Global datasets harmonized to a 5 arcmin (0.0833°×0.0833°) spatial grid were first divided by land-cover type and then assigned to climate event classes. Figure 1 (left) outlines the workflow for land-cover classification, and Fig. 1 (right) summarizes the datasets and their functions. The datasets used in the analysis were divided into three groups: (1) those used to derive the response variables (resistance and resilience), (2) those used to compute explanatory variables, and (3) those used for classification (land cover and SPEI).

All datasets were harmonized to a 5 arcmin spatial resolution. Time-series analyses were conducted over the period 1982–2015, excluding 1990–1994. NDVI during this interval was influenced by NOAA-11 orbital drift, which shifted the satellite's equatorial crossing time by ∼ 3.7 h (Li and Brown, 2017) and altered the solar geometry during acquisition. This period also included the Pinatubo eruption, thereby introducing additional radiative anomalies. To avoid these confounding effects, NDVI data for 1990–1994 were removed, and all analyses were conducted for 1982–1989 and 1995–2015. As NDVI quality declines at high latitudes, the analysis domain was restricted to 60° N–60° S (Beck et al., 2006; Justice et al., 1985).

Terrestrial regions worldwide were included, except for land-cover types represented by very few grid cells (Channan et al., 2014). The excluded classes were Closed Shrublands (CS), Permanent Wetlands (PW), Urban and Built-up areas (U), Barren land (B), and Snow and Ice (SI). The retained land-cover classes included Evergreen Needleleaf Forest (ENF), Evergreen Broadleaf Forest (EBF), Deciduous Needleleaf Forest (DNF), Deciduous Broadleaf Forest (DBF), Mixed Forest (MF), Open Shrublands (OS), Woody Savannas (WS), Savannas (S), Grasslands (G), Croplands (C), and Cropland/Natural Vegetation Mosaic (C/N). Land-cover data were used at a 5 min spatial resolution.

The climate conditions for each grid cell were classified using the Standardized Precipitation–Evapotranspiration Index (SPEI, Vicente-Serrano et al., 2010) at 3-, 6-, and 12-month time scales for 1982–2015. Seven categories were defined (Fig. 3): extremely dry (SPEI < -1.645), severely dry (SPEI < -1.28), moderately dry (SPEI < -0.67), normal (-0.67 ≤ SPEI ≤ 0.67), moderately wet (SPEI > 0.67), severely wet (SPEI > 1.28), and extremely wet (SPEI > 1.645). These thresholds correspond to the probabilities under a standard normal distribution.

SPEI was computed using the SPEI R package (Begueria and Vicente-Serrano, 2017) using precipitation from GPCC v7.0 (1° resolution; Schneider, 2015) and potential evapotranspiration from CRU-TS 4.01 (0.5° resolution; University of East Anglia Climatic Research Unit et al., 2017). To obtain complete time series across all SPEI scales, climate data from 1981 to 2015 were used. SPEI results were then aligned with the NDVI datasets after the removal of 1990–1994.

Resistance and resilience were derived from the NDVI3g dataset (GIMMS; Pinzon and Tucker, 2014) at a 5 arcmin resolution. NDVI from 1990 to 1994 was excluded. The maximum biweekly NDVI value was retained for each month. For each grid cell, the month that most frequently contained the annual NDVI maximum between 1982 and 1989 and between 1995 and 2015 was identified (Fig. 4). This “peak month” was treated as a proxy for annual vegetation productivity.

In this study, resistance and resilience were defined as event-based proximity indices to the normal-year vegetation condition. Resistance quantifies how close vegetation activity during the climate-event year was to the normal-year level, whereas resilience quantifies how close vegetation activity in the year following the event was to the normal-year level. These indices were calculated as follows: 1Resistance=1NDVIm-NDVIe+12Resilience=1NDVIm-NDVIe+1+1 where NDVIm  is the mean NDVI in normal years (-0.67 ≤ SPEI < 0.67), NDVIe is the NDVI during the climate-event year, and NDVIe+1 is the NDVI in the year after the event.

Under this formulation, a value of 1 indicates vegetation activity comparable to the normal-year condition. Values below 1 indicate reduced vegetation activity relative to normal years, whereas values above 1 indicate above-normal vegetation activity. Thus, the indices are centered around 1 and provide a direct interpretation as proximity to normal vegetation activity. This property is useful for global pixel-wise mapping, land-cover comparisons, and machine-learning analyses because the same reference value can be used across pixels, event categories, and land-cover types.

We adopted this formulation because the objective of this study was to compare global vegetation responses across land-cover types and hydroclimatic-event categories using response variables with a direct and consistent interpretation. Alternative stability metrics emphasize different aspects of disturbance response. For example, ratio-based metrics such as the Isbell-type formulation can become extremely large or undefined when the event-year or post-event vegetation index is very close to the normal-year condition (Isbell et al., 2015). Bounded metrics such as the Orwin and Wardle-type formulation are useful for evaluating departure from a control condition, but above-normal vegetation activity is treated as a deviation from the normal state rather than as enhanced vegetation activity (Orwin and Wardle 2004; Liu et al., 2022). Log-response metrics are useful for quantifying proportional change, but they are centered around 0 and do not directly indicate whether vegetation activity has returned to the normal-year condition (Xu et al., 2022; Yan et al., 2025).

Therefore, our indices should be interpreted as proximity-to-normal vegetation activity indices, rather than as direct substitutes for all previously proposed resistance and resilience metrics. To clarify the behavior of our metrics relative to commonly used alternatives, we provide numerical comparisons with Isbell-type, Orwin and Wardle-type, and log-response formulations in Tables S1 and S2 in the Supplement. These comparisons show how each metric responds when vegetation activity is equal to the normal condition, declines below it, recovers to it, or exceeds it.

Autoregressive formulations were not used because the analysis incorporated categorical and non-time-series explanatory variables.

NDVI was used rather than EVI because the main objective of this study was to analyze resistance and resilience over a long temporal period and to capture as many extreme dry and wet events as possible. The GIMMS NDVI3g dataset provides a globally consistent vegetation record beginning in 1982, whereas a directly comparable long-term EVI product is not available for the full study period. Because EVI requires spectral bands that are not available from the AVHRR-based GIMMS NDVI3g record, an EVI-based analysis could not be conducted over the same temporal extent.

We acknowledge that EVI has important advantages over NDVI, including reduced saturation in high-biomass vegetation and improved correction for soil and atmospheric background effects. Therefore, the use of NDVI may affect the estimated magnitude of resistance and resilience, particularly in dense forests, sparsely vegetated systems, and regions where soil background effects are important. To evaluate this limitation, we conducted an additional MODIS-based comparison using both NDVI and EVI over the same period and added the results to the Supplement (Figs. S2 and S3). This comparison was used as a sensitivity analysis to assess how the choice of vegetation index may influence the estimated resistance and resilience values.

Twenty-two explanatory variables were compiled and grouped into four categories: climatic, abiotic, biodiversity-related, and land-use history/irrigation variables (Table 1 and Supplement).

Climatic Factors: Annual precipitation (P: Schneider, 2015), temperature (T: University of East Anglia Climatic Research Unit et al., 2017), and shortwave radiation (R: European Centre for Medium-Range Weather Forecasts, 2012) were obtained for the event year, the preceding year (-1), and the following year (+1). The +1 variables were used only for resilience models, yielding 19 explanatory variables for resistance and 22 for resilience models.

Abiotic Factors: Elevation (Fischer et al., 2008) and six soil properties (Global Soil Data Task Group, 2000) – bulk density, wilting point, plant-available water capacity, thermal capacity, soil organic carbon, and soil nitrogen – were included because of their roles in water and nutrient regulation.

Biodiversity-Related Factors (Ellis et al., 2012): Anthropogenic species increase (ASI) was estimated using alien species distributions, crop species data, and urban area extent. Anthropogenic species loss (ASL) was derived from vegetation patterns and habitat-based models. Anthropogenic species richness (ASR) was calculated as follows: 3ASR=N-ASL+ASI where N is the potential species richness, estimated using the global model by Kreft and Jetz (2007). These variables are modeled biodiversity-related proxy variables and should not be interpreted as direct observations of local biodiversity change.

Land-Use History and Irrigation: HYDE 3.1 data (Klein Goldewijk et al., 2011) were used to estimate the duration (years) that each grid cell had been under cropland or pastureland for the past 8000 years. The extent of irrigation was obtained from the Global Map of Irrigation Areas (Siebert et al., 2013).

Non–time-series variables (soil, elevation, land-use history, ASR/ASI/ASL) were normalized to a 0–1 scale using the global minimum and maximum values: 4Y=X-Xmin⁡XMax-XMin

X: Data for any given grid

Y: Normalized data (0–1 range) for the corresponding grid

XMax: Maximum value across all grids

XMin: Minimum value across all grids

Time-series climatic variables (T, P, R) were aggregated to annual values and normalized within each grid cell across 1982–1989 and 1995–2015 as follows: 5Yi=Pi-Pmin⁡PMax-PMin

Pi: Data for year i

Yi: Normalized data (0–1 range) for year i

PMax: Maximum value across all years

PMin: Minimum value across all years

Grids outside 60° N–60° S were excluded because of NDVI reliability limitations.

The relationship between the explanatory variables and vegetation resistance or resilience was assessed using a Light Gradient Boosting Machine (LightGBM) model (Shi et al., 2025). LightGBM was selected because it can efficiently capture nonlinear relationships and interaction effects among multiple climatic, edaphic, topographic, biodiversity-related, and land-use predictors without requiring a priori assumptions about functional forms, which is particularly important for ecosystem–climate analyses where threshold behavior and scale-dependent responses are common. In addition, its histogram-based learning algorithm and leaf-wise tree growth strategy enable high computational efficiency, making it well suited for large-scale, high-dimensional environmental datasets.

Another advantage of LightGBM is its robustness to multicollinearity among predictors, such as correlated climate variables or indices at different accumulation time scales, allowing all explanatory variables to be retained without prior dimensionality reduction. Furthermore, LightGBM can internally handle missing values by learning optimal split directions for missing observations, which reduces the need for explicit gap-filling procedures and improves robustness when working with long-term remote-sensing and climate datasets. In the present analysis, however, rows containing missing or infinite values were removed before model fitting to ensure consistent input data across repeated model runs.

For each model configuration, the data were randomly divided into training and testing datasets using an 80/20 split. The models were trained with up to 500 boosting iterations, and early stopping was applied if the prediction accuracy did not improve for 50 consecutive iterations. Model performance was evaluated using the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R2). These settings follow the attached robustness scripts, in which TEST_SIZE = 0.20, NUM_BOOST_ROUND = 500, and EARLY_STOPPING_ROUNDS = 50 were used.

To quantify the relative contribution of individual explanatory variables, feature importance was evaluated using the gain metric: 6Feature Importance (Gain)j=∑t∈TjΔLosst where Tj denotes the set of decision tree splits that involve feature j, and ΔLosst represents the reduction in the loss function attributed to split t. Because this metric directly reflects the degree to which each variable contributes to reducing the model error, it is a practical and interpretable indicator of variable influence in gradient boosting models. In each model run, gain-based feature importance was calculated for all predictors, normalized by the total gain, and the top five predictors were retained for the robustness summary.

To assess the robustness of the LightGBM results, we repeated the model training under multiple random seeds, tree-structure hyperparameter settings, and learning-rate settings. Learning rate is also a hyperparameter in LightGBM; here, we treated it separately from the two tree-structure parameter settings in order to explicitly evaluate sensitivity to the step size of boosting updates. For each learning-rate setting, we used three random seeds and two tree-structure hyperparameter settings, resulting in six independent model runs per model configuration. The three random seeds were 0, 1, and 2. The two tree-structure parameter settings were an original setting and a simpler-tree setting. The original setting used num_leaves = 23, min_data_in_leaf = 1, feature_fraction = 1.0, bagging_fraction = 1.0, bagging_freq = 0, and max_depth = -1. The simpler-tree setting used num_leaves = 15, min_data_in_leaf = 20, feature_fraction = 0.9, bagging_fraction = 0.9, bagging_freq = 1, and max_depth = 6.

We further evaluated the sensitivity of the results to the learning-rate hyperparameter by repeating this six-run procedure under three learning-rate settings: 0.1, 0.2, and 0.3. This procedure was applied to all 462 model configurations, defined by the combination of three SPEI timescales, 11 land-cover types, seven event categories, and two response variables. Thus, each model configuration was evaluated using 18 runs in total, resulting in 8316 model runs across all configurations.

For each learning-rate setting, we summarized the six runs and identified the variable that most consistently showed the highest gain-based importance. We then compared the selected top variables across the three learning-rate settings and adopted the variable that was most frequently selected as the dominant predictor. If the same variable was selected under all three learning-rate settings, it was regarded as a robust top predictor. If the same variable was selected under two of the three learning-rate settings, the result was marked with a triangle symbol to indicate moderate robustness. If different variables were selected under all three learning-rate settings, the result was recorded as “nan,” indicating that no single dominant predictor was consistently identified.

This repeated-modeling procedure was used to reduce the dependence of feature-importance interpretation on a single random seed, tree-structure hyperparameter setting, or learning rate, and to provide a more systematic assessment of uncertainty in the selected top predictors.

Figure 5 shows the resistance and resilience estimated from SPEI-3, separated by climatic conditions, with results for extreme drought and extreme wet conditions across all land cover types. To improve the readability of the displayed distributions, the y-axis range in Fig. 5 was restricted to 0.8–1.2. However, values outside this range accounted for only a very small proportion of the data: 0.418 % for SPEI-3, 0.417 % for SPEI-6, and 0.412 % for SPEI-12. Thus, most resistance and resilience values were distributed within the 0.8–1.2 range (Table S3 in the Supplement).

Figure 6 shows the spatial distribution of resistance and resilience based on SPEI-3 for all climate categories. The full set of results for all SPEI timescales and climatic conditions is presented in Fig. S4 in the Supplement.

In Fig. 5, the median values indicate that, under both drought and wet conditions, deciduous needle-leaf forests (DNF) exhibit low resistance and low resilience under extreme drought conditions. Open shrublands (OS), grasslands (G), and croplands (C) also showed resistance and resilience values below 1 under extreme drought conditions, similar to DNF, but their resilience values were not as low as those of DNF.

Drought-related patterns in Fig. 6a and b show that resistance declines most strongly in dryland regions, especially in areas classified as grasslands (light purple) and savannas (orange) in Fig. 2. These regions include sub-Saharan Africa, the interior of the Eurasian continent, and the central Great Plains of North America, where resistance markedly decreases. Reductions in resilience were weaker and less spatially coherent than those in resistance. Under wet conditions, partial decreases in resistance appear in some high-latitude and temperate regions. By stratifying resistance and resilience according to the climate category, the maps revealed that ecosystem responses to extreme events vary strongly among regions (Fig. 6), underscoring the need to compare ecosystem stability under comparable climatic conditions rather than assuming spatially uniform responses.

For each of the 11 land-cover types, the most important variables in the LightGBM models for resistance and resilience were identified across three SPEI time scales (3, 6, and 12 months) and seven water-balance categories (Table 2). In total, 462 (11 × 2 × 3 × 7) model configurations were evaluated. Values below 1.00 are indicated by a dagger symbol (†), whereas values equal to or greater than 1.00 are shown in bold.

The analysis was stratified by SPEI time scale, land-cover type, climatic event category, and response variable. The feature importance of each explanatory variable in the LightGBM models for resistance and resilience was summarized for each land-cover type and climate category (Tables S4 and S5 in the Supplement).

To assess the robustness of model performance, six model runs were treated as one set, and this procedure was repeated three times using different learning-rate settings. Across these three sets of analyses, the model that consistently showed the highest R2 was the SPEI-3 model for resistance under extreme drought in DNF, with an R2 of 0.717. In contrast, the model that consistently showed the lowest R2 was the SPEI-6 model for resistance under normal conditions in EBF, with an R2 of 0.097. Because each set included 462 model results, a total of 1386 model-performance records were generated. These results are provided through the link included in the Supplement.

Under Normal climatic conditions, temperature-related variables were consistently identified as the leading predictors of ecosystem stability, with the exception of resilience in DNF under SPEI-12, for which R+1 had the highest importance. For resistance, the temperature in the event year (T) was the most frequently selected driver, appearing in 18 of the 33 combinations (≈ 54.5 %). For resilience, the temperature in the post-event year (T+1) was selected in 18 combinations (≈ 54.5 %).

These patterns indicate that temperature-related variables were the leading predictors of NDVI-based vegetation stability metrics under normal conditions. The temporal shift in the selected temperature predictors is consistent with the definitions of the two indices: resistance, which represents proximity to normal vegetation activity during the event year, was more strongly associated with event-year temperature, whereas resilience, which represents proximity to normal vegetation activity in the year following the event, was more strongly associated with post-event temperature.

For ENF, EBF, DNF, DBF, and MF, across these forest-dominated land cover types, climate variables were consistently identified as leading predictors, with temperature most frequently showing the highest gain-based importance. As shown in Table 2, DNF often showed temperature or radiation as the variable with the highest feature importance, whereas DBF frequently showed both temperature and precipitation among the top-ranked drivers. These feature-importance patterns suggest that climatic variables provide substantial predictive information for NDVI-based resistance and resilience in forest systems, with radiation showing high predictive importance in DNF and precipitation showing relatively high importance in DBF.

The effects of drought on OS, S, G, C, and C/N were especially pronounced under severe and extreme drought conditions. Under Severe or Extreme droughts, non-climatic variables – ASI, ASL, ASR, and elevation – often emerged as the highest feature importance rather than climate variables. These land cover types share relatively sparse or structurally constrained vegetation, which may limit their buffering capacity against water stress.

In OS and G, ASL had the highest feature importance under extreme drought at the SPEI-3 timescale, and resistance values were below 1. In C, ASI was selected as the leading predictor under extreme drought (Table 2), suggesting an association between modeled anthropogenic species-increase proxies and NDVI-based stability in croplands. In C/N, ASR showed the highest contribution to resistance under SPEI-6.

Under Extreme or Severe wet conditions, both resistance and resilience exceeded 1 in these sparsely vegetated or managed land cover types, indicating enhanced vegetation functioning. In such cases, temperature-related variables are the most frequently identified leading drivers. ASR was the most influential variable for WS resilience under SPEI-3 and SPEI-6, and elevation showed the highest feature importance for G resistance under SPEI-3.

Our results revealed that vegetation resistance declines markedly during droughts, particularly in arid and semi-arid regions (OS, S, G), whereas resilience exhibits weaker and less spatially coherent declines. Deciduous needle-leaf forests (DNF) consistently show low resistance and resilience, indicating high sensitivity to hydroclimatic variability. This finding is consistent with previous studies reporting reduced ecosystem stability or increased ecosystem degradation in DNF under drought and precipitation variability (Hai et al., 2025; Li et al., 2021; Nogovitcyn et al., 2023). Regions where both stability components are low are likely to experience heightened ecological vulnerability, as reduced resistance and resilience imply a diminished capacity to withstand and recover from climatic disturbances (Yao et al., 2024). Accordingly, our findings suggest that DNF may become increasingly vulnerable under future climate change scenarios.

Sparse vegetation systems, such as open shrublands (OS), grasslands (G), and croplands (C), also exhibit low resistance but maintain comparatively high resilience (Fig. 5b and d). Mapping stability across climatic categories demonstrates substantial spatial heterogeneity in ecosystem responses, highlighting that ecosystems do not respond uniformly to droughts or heavy rainfall. Even under similar climatic conditions and within the same land-cover type, resistance and resilience responses to extreme climatic events can vary markedly. This spatial heterogeneity is consistent with previous findings that sparse vegetation, although highly vulnerable to immediate drought impacts, often exhibits rapid post-drought recovery (Ponce-Campos et al., 2013; Seddon et al., 2016). The pronounced spatial heterogeneity observed in this study underscores the necessity of evaluating ecosystem stability within comparable climatic contexts rather than generalizing responses across biomes or land cover types.

Temperature-related variables were identified as leading predictors of both resistance and resilience under normal climatic conditions. Although this result partially overlaps with the findings of Huang and Xia (2019), a direct comparison is difficult because the climatic gradient classifications differ. The high predictive importance of temperature-related variables is consistent with studies reporting temperature-driven vegetation shifts and abrupt transitions in ecosystem functioning (Berdugo et al. 2022). Resistance was most strongly associated with temperature in the event year (T) in forest-type land-cover areas. This pattern is consistent with Vitasse et al. (2019), who showed that late spring frost can substantially reduce current-year growth in forest tree species, indicating that temperature anomalies during the event year may be associated with changes in growth-related resistance. At the same time, antecedent temperature may also be important, given the concept that ecosystems undergo “pre-conditioning” through prior thermal and soil moisture conditions before extreme events (De Keersmaecker et al., 2015; Fu et al., 2024). Thus, resistance may reflect both the direct impact of event-year temperature and the legacy effects of climatic conditions in the preceding year.

Resilience, defined here as proximity to normal vegetation activity in the year following the event, was most strongly associated with temperature in the year after the extreme event (T+1). One major reason for this is likely that the resilience metric explicitly incorporates NDVI in the year following the event. In addition, although this interpretation remains hypothetical, the result may partly reflect the temperature sensitivity of root and shoot regrowth processes during post-disturbance recovery (Wu et al., 2025). Taken together, these findings suggest that the timing of temperature-related predictors differ between resistance and resilience. However, because this study considered only a one-year window before and after each extreme event, longer-term evaluations are needed to better capture ecosystem-specific response times and delayed recovery processes.

Under Severe and Extreme drought conditions, non-climatic variables, particularly modeled anthropogenic species increase (ASI), anthropogenic species loss (ASL), anthropogenic species richness (ASR), and elevation, frequently emerged as dominant predictors in sparsely vegetated areas. This pattern suggests that these variables provide additional predictive information when water availability is extremely limited. In such environments, even small changes in species composition or vegetation structure may be associated with differences in resistance and resilience to severe climatic stress. This predictive pattern is broadly consistent with previous research linking higher biodiversity to enhanced ecosystem stability under extreme climatic events (Isbell et al., 2015), but modeled anthropogenic species-loss proxies should be interpreted as predictive associations rather than direct evidence of increased vulnerability in sparse vegetation systems. The study by Isbell et al. (2015) was based primarily on grassland data from Europe and North America in the Northern Hemisphere, and therefore does not provide direct support for the global-scale patterns observed in the present study. In addition, the biodiversity-related variables used here may have been influenced more strongly by relatively recent land-cover conditions than by the long-term land-use history considered in this study. Thus, the apparent importance of ASI, ASL, and ASR should be interpreted as an indication of potential biodiversity-related effects rather than as direct evidence of causal mechanisms.

Taken together, the results indicate that more detailed monitoring and analysis are needed in sparsely vegetated regions, where small changes in vegetation structure and species composition may strongly affect ecosystem stability under severe drought. Furthermore, as shown in Fig. S2, the comparison between EVI and NDVI indicates relatively large differences between the two vegetation indices in OS and G. This suggests that the choice of vegetation index may affect the patterns observed in sparsely vegetated areas. To enhance the robustness of such interpretations, future analyses should incorporate multiple vegetation indices where possible. In the present study, however, NDVI was selected because its long-term availability enabled us to include the largest possible number of extreme events within the analysis period.

In croplands, ASI emerged as the most important variable under Extreme drought conditions. This result may suggest that management-dependent species turnover, such as changes in crop varieties or cultivated species composition, can influence resistance during drought events. This interpretation is broadly consistent with agronomic studies showing that cultivar selection can play an important role in modulating crop productivity and stability under drought conditions (Lobell et al., 2014). Collectively, these results point to a possible interaction between climatic and biotic factors and suggest that species composition and land management practices should be considered when assessing ecosystem stability.

Overall, the results indicate that ecosystem stability shifts substantially along the climatic severity gradient. Under Normal conditions, temperature-related variables were frequently selected as leading predictors of both resistance and resilience. Under severe and extreme droughts, the relative predictive importance of climatic variables decreased in several sparsely vegetated systems, whereas non-climatic variables, including modeled anthropogenic species-change proxies and elevation, showed increased predictive importance for vegetation stability metrics (Fig. 7). These findings demonstrate that climatic variables alone are insufficient for predicting vegetation stability under intense water limitation and that ecological and structural traits must be incorporated into the predictive frameworks.

Despite the strengths of this study, several limitations should be acknowledged. First, the temporal discontinuity caused by excluding NDVI observations from 1990 to 1994 may have reduced the continuity of long-term trends. However, this exclusion was necessary to avoid potential artifacts associated with NOAA-11 orbital drift and the eruption of Mt. Pinatubo.

Second, the biodiversity-related indicators used in this study, namely anthropogenic species increase (ASI), anthropogenic species loss (ASL), and anthropogenic species richness (ASR), as well as the long-term land-use reconstruction dataset HYDE 3.1, are based on modeled estimates and therefore contain inherent uncertainties. These uncertainties may have influenced the relative importance of explanatory variables estimated by the models. More specifically, ASI and ASL are model-based indicators of anthropogenic species gains and losses estimated by Ellis et al. (2012), using native vascular plant species richness (N) derived from Kreft and Jetz (2007) as a baseline. ASL was estimated mainly from habitat loss using species–area relationships, whereas ASI was calculated as the sum of species increases associated with exotic species invasions, crop species, and ornamental plants. Therefore, these variables should not be interpreted as direct differences between observed and predicted current species richness. Rather, they should be regarded as modeled indicators representing potential anthropogenic changes in plant species richness.

Third, LightGBM identifies statistical associations rather than causal mechanisms. Although the present analysis used a large global dataset, spatial autocorrelation among grid cells may still have affected model performance and the estimated importance of explanatory variables. Therefore, the identified drivers should be interpreted as variables statistically associated with resistance and resilience, rather than as direct causal factors.

Fourth, relying solely on NDVI as a proxy for vegetation functioning may underestimate productivity changes in high-biomass ecosystems because of canopy saturation. To address this concern, we compared the results based on NDVI and EVI and presented these comparisons in Figs. S2 and S3 in the Supplement. Nevertheless, the choice of vegetation index may still influence the observed patterns, particularly in ecosystems where vegetation signals are weak or where index-specific sensitivities differ.

Fifth, climatic variables were aggregated at the calendar-year scale. This may introduce uncertainty in regions with pronounced wet–dry seasonality, because the use of calendar years can artificially split wet or dry seasons across different years. As a result, climate signals relevant to ecosystem responses may be weakened or distorted. This issue is particularly important in monsoonal and seasonally dry ecosystems, where ecological responses are often closely linked to seasonal water availability.

Finally, the resistance and resilience metrics used in this study represent only part of the broader concept of ecosystem stability. Ecosystem stability can encompass a wide range of dimensions, including temporal variability, recovery time, persistence, and potential regime shifts. In contrast, the present study evaluated the impact of each climatic event independently, without explicitly considering the continuity or sequence of multiple events. In addition, although ecosystem responses should ideally be evaluated over multiple years before and after each event, this study focused only on the year preceding and the year following each event in order to include as many analyzable extreme events as possible. Therefore, the aspects of ecosystem stability that can be discussed in this study are necessarily limited to event-based resistance and short-term resilience.

Future research incorporating longer observational records, improved biodiversity and soil datasets, alternative vegetation indices, and more detailed land-use histories will help address these limitations. Such efforts will enable more robust assessments of ecosystem stability under extreme climatic events and provide a more comprehensive understanding of the mechanisms governing ecosystem responses to climate extremes.