Introduction

Drylands are characterized by climate variability and water scarcity because of the low rainfall and high evapotranspiration from surface1. The United Nations Convention to Combat Desertification defined drylands as the areas where the Aridity Index (AI) is less than 0.651. Drylands account for about 41% of the earth’s land surface2 and up to 44% of cultivated systems in the world are in drylands3. Drylands support more than 33% of the world’s population2. In addition to the immediate provision of food, drylands provide a broader set of ecosystem services. The ecosystems of drylands are the main influence on the terrestrial carbon sink interannual variability and trend4. Dryland ecosystems are quite vulnerable and sensitive to external influences, for example, human activities and climate variabilities5. Damage to dryland ecosystems will have a considerable influence on human life. Meanwhile, due to the increased human demands and climate change, dryland ecosystems are facing severe problems5,6,7. It is estimated around 10%-20% of drylands have already degraded and new degradation is still happening every year8. Degradation of drylands will lead to the loss of biodiversity, damage to ecological integrity, and threat to food security9. Especially poor people in dryland areas with a high-risk of desertification will suffer more natural disasters10.

AI is the ratio of annual precipitation to potential evapotranspiration, which is widely used to define drylands and measure atmospheric aridity11,12. AI shows a decreasing trend in the past few decades13. Which indicates drier atmospheric conditions. Drylands expansion has also been observed during the past sixty years14. Further, over drylands the atmosphere is projected to become more arid according to the AI under all future climate scenarios11,14,15. This, in turn, is projected to increase the dryland extent with higher emission scenarios leading to larger dryland expansions16. The high emission scenario (RCP8.5) in CMIP5 (Coupled Model Intercomparison Project Phase 5) is projected to produce a 10% expansion in dryland area by 2100 compared to 1961 to 199014. However, whether dryland expansion will have negative impacts on vegetation productivity has been argued and still debated17,18. The AI changes in the past are in apparent contradiction to the vegetation trend which is a greening trend in most dryland areas according to satellite records19,20,21,22. Based on satellite derived vegetation index, 41% of drylands became greener between 1982 and 201523. CO2 was the main contributor of this change23. Although on average drylands are greening, 6% of drylands have undergone desertification between 1982 to 2015, and negative land use effects played a key role in 79.9% of areas, followed by climate change and climate variability23. Understanding whether, or to what extent, drylands will degrade under future climate change is critical to planning and managing them in advance.

Compared to earlier climate projections CMIP6 models boast finer spatial resolutions and employ more robust parameterization schemes for major physical processes, enhancing their predictive capabilities24. The historical forcings in CMIP6 are primarily based on observations24. For the future scenarios, near-term climate predictions were improved in CMIP624. Spatial patterns for the interannual variability of precipitation and temperature in drylands were well reproduced in CMIP6 models25.

Given the past inconsistency between AI trends and vegetation trends it is unclear whether future dryland vegetation changes will reflect these projected changes in atmospheric aridity. Here we use a data driven approach, TSS-RESTREND (Time Series Segmentation and Residual Trend), which considered the CO2 fertilization effects, under four CMIP6 future climate scenarios to examine the projected changes in dryland vegetation due to future anthropogenic climate change and compare these to the projected changes in atmospheric aridity represented by the AI in this study.

Results and discussion

Global drylands mean annual maximum Normalized Difference Vegetation Index (NDVImax) continues to rise over the coming three decades

Within dryland areas, NDVImax was observed to increase since 1982 (Fig. 1a). According to 24 CMIP6 models, as presented in Supplementary Table 1, the simulation results among models exhibit a consistent trend under each scenario. Under four future climate scenarios (SSP1-2.6, SSP2-4.5, SSP3-7.0, SSP5-8.5), we found that the global mean NDVImax in drylands is projected to continue rising over the next 27 years (2024 to 2050) (Fig. 1a). Notable, differences among the four scenarios emerge after the year 2030. With the SSP1-2.6 scenario showing a lower growth trend compared to other scenarios. While the SSP5-8.5 scenario shows a stronger upward trend. By the year 2050, there could be a difference of 0.027 in NDVImax (~6% of current values) between the high emission and low emission scenario.

Fig. 1: Global dryland mean NDVImax and AI trends of models under different scenarios.
figure 1

a Global dryland mean NDVImax trends from the past to the different future scenarios. b Global dryland mean AI trends from the past to the different future scenarios. Each light-colored line represents the simulation result from an individual model (same color represents same scenario: light green (historical), light blue (SSP-1.26), light pink (SSP-2.45), light red (SSP-5.85)), the dark-colored lines (observation, his_mean, 126_mean, 245_mean, 370_mean and 585_mean) represent global dryland mean of observed data, multi-model mean drive from 24 models of historical scenario and different future scenarios respectively.

Full size image

To gain further insights into the rising trend of NDVImax, we examined the global mean trends of precipitation and temperature in drylands, and CO2 levels (Supplementary Fig. 1). The rising trends in NDVImax across different scenarios and the variations between scenarios align with the corresponding changes in CO2 concentration (Supplementary Fig. 1d). Additionally, temperature displays an increasing trend and noticeable differences between scenarios emerge around the year 2030 (Supplementary Fig. 1b). In contrast, precipitation shows no clear trend in drylands during the period from 2015 to 2050 (Supplementary Fig. 1a). The precipitation discrepancy in land area among scenarios can be seen after the year 205026, beyond our study period. If dryland experience a similar situation, differences in NDVImax after 2050 could be even more pronounced. However, our future simulations did not consider the impact of breakpoints in the Normalized Difference Vegetation Index (NDVI) time series. Significantly changing climate factors may trigger vegetation changes, leading to breakpoints in linear relationships and inaccurate projections. To better simulate NDVImax in the far future, understanding the climate conditions that cause breakpoints requires further exploration.

The hotspots of NDVImax changes in drylands

We conducted a comparative analysis between the time periods of 2031 to 2050 and 1982 to 2001 to assess the changes under different climate scenarios (refer to Supplementary Fig. 2 for detailed results). The results of these scenarios revealed similar changes in vegetation dynamics. Greening hotspots were found in western India, southeastern Australia, northern East Asia, southeastern Africa and southern Sahel (Fig. 2a). Although the changes in dryland were broadly positive according to all scenarios simulation results, there are still some regions that exhibited browning, indicating degradation. After testing for statistical significance and agreement amongst models (see methods) we find that 3.61%, 3.44%, 3.00% and 3.31% of dryland areas were projected to experience significant degradation under SSP1-2.6, SSP2-4.5, SSP3-7.0, and SSP5-8.5 respectively. Our findings projected areas in the north-east Brazil and central Asia are the primary areas facing significant degradation and hence desertify (Fig. 2a) in the future. These future changes reflect recent trends in these regions27,28. Figure 2b provides insights into the model contributions to degradation, it is easy to see most models project degradation in the North-east Brazil, Namibia, Horn of Africa, Central Asia and north of East Asia.

Fig. 2: NDVImax changes and model contribution under scenario SSP5-8.5 between period 2031–2050 and 1982–2001.
figure 2

a NDVImax changes. Pixels where at least 50% of models indicate significant changes (p < 0.05), and 75% of models agree on the direction of change are plotted as green to brown. Pixels where 50% of models show significant changes, but less than 75% of models agree on the direction of change are plotted as white. Nonsignificant positive changes are plotted as cyan, nonsignificant negative changes are plotted as violet. b Percentage of models simulating negative changes. All non-dryland areas and areas that failed to build linear relationship (e.g., no rainfall or lack NDVI values in the time series) are masked as gray.

Full size image

To investigate the consistency of projected degradation across models and quantify the magnitude of disparities among them, we conducted an analysis to calculate the percentage of pixels showing significant negative change for each model. As depicted in Supplementary Fig. 3, the outcomes of four future scenarios across models revealed small disparities, especially scenario SSP3-7.0 which has the smallest variability. SSP1-2.6 exhibited the widest range among the models. In the SSP2-4.5 scenario, the median line is conspicuously situated towards the lower end of the box, indicative of a prevalent trend among models to project a lower degree of degradation compared to the median value. However, across all scenarios, the median values, and majority of models, consistently remained below 4%.

Vegetation productivity is strongly related to the availability of water resources29. Our analysis reveals 99% pixels exhibit positive responses to precipitation (Supplementary Fig. 4a). Regions such as western Indian and southern Sahel stand out as greening hotspots, primarily boosted by heightened precipitation levels (Supplementary Fig. 5a). Moreover, a substantial 67% of the studied pixels also demonstrate sensitivity to temperature fluctuations. 27% display positive reactions to temperature, particularly discernible in some areas in southeastern Africa and Atlas Mountain in northern Africa. NDVImax changes underscores the influence of temperature (Supplementary Fig. 4b), a finding that aligns with prior investigations conducted in semi-arid areas by satellite-derived observations30. Discernible degradation areas, such as northeastern Brazil and northwestern Namibia are locations with increased precipitation sensitivity31, where diminishing precipitation and escalating temperature manifest adverse impacts on vegetation. Increased precipitation is the main dominant factor driving greenness variation in dryland in the past decades32. As future climate changes, the relationship of vegetation with precipitation and temperature could be modified by elevated CO2 and increasing precipitation variability29,33,34,35. When considering the relation on a local scale, vegetation type can be important29.

The degradation assessment was calculated by contrasting the initial two decades of historical period with the last two decades of the future period. The results may exhibit variability contingent upon the periods chosen given that the analysis of long term vegetation changes is sensitive to the environmental conditions over the study period30,36. Nonetheless, the projected changes were driven by climate conditions. Fensholt et al.30 have found NDVI trends in semi-arid areas which are precipitation constrained areas, are sufficiently coherent and consistent with correlated climate variables.

NDVI changes are opposite to AI changes

Previous studies have shown that the AI trend contradicts the dryland NDVI changes in the past11,23. Our findings corroborate this result, as depicted in Fig. 1. Specifically, our results indicate that NDVImax will continue to increase in the future, further contrasting with the trend simulated in the AI. Notably, the disparity in AI trends between higher emission scenarios and low emission scenarios is not as pronounced as that shown in NDVImax, particularly in the case of scenarios SSP3-7.0 and SSP5-8.5. Broadly under drier atmospheric conditions, our study reveals a widespread tendency for dryland areas to exhibit greener land cover (Fig. 2a). This trend is particularly noticeable in southeastern Australia, southeastern Africa, eastern East Asia, and some midwestern regions in the south of Canada. AI changes suggests these same areas will experience increased aridity (Fig. 3). However, there are still regions where changes in both NDVImax and AI align. For instance, in the eastern corner of Brazil and southwestern of Namibia, both AI and NDVImax indicate a negative change in the future. Similarly, for western India, both indices project a positive change in the future (Results of all scenarios see Supplementary Fig. 2a and Supplementary Fig. 5d).

Fig. 3: AI changes under scenario SSP5-8.5 between period 2031–2050 and 1982–2001.
figure 3

An increase in AI indicating wetter conditions (plotted in pink), a decrease in AI indicating drier conditions (plotted in blue). All non-dryland and failed to build linear relationship areas (e.g., no rainfall or lack NDVI values in the time series) are masked as gray.

Full size image

It is noteworthy that many studies primarily evaluate dryland with AI, often calculated via the Penman–Monteith method, which suggests an expansion of dryland area under future scenarios14,16. However, these methods, including Penman–Monteith and Thornthwaite, commonly utilized for dryland evaluation do not account for the potential benefits to vegetation from elevated CO2 levels having a water saving effect37. Some studies have examined this showing that Penman–Monteith overestimated dryland expansion38 and was not suitable to estimate drought in the future climate39. Research also indicated that increased leaf area (LAI), a response to elevated CO2 concentration, offsets around 30% of global average water-saving effects, as measured by physically based evapotranspiration model compared to conventional methods40. Elevated CO2 concentration enhances photosynthesis rates and decrease water demand for most vegetation. Inconsistencies in trends between increased LAI and decreased AI, attributed to CO2 fertilization effects, have been observed in drylands from 1982 to 201441. Our future projections show a broadly greening in dryland areas despite increasing aridity evaluated by Penman–Monteith and Thornthwaite methods in future climates. This highlights the importance of considering CO2 effects in future assessments to provide more accurate projections of dryland dynamics.

In comparison to global changes, climate variables in drylands exhibit relatively minor fluctuations

The trends and changes in precipitation, temperature and PET were calculated individually during the same period (Results of all scenarios see Supplementary Fig. 1 and Supplementary Fig. 5). Although the global mean precipitation shows a mild rising trend42, this trend is more pronounced than in dryland areas (Supplementary Fig. 1a). Temperature will increase under all scenarios across global drylands (Fig. 4b). And the multi-model mean temperature for drylands demonstrates that the difference between high and low emission scenarios may reach 0.8 °C by the end of the year 2050 (supplementary Fig. 1b). While for global land regions, the difference in the warming trend can amount to 0.4 °C per decade43. Certain regions, including the north of North Asia, middle of Africa, and west central North America will experience a discernible rise in temperature, especially under scenario SSP5-8.5 (Fig. 4b, Supplementary Fig. 5b). Additionally, PET demonstrates a general increase across most dryland regions, except for certain areas in the west of East Asia (Fig. 4c). Notably, the time series analysis of dryland mean PET indicates a rising trend, with the trend being more pronounced under high emission scenarios (Supplementary Fig. 1c).

Fig. 4: Climate variables changes under scenario SSP5-8.5 between period 2031–2050 and 1982–2001.
figure 4

a Monthly total precipitation changes (blue represents increase, brown represents decrease). b Monthly mean temperature changes (red represents increase). c Monthly total PET changes (red represents increase, white represent decrease). All non-dryland and failed to build linear relationship areas (e.g., no rainfall or lack NDVI values in the time series) are masked as gray.

Full size image

In summary, projection of vegetation productivity within drylands is very important and constitutes a pivotal phase in addressing forthcoming challenges posed by climate changes, and offers a strategic avenue for formulating adaptive strategies within drylands. Our findings reveal a persistent contradiction between vegetation changes and Aridity Index changes across all future climate scenarios. Globally, a general greening on drylands in future climate scenarios was simulated, less than 4% (~3.41 million km2) of drylands are projected to desertify. SSP3-7.0 projected the least degradation among these four scenarios. Further, our results showed that North-east Brazil, Namibia, western Sahel, Horn of Africa and central Asia are projected to degrade by most models in all future climate scenarios. This indicates that there is no “safe” level of climate change for these locations. Overall, our findings indicate persistence of elevated vegetation productivity in most drylands, driven by the fertilizing effects of increased CO2 concentrations within a broadly desiccated atmospheric context. This projection serves as an essential groundwork for making mitigation strategies, especially within high-risk areas.

Methods

Observation datasets

TERRACLIMATE data offering monthly high-spatial resolution (0.04°) coverage for global terrestrial areas, spanning from 1958 to 201544. It has higher spatial resolution compared with other observation datasets, such as CRU4.07 (0.5°) and ERA5 (0.25°)44. TERRACLIMATE also provides consistent datasets covering a longer time span compared to ERA5. TERRACLIMATE precipitation and temperature data were used as input climate variables for running TSSRESTREND method, potential evapotranspiration (PET) and precipitation datasets were used to calculate AI which was used to generate a dryland mask (AI < 0.65)23. To ensure the compatibility of the study period, we selected the period from 1960 to 2014, as climate variable data should encompass a longer time frame compared to the NDVI dataset23. The NDVI dataset employed in this study is GIMMSv3.1 g, which represents version 3.1 of Global Inventory for Mapping and Modeling Studies (GIMMS). GIMMSv3.1 g provides NDVI data derived from the Advanced Very High-Resolution Radiometer (AVHRR) aboard the National Oceanographic and Atmospheric Administration (NOAA) satellites, spanning the period from 1982 to 2015. GIMMS dataset is widely recognized as one of the most commonly used long-term vegetation index products45 and it demonstrates good concordance with higher resolution NDVI datasets such as MODIS46. The resolution of GIMMSv3.1 g dataset is 15 days at 1/12°, in this study, it was aggregated to monthly data between 1982–2014 using the mean of the valid values47. Then the NDVImax in a calendar year is utilized as a proxy for vegetation growth23. Which serves as a well-established indicator of vegetation productivity48. The fraction coverage of C3 and C4 vegetation was obtained from the North American Carbon Program (NACP) Global C3 and C4 SYNergetic land cover MAP (SYNMAP)23,49. Which are used in TSS-RESTREND to account for CO2 effects. Subsequently, all data were resampled to match the spatial resolution of the NDVI dataset using the First-Order Conservative Remapping method50 implemented in the Climate Data Operators (CDO) software, facilitating consistent analyses and interpretations. And the non-dryland regions in all datasets are masked using the generated dryland mask in the CDO software.

CMIP6 data

37 CMIP6 precipitation and near-surface air temperature were used before the model selection, each with its unique spatial resolution. To ensure consistency, all datasets were interpolated to a uniform spatial resolution of 1/12°, matching the resolution of NDVI. The First-Order Conservative Remapping method50 implemented in the CDO software was employed for the resample purpose. For the historical scenario analysis, data spanning the period from 1982 to 2014 were selected, aligning with the observational data to maintain consistency. Looking towards future scenarios, this study focused on the projection period from 2015 to 2050.

Calculating the aridity index

The AI, defined by the United Nations Convention to Combat Desertification (UNCCD), is widely used to evaluate dryland conditions. In this study, drylands are identified based on an AI threshold of 0.65. The estimation of potential evapotranspiration (PET), a crucial variable for AI calculation, is commonly accomplished using Thornthwaite equation or FAO Penman–Monteith method51. However, limitation in variable availability with some CMIP6 models render it impractical to obtain PET using Penman–Monteith method for each model, we selected Thornthwaite equation to calculate PET. Thornthwaite has lower data requirements compared to other methods, it can be applied to all GCMs to simulate PET at a large-scale and assess aridity over different landscapes52. It was found to have the lowest uncertainty among 11 different methods53. It exhibits broadly similar features in the interdecadal variability of most dryland areas and has high correlation with Penman–Monteith results51. The formulars used to calculate PET and AI are shown as followings51:

$${PET}=16\times \frac{L}{12}\times \frac{N}{30}\times {\left(\frac{10T}{I}\right)}^{a}(0\le T < 26.5)$$
(1.a)
$$I=\mathop{\sum }_{k=1}^{12}{\left(\frac{T}{5}\right)}^{1.514}$$
$${PET}=\frac{L}{12}\times \frac{N}{30}\times (-415.85+32.24T-0.43{T}^{2})(T\ge 26.5)$$
(1.b)
$${AI}=\frac{P}{{PET}}$$
(2)

PET: Potential evapotranspiration (mm month−1)

L: The average day length (hours) of the month being calculated

N: The number of days in the month being calculated

T: The average monthly temperature (°C; negatives were set to 0)

$$a:a=\left(6.75\times {10}^{-7}\right){I}^{3}-\left(7.71\times {10}^{-5}\right){I}^{2}+\left(1.792\times {10}^{-2}\right)I+0.49239$$

AI: Aridity index

P: Precipitation (mm month−1)

To enhance the reliability of our findings, we additionally calculated PET by FAO Penman–Monteith method for models with the required variables. The calculation formula, models that we used (Supplementary Table 1), and results (Supplementary Fig. 6) are shown in the supplementary materials.

Bias correction of datasets

Prior to CMIP6 application, it is necessary to address the differences in the bias54 through bias correction techniques. Considering the non-negative values associated with precipitation, it is important to account for relative changes. To address bias in precipitation data, a widely adopted method known as linear scaling was used54,55. The CMIP6 precipitation dataset was corrected using the following equation:

$$P=X\frac{{Y}_{{mean}}}{{X}_{{mean}}}$$
(3)

P: precipitation data after bias correction; \(X\): CMIP6 precipitation data; \({Y}_{{mean}}\): monthly mean of observed precipitation; \({X}_{{mean}}\): monthly mean of CMIP6 historical precipitation.

And temperature was corrected using this equation55:

$${{{{{\rm{T}}}}}}=\frac{X-{X}_{{mean}}}{{X}_{{std}}}\times {Y}_{{std}}+{Y}_{{mean}}$$
(4)

T: temperature date after the bias correction; X: CMIP6 temperature data; \({X}_{{mean}}\): mean of CMIP6 historical temperature; \({X}_{{std}}\): standard deviation of CMIP6 historical temperature; \({Y}_{{mean}}\): mean of observed temperature; \({Y}_{{std}}\): standard deviation of observed temperature.

Model selection

Climate models are our best tools to understand how climate has changed in the past and what may happen in the future. They provide useful information, but uncertainties from many aspects affect the output of these models. Different climate models show different discrepancies when comparing with observed datasets. Choosing the right models is a key step in climate change assessment and will influence the study results. Climate models should be selected based on the research purpose, case-specific metrics and intended use56.

In our study, precipitation and temperature are the primary climate variables of interest. To evaluate the performance of the precipitation and temperature datasets, we employ the Root Mean Square Error (RMSE) and Bias metrics. Then ranking models through RMSE and Bias scores respectively. The ranking results are presented in Supplementary Fig. 7. The formulas used to calculate Bias and RMSE are as follows:

$${Bias}=\frac{{\sum }_{1}^{12}({S}_{M}-{O}_{M})}{12}$$
(5)
$${RMSE}=\sqrt{\frac{{\sum }_{1}^{12}{({O}_{M}-{S}_{M})}^{2}}{12}}$$
(6)

\({S}_{M}\): Monthly mean of CMIP6 data.

\({O}_{M}\): Monthly mean of observation data.

Ten models are excluded from further analysis based on these ranking results. Additionally, during the bias correction process for precipitation, it was observed that extremely low and zero values in the historical scenario, combined with higher values in future scenarios, could result in substantial values after bias correction. And this situation is particularly likely to occur in hyper-arid areas. Considering statistical significance, to mitigate this issue, we apply a masking technique, removing areas where precipitation is less than 1 mm in historical time series. The models are then ranked based on the remaining number of pixels. Both all dryland areas and dryland areas excluding hyper-arid regions are considered. The results reveal that models “MPI-ESM1-2-LR”, “MPI-ESM1-2-HR” and “AWI-CM-1-1-MR” exhibit deficient performance and are thus removed from further study (Supplementary Fig. 8). Consequently, a total of 24 models remains for use in this study, as presented in Supplementary Table 1.

Masks for each model and common mask

Considering the inherent limitations of bias correction, the models are subject to varying degrees of masking. The number of model contributions is depicted in Supplementary Fig. 9. Regions with dark blue have a smaller number of models to contribute to the following results analysis. To facilitate a comparison of the time series development, a common mask was employed, where regions with a model contribution value lower than 24 were excluded from the analysis. However, when investigating spatial changes, each model was assigned its own unique mask, and subsequently, the spatial changes were averaged based on the number of models employed. This approach allows for a comprehensive examination of both temporal and spatial dynamics while accounting for the specific characteristics and contributions of individual models.

Applying TSS-RESTREND

TSS-RESTREND method47,57 was used in this study to derive the linear relationship coefficients and other essential parameters for each grid. TSS-RESTREND method combines RESTREND method58 and BFAST method59,60, offering an extended approach to address the limitation of RESTREND in the presence of structural changes within dryland ecosystem47. This means it can successfully capture the long-term trend and step change of NDVI in regions where the ecosystem has undergone significant structural changes23. Once step changes (breakpoints) are detected, TSS-RESTREND method rebuilds the relation after the breakpoint. Moreover, to ensure robustness, the selection of data was tested and a single dataset was found capable of detecting changes in over 95% of areas61. Furthermore, multiple ensembles were utilized to test the application of TSS-RESTREND, with the total change captured by TSS-RESTREND exhibiting a deviation of only 3% from observed change4. In this study, we focused on the relationship between NDVImax and accumulated climate variables. We used observation precipitation, temperature and NDVI as the input data to TSS-RESTREND. TSS-RESTREND will find the NDVImax and remove the influence of CO2. Then build linear relationship between NDVImax and accumulated climate variables for each pixel. After that, the coefficients of each linear relationship and the best combination of parameters used to calculate accumulated climate variable were outputted by TSS-RESTREND model.

Simulate NDVImax

To simulate NDVImax, we employed a reversed version of TSS-RESTREND method. The following steps were undertaken to achieve this simulation: 1. Selection of peak months: from the time series spanning the period 1982 to 2014, we identified the months in which NDVImax occurred most frequently for each pixel. These peak months served as the basis for subsequent calculations. 2. Calculation of optimal accumulated precipitation and temperature: using the identified peak month and accumulate and offset periods output from TSS-RESTREND57 to calculate optimal accumulated precipitation and temperature. 3. Simulate NDVImax: utilizing linear relationship coefficients derived from the TSS-RESTREND, and accumulated precipitation and temperature to get simulated NDVImax. 4. Incorporation of CO2 effects: To account for the influence of CO2 on NDVImax, we applied the necessary adjustments to the simulated values. A theoretical relationship between increasing CO2 and photosynthesis was used to remove the CO2 contribution on NDVI23,62. The figure provided in supplementary Fig. 10 illustrates the overall processing workflow for the simulation of NDVImax in our study.

Test statistical significance

To evaluate the significance of the NDVI changes, we performed a t-test for each pixel across all models under different scenarios. Pixels were identified as significant changes if at least 50% of the models got p-values smaller than 0.05. Subsequently, an agreement test (more than 75% models agree on the direction of changes) was employed to assess the consensus among the areas with significant changes.