Drought indices are essential for studying drought events, as they allow for consistent comparison of drought characteristics across space and time. In this study, we used the Standardized Precipitation Evapotranspiration Index (SPEI), which accounts for both water supply (precipitation, P) and atmospheric water demand (potential evapotranspiration, PET) . SPEI is particularly suitable for assessing drought impacts on vegetation as vegetation is not only impacted by lack of water supply via P, but also the increase in atmospheric water demand via PET.

SPEI was computed at multiple timescales (1–24 months) using P and PET derived from W5E5 reanalysis data (see Table , ). Both P and PET were aggregated to monthly values to generate monthly SPEI outputs. A detailed description of the SPEI calculation is provided in Sect. S1 in the Supplement.

To define MYDs and NDs, we focused on SPEI calculated at a 12-month timescale (SPEI-12), as it captures multi-year drought conditions while minimizing seasonal variability. Following , MYDs are defined as periods where SPEI-12 remains below -1 for at least 12 consecutive months. The threshold of -1 was chosen to represent moderate to extreme droughts . While the start date of the drought is only identified as the first month where SPEI-12 dips below -1, this definition already takes into account the accumulated P and PET over the 12 months leading up to the start of the drought. evaluated the impact of different SPEI aggregation periods and concluded that the SPEI-12 is suitable for capturing multi-year drought dynamics.

To distinguish MYDs from shorter drought events, we define NDs as periods where SPEI-12 remains below -1 for less than 12 consecutive months. This definition ensures that each drought event is consistently categorized as either an MYD or ND, allowing us to systematically compare their respective impacts on vegetation.

In this study, the DGVM LPJmL-5 (v5.7.9, ) was used to simulate vegetation responses to MYDs. LPJmL-5 builds on LPJmL-4 and simulates both natural and agricultural vegetation, linking their growth and productivity through consistent water, carbon, and energy fluxes . LPJmL-5 makes use of the big-leaf representation for different Plant Functional Types (PFTs) that can coexist within a grid cell and compete for resources such as light, water, and nitrogen . Land cover can either be simulated dynamically or be prescribed, both of which are used in this study. In contrast, Crop Functional Types (CFTs) are always prescribed, with a distinction between irrigated and rainfed crops. Each grid cell can contain both managed and natural vegetation fractions, allowing for a realistic subgrid representation of land-use. CFTs only appear when crops are sown and grow on soil areas separate from natural vegetation . The model also includes key physiological processes such as photosynthesis, gross primary production, evapotranspiration, and plant responses to soil moisture and drought stress . The terrestrial nitrogen cycle is incorporated to account for nutrient limitations to plant growth . This makes that LPJmL-5 contains all the necessary processes to study vegetation responses to drought, though its performance may still vary across regions and vegetation types.

After evaluating how well LPJmL-5 simulates general patterns of GPP compared to MODIS GPP, we examined how well LPJmL-5 can model vegetation responses to drought. Here, we aim to understand modelled vegetation sensitivity to drought without differentiating between “normal” and “multi-year” droughts. To ensure a fair comparison of vegetation responses across regions and to remove seasonal variation, the GPP data was normalized and standardized using z-score normalization resulting in a standardized GPP anomaly (GPP_SA). The reference normalization period was chosen to match the MODIS dataset (2000–2019) to avoid errors resulting from GPP trend differences between modelled and observed data.

Vegetation can respond differently to drought events of different durations. To capture this, we used the extreme-based method developed by . This method examines how different drought durations affect vegetation across various SPEI timescales (ranging from 1 to 24 months). Specifically, we calculated the average GPP_SA response for each timescale and identified which one has the strongest negative impact during drought periods. This timescale is referred to as the dominant drought timescale and reflects how quickly vegetation responds to drought stress. Shorter timescales indicate faster, more sensitive responses, while longer timescales suggest a slower response and less drought-sensitive vegetation. For this analysis, a SPEI threshold of -1 was chosen to define atmospheric drought conditions, as suggested by .

After determining how sensitive vegetation is to drought at different timescales, we further investigated how vegetation in LPJmL-5 responds to MYDs specifically. We compared MODIS observations and LPJmL-5 simulations by analysing global spatial patterns together with temporal dynamics in selected focus regions. Vegetation responses during MYDs were quantified by using the mean GPP_SA during MYD events.

While this study primarily examines global drought patterns, it also focuses on selected focus regions to better understand local drought dynamics and how regional differences affect vegetation responses. The focus regions used in this study are: California (CAL), the Rhine-Meuse delta in western Europe (WEU), the Brahmaputra River basin spanning Bangladesh, India, Bhutan, and China (BRA), central Argentina (ARG), the Orange River basin in southern Africa (SA), and the Murray-Darling basin in Australia (AUS) (Appendix B). These regions correspond to the focus regions used by examining MYD patterns. The focus regions were selected based on the number of MYDs that have occurred between 2000–2019, geographic and climatic diversity, and variation in vegetation types, ensuring a broad representation of both water- and energy-limited regions across different climate zones.

To evaluate LPJmL-5's ability to simulate GPP, we compared modelled GPP to MODIS-derived GPP using two simulation setups: dynamic and prescribed land cover. The dynamic land cover setup allows vegetation fractions to shift in response to climate and includes background mortality, capturing potential vegetation loss . However, dynamic vegetation does not necessarily match observed vegetation types, which can lead to differences in GPP due to mismatched vegetation types. The prescribed land cover simulation maintains fixed vegetation fractions based on observed maps. This approach ensures that variations in GPP result from the model's ability to simulate vegetation productivity. However, this setup does not include background mortality, potentially causing an underestimation of vegetation loss. By comparing both setups, we can determine which one provides a better representation of satellite-derived GPP and choose the setup to be used in later analyses. Model performance is evaluated by comparing LPJmL-5 and MODIS GPP using the KGE metric (Sect. ) and time series analysis (Sect. ).

To evaluate the ability of LPJmL-5 to capture vegetation drought response, we applied the extreme-based method (Sect. ). This method identifies the dominant drought timescale, defined as the SPEI timescale (ranging from 1 to 24 months) that has the most negative impact on vegetation during drought periods.

The spatial distribution of dominant drought timescales, shown in Fig. , shows a clear difference between MODIS (a) and the LPJmL-5 simulations (b). Compared to MODIS, LPJmL-5 exhibits shorter dominant timescales, indicating a faster GPP response to drought. This difference is most pronounced in the Southern Hemisphere. LPJmL-5 also shows a skewed distribution with a clear peak around 2 months (panel c), whereas the MODIS response is more evenly distributed. Overall, these patterns highlight differences in temporal sensitivity between LPJmL-5 and the satellite-derived reference, with LPJmL-5 showing shorter response times during drought periods.

After evaluating LPJmL-5's response to droughts in general, this study now shifts its focus to MYDs and NDs.

Different choices in model parametrisation and setup can lead to significant differences in modelled GPP. For example, both the choice of meteorological forcing dataset (ERA5 vs. W5E5 ) and the temporal resolution (monthly vs. daily) have a significant impact on simulated GPP patterns (Sect. S4). This indicates that it is important for future drought oriented studies to make clear choices with regard to the modelling setup. In this study, we selected the daily W5E5 dataset , as it is bias-corrected and offers a high temporal resolution and the total global GPP was closest to that found in observational datasets.

Our analyses also showed that GPP output is highly sensitive to the radiation inputs (Sect. S4). GPP values varied substantially depending on whether the model calculated radiation internally or used observed radiation data. We found that errors in GPP are higher from internal estimation, compared to directly providing longwave and shortwave radiation from observations.

Using dynamic land cover in LPJmL-5 resulted in systematically overestimated GPP values across most regions (Fig.  and Sect. S5), also noted in . We argue that this difference is the result of a mismatch between modelled and observed vegetation distributions (Appendix C and Sect. S5), which is to be expected as the land cover is not constrained to the observed land cover. For example, LPJmL-5 simulates tree cover in parts of Western Europe where forests could exist naturally, but are absent today due to long-term human management. To better match satellite-derived GPP and reduce these discrepancies, we prescribed land cover fractions, rather than letting the model simulate land cover dynamically.

Prescribing land cover in LPJmL-5 improved the agreement with MODIS GPP, as indicated by higher KGE values in Fig. . This is consistent with studies showing that accurate land cover representation is essential for reproducing observed productivity patterns , and that constraining DGVM simulations with satellite-derived observations (including land cover) enhances their ability to capture observed spatial patterns of vegetation and carbon dynamics .

Breaking down the differences between LPJmL-5 and MODIS GPP during MYDs by vegetation type further highlights a varying model performance (Fig. ). Tropical trees show the largest variability in drought response, showing both under- and overestimations of impacts. Similar results were found by , who showed that terrestrial biosphere models underestimate drought impacts on tropical forest biomass. Temperate trees generally align more closely with observations but still underestimate negative drought impacts. This is consistent with , who found that land carbon models underestimate the severity and duration of drought effects on productivity across temperate forests in the US and Europe. Similarly, observed that models frequently overpredict GPP under dry conditions, further pointing to an underestimation of drought stress. Boreal trees and grasses are more prone to underestimated positive responses. Croplands, despite poorer agreement in absolute GPP dynamics, show comparatively better agreement in drought-induced anomalies. This suggests that while LPJmL-5 struggles to reproduce the magnitude of the GPP, it is able to capture relative deviations during drought events. Such behaviour indicates that biases in absolute GPP do not necessarily translate into errors in anomaly-based responses.

A key limitation underlying these issues appears to be how vegetation mortality is represented in LPJmL-5. In the model, mortality is only driven by low productivity and heat stress, and can not be directly induced by prolonged water deficits. Drought-induced mortality occurs only if plants fail to produce enough biomass to sustain themselves. However, this natural mortality is capped by a maximum annual rate , which restricts how much vegetation can be lost due to die-off each year. Moreover, this mortality mechanism is only active in the dynamic land cover setting, and is not applied when land cover is prescribed. Since the simulations in this study use the prescribed land cover setting, no vegetation is lost due to this background mortality. This means that while GPP may decline sharply during droughts, vegetation does not die, causing unrealistically rapid recovery periods when precipitation (temporarily) returns. Although LPJmL-5 includes a heat-stress mortality mechanism, it only applies to boreal trees . As a result, vegetation decline during MYDs is not directly driven by water shortages, which prevents large-scale die-off and causes the model to underestimate total vegetation loss. These limitations reduce the model's ability to simulate the full severity of vegetation decline, especially during MYDs. This issue has been noted by , who reported that the lack of realistic mortality processes leads to an overestimation of simulated biomass and contributes to biases in GPP. Similarly, found that models which combine carbon starvation with hydraulic failure mechanisms are better able to reproduce observed drought-induced tree mortality, and highlighted the need to account for both stomatal and non-stomatal limitations to realistically capture plant responses to water stress. further showed that including hydraulic processes in LPJ-GUESS improves simulations of drought responses. Together, these findings suggest that LPJmL-5 likely underestimates the true impact of drought on vegetation.

Despite underestimating long-term vegetation loss, LPJmL-5 tends to overestimate vegetation decline at the start of droughts, as shown in Figs.  and . LPJmL-5 does include a water stress factor that influences the growth, which results in a rapid decline in GPP. This stress is determined by soil moisture availability and fixed rooting depths for each PFT, without accounting for dynamic rooting strategies, capillary rise or explicit groundwater processes . However, in reality, vegetation often withstands short-term stress periods through drought survival strategies . At the same time, LPJmL-5 also simulates unrealistic recoveries during the drought period. Since there is no mortality, vegetation can survive during a drought period, resulting in no loss of biomass. When conditions improve temporarily, for example due to brief rainfall, the model quickly simulates regrowth. This leads to an unrealistic drought recovery effect. This is in contrast with real vegetation, which often suffers from long-term impacts . Both the quick decline and recovery are consistent with findings by , who show that land carbon models underestimate the severity and duration of drought impact on plant productivity.

Together, the combination of overly sensitive short-term GPP decline, limited representation of long-term vegetation loss, and unrealistically fast recovery during drought periods results in a mismatch in LPJmL-5's simulated vegetation drought response. While prescribing land cover improves the overall GPP performance, these structural issues in the mortality representation limit the model's ability to simulate realistic vegetation dynamics during drought events. This is especially limiting when the analysis is focused on multi-year drought events, which have been shown to cause higher vegetation mortality rates and are expected to become more frequent under climate change .

To improve the overall quality and reliability of our simulations, we used the state-of-the-art dynamic vegetation model LPJmL-5. This model is publicly available with an extensive model description and evaluation . LPJmL-5 provides a multitude of vegetation-related outputs on ecological, hydrological, and agricultural components, making it suitable for studying vegetation responses at the global scale.

Drought events were identified using the Standardized Precipitation Evapotranspiration Index (SPEI), which incorporates both precipitation (P) and potential evapotranspiration (PET), allowing it to better reflect the impact of temperature on drought severity . As a standardized index, SPEI also enables consistent spatial comparisons of drought conditions at the global scale. By using a 12 month aggregation period for the SPEI, we also remove seasonal variations and focus on capturing MYDs.

As a standardized metric, SPEI characterizes drought conditions relative to the local climatic baseline, reflecting periods that are characterized by below normal water availability. In energy-limited regions, such as the tropics or boreal forests, negative SPEI values may occur even when absolute water availability remains relatively high . To ensure these droughts correspond to meaningful reductions in water availability, we compared soil water content (SWC) during droughts and normal conditions (Sect. S2), finding consistently lower SWC across soil layers. This supports the interpretation that identified MYDs correspond to periods of meaningful vegetation stress, even in energy-limited regions.

While there is currently no universally agreed-upon definition of MYDs, we follow the approach of (see Sect. ), which was also adopted in and . SPEI values are sensitive to the method and input data used for calculating PET. We compared results using ERA5 and W5E5 and found differences in the number of MYDs identified for the period 2000–2019 (Sect. S4). Nonetheless, the MYDs identified in our focus regions are broadly consistent with those reported in other studies , despite minor differences in timing.

To evaluate vegetation responses, we used monthly MODIS GPP as a proxy for ecosystem productivity, representing the amount of carbon captured from the atmosphere through photosynthesis . Previous drought studies have often relied on the EVI , but EVI is satellite-derived and not directly available from LPJmL-5. Because MODIS GPP and EVI are strongly correlated (; see also Sect. S3), including during multi-year droughts , MODIS GPP can be considered a suitable proxy for vegetation activity.

Several studies further indicate that MODIS GPP provides a meaningful representation of productivity across biomes, with most errors linked to input datasets rather than the algorithm itself. Performance is generally good in temperate and mixed forests, while errors are larger in evergreen broadleaf forests and tropical regions . showed that much of the error originates from FPAR, land-cover misclassification, and light-use efficiency parameters rather than meteorology. Systematic underestimation has also been reported in grasslands . These studies agree that cloud contamination and uncertainties in tropical regions remain key challenges. Taken together, the strong correspondence between MODIS GPP, EVI, and solar-induced fluorescence (Sect. S3) supports the use of MODIS GPP as our validation dataset for LPJmL-5 GPP, while acknowledging that performance is weaker in tropical ecosystems and that results should be interpreted with these limitations in mind.

Although other sources of GPP data exist, such as site-level flux tower measurements, we choose to use satellite-derived MODIS GPP because it provides consistent, observation-based estimates at the global scale. To maintain observational consistency and avoid mixing in other models, we did not include GPP estimates derived from other land surface models as they are likely based on similar assumptions and equations as the LPJmL-5 model. Nonetheless, MODIS GPP is subject to uncertainty, particularly in tropical regions (Sect. S3), and this should be kept in mind when interpreting the results.