibit higher sensitivity to CDHEs due to the intensified water evaporation and vegetation water stress experienced under such conditions.Although an increase in VPD may lead to higher sensitivity of vegetation to drought, once VPD reaches a certain level, the basic trend of increased VLP may plateau or slow down.This likely because as VPD continues to rise, the main drivers of VLP change to others (e.g., precipitation, biomass).
Biodiversity is a key biotic factor influencing VLP among vegetation types, for ecosystems with richer biodiversity tend to be less vulnerable to CDHEs, which has also been verified for grasses and forests using experiments and remote sensing data sets (De Boeck et al., 2018;Z. Liu et al., 2023).Therefore, restoration of species diversity emerges as a viable strategy for alleviating the negative impacts of extreme CDHEs on vegetation, particularly in drought-susceptible areas.AGB has similar effects as biodiversity against the impact of CDHEs, such that higher AGB tends to reduce VLP.Soil structures have smaller impacts on VLP for various vegetation types.The reason for this phenomenon may be that vegetation has adapted to the soil texture.In areas with low clay content, vegetation roots are more likely to extend to absorb deep soil water to resist CDHEs, while in areas with high clay content, soil moisture is less prone to loss, ensuring vegetation water supply (Gao et al., 2015;Huang & Hartemink, 2020).However, the impacts of soil textures on VLP should receive more attention.The results of VLP estimated with respect to different factors based on kNDVI, EVI, and NIRv are also shown in Figures S10-S12 in Supporting Information S1.Despite differences in the ranking of key influencing factors, the VLP and the partial dependency of each influencing factor are generally consistent, which demonstrate the results are reliable.

Conclusions
Compound drought-hot extremes (CDHEs) seriously affect the vegetation productivity of terrestrial ecosystems, which can potentially cause vegetation to shift from carbon sinks to carbon sources.Between different vegetation types examined, the VLP of grasses is the highest, followed by shrubs, savannas, and evergreen forests.Climatic factors (precipitation, aridity index, temperature, and radiation) dominate the VLP of all vegetation types examined.As expected, the drier the climate, the higher the vulnerability of vegetation to CDHEs.Energy is the dominant limiting factor for the growth of evergreen forests, while water availability dominates vegetation dynamics for grasses and shrubs.Biodiversity and biomass are two key biotic factors that influence the VLP among all vegetation types, such that ecosystems with richer biodiversity and higher biomass are less vulnerable to the impact of CDHEs.
Although the VLP shows a simple decreasing trend with the increase of AGB, the theoretical relationship between VLP and biomass could be complicated.For instance, a favorable climate (e.g., wet) in previous periods may have spurred vegetation growth beyond the ecosystem's capacity, rendering it susceptible to climate-induced stresses (Jump et al., 2021;Y. Zhang et al., 2021).Consequently, ecosystem overshoot could exacerbate drought or CDHEs, causing rapid vegetation decline and higher VLP.In the context of global warming, vegetation growth seasons advance (i.e., spring phenology changes), driving earlier consumption of soil moisture by vegetation evapotranspiration and reducing the available water for sustaining vegetation growth in subsequent seasons (Buermann et al., 2018).Spring phenology changes can also influence the VLP under drought or CDHEs through biological processes.This occurs as alterations in spring phenology directly affect the health and physiology of vegetation, consequently affecting its capacity to resist or recover from CDHEs (Li et al., 2023;Lian, Jeong, et al., 2022;Lian, Piao, et al., 2022).In addition, land-atmosphere feedbacks and other processes also play potentially crucial roles in the responses of terrestrial vegetation to CDHEs Therefore, it is needed to conduct more research with continuous satellite monitoring and improved model simulations to help better understand the ecological impacts of CDHEs.

Figure 1 .
Figure 1.Assessment and attribution framework of vegetation vulnerability responding to CDHEs using vine copula and machine learning methods.

Figure 3 .
Figure3.Vegetation loss probability distributions for different vegetation types with respect to climatic factors, biotic factors, biodiversity and soil structures, and their partial dependency to each independent variable based on a random forest analysis.Boxplots in the first row show the relative importance (%IncMSE) of different factors for five vegetation types.The second and third rows show the partial dependence between VLP and each variable estimated using random forest regression, remaining other variables unchanged.Five climatic variables are selected, including annual precipitation (Pre), temperature (Tmp), vapor pressure deficit (VPD), aridity index (AI), and net surface radiation.Three biotic variables include biodiversity (number of native species), aboveground biomass (AGB), and maximum root depth (Maxroot).Soil clay fraction (Clay_frac) is also selected.
Hence, precisely evaluating how terrestrial vegetation responds to CDHEs is pivotal in gaining crucial insights for effectively mitigating and adapting to the impacts of climate change on ecosystems.In this study, we have constructed a statistical model based on vine copula to appraise responses vegetation dynamics' reactions to CDHEs, and attributed the VLP to different climatic, biotic, and soil structure factors using a RF regression.The cumulative effects of SPEI (STI) on four different detrended and VIs_de, including kNDVI, EVI, NIRv, and SIF, are widespread, with average cumulative effect durations ranging from 4.65 to 5.70 months (3.55-4.83months)for different vegetation indices.Based on maximum correlations for SPEI-VIs and STI-VIs, the SPEI and STI at m-month time scale (SPEI_m and STI_m) are identified at each pixel.Then a statistical model is used to evaluate VLP distributions based on SPEI_m, STI_m, and four VIs_de, and the separate RF models were constructed to attribute the VLP for each vegetation type to climatic, biotic, and soil factors.The results show that high VLP is mainly found in central and southern regions of North America, eastern and southern regions of South America, Southern Africa, northern and western Europe, and northern and eastern Australia.