We use output from 12 CMIP6 ESMs for both the historical experiment and the SSP3-7.0 scenario (; Table S1). SSP3-7.0 represents a regional rivalry pathway (Shared Socioeconomic Pathway 3) with a radiative forcing of 7 Wm-2 by 2100 (Representative Concentration Pathway 7.0; ). We select SSP3-7.0 because the warmest scenario, SSP5-8.5, is unlikely to be realised under current and projected emission trajectories . SSP3-7.0, as the second-highest emission scenario, is therefore particularly relevant for current climate impact assessments. However, we acknowledge that hydroecological responses are scenario-dependent . Given the impact focus of our analysis, we prioritise a scenario whose coupled forcings are more directly policy-relevant outcomes than idealised experiments.

The ESMs are selected to provide the variables required to compute the BGWS metric (Sect. ) and the climatic and hydroecological variables used in the subsequent analyses (Sect. ). Additionally, only models with time-varying leaf area index (LAI) were considered (Table S1). For each model, we select the ensemble member with the lowest available indices along the CMIP6 ensemble axes (“ripf”: realisation, initialisation, physics and forcing; member IDs in Table S1). This approach ensures that each model contributes equally to the analysis and avoids over-representing models with a larger number of ensemble members.

To evaluate the performance of the selected CMIP6 models in representing the BGWS, we compare the ESM results against observation-based and reanalysis datasets. Specifically, we compute the BGWS from GPCC precipitation , G-RUN runoff , and GLEAM transpiration , providing an observation-based reference for model evaluation. These datasets are derived using different methodologies: GPCC is a gauge-based precipitation dataset, G-RUN reconstructs global runoff using a machine-learning approach trained on streamflow observations, and GLEAM is a hybrid observation–model product that estimates land-surface evaporation and transpiration from satellite-based data. Thus, these reference datasets are not equally direct observations and, due to their independent origins, are not fully consistent with one another or with a closed global water balance . To account for this, we also use ERA5-Land, a physically consistent reanalysis dataset providing all required variables from a single modelling framework . However, reanalysis data are not direct observations and inherit uncertainties from model physics, model parametrisations and input data quality. While ERA5-Land offers a spatially and temporally coherent dataset, its ability to accurately represent long-term hydrological changes remains uncertain . We therefore use these global-scale gridded datasets as a first-order consistency check of ESM simulated BGWS patterns, rather than as a full benchmarking of ESM skill. In addition to comparing climatological BGWS fields, we also assess whether the ensemble mean reproduces observed changes in BGWS over the recent historical period. For this purpose, we compare BGWS changes between 1985–1999 and 2000–2014 in the CMIP6 ensemble mean and the reference datasets.

To enable spatial comparison across models and datasets, we regrid all fields to a common 1°×1° latitude–longitude grid using conservative interpolation. After regridding, we compute 30-year climatological means for a historical period (1985–2014) and a far-future period (2071–2100). Future-minus-historical changes are then calculated for all variables used in the subsequent analyses.

We apply a fixed historical mask derived from the ensemble mean over 1985–2014. The Greenland/Iceland and Antarctic regions are excluded a priori using the reference regions of the Sixth Assessment Report of the Intergovernmental Panel on Climate Change (IPCC AR6; ), because they are dominated by permanent ice cover and fall outside the vegetated runoff–transpiration domain targeted here. The mask restricts the remaining analysis to vegetated and hydrologically active land areas where the subsequent runoff–transpiration partitioning analysis is meaningful. Grid cells are retained only where mean precipitation exceeds 0.822 mmd-1 (≈ 300 mmyr-1) and where both mean runoff and mean transpiration exceed 0.05 mmd-1. These values are conservative screening thresholds rather than hydroclimatic regime definitions. They exclude grid cells with low water input or negligible runoff or transpiration, where a runoff–transpiration partitioning interpretation is less robust.

The precipitation threshold follows the order of magnitude used in the IPCC description of arid zones , and the runoff threshold matches the value used to mask very low-runoff grid cells in a comparable large-scale impact assessment . We apply the same low-flux threshold to transpiration because both fluxes enter the BGWS metric directly and symmetrically. Because non-negligible transpiration requires active vegetation, these criteria also exclude bare and sparsely vegetated land. The same mask is applied consistently to all historical fields, future changes, and reference datasets. It is also applied identically to all 12 ESMs, so that ensemble statistics and individual-model attributions are evaluated on the same spatial domain.

We employ an unweighted multi-model ensemble (hereafter referred to as the ensemble mean) to mitigate uncertainties associated with individual model biases and initial-condition variability. For each model, climatological means and changes are first computed at the model level and then averaged across models to obtain the ensemble mean. This ordering ensures that ensemble statistics reflect individual model responses before averaging. We do not apply performance-based weighting, as our objective is to assess broad climate impacts rather than to optimise the ensemble for a specific variable. Spatial means are always calculated as area-weighted averages to account for the convergence of meridians with latitude. Each grid cell is weighted by the cosine of its latitude, cos⁡(φ), where φ denotes latitude.

In this study, we analyse blue–green water partitioning using precipitation (P; mmd-1) as the incoming water supply, runoff (R; mmd-1) as blue water flow and transpiration (Et; mmd-1) as the vegetation-mediated green water flow in vegetated and hydrologically active land areas. We focus on transpiration because it represents the direct vegetation water-use response, whereas non-transpiration evaporation may respond differently to climate change than vegetation-mediated water use . This distinction matters because total ET can reflect offsetting changes in its components. For example, transpiration can decrease while soil evaporation increases, so the ET response emerges from opposing biological and physical processes . Using ET instead of transpiration would therefore address a different question, namely runoff versus total evaporative loss rather than runoff versus vegetation-mediated water use .

We introduce BGWS as a flow-based partitioning metric for runoff and transpiration rather than as a complete representation of all blue and green water or water-balance components. BGWS thus does not represent total green water flow, but specifically the vegetation-mediated component of green water use captured by transpiration. The historical mask defined in Sect.  ensures that the metric is evaluated only in regions where this transpiration-based interpretation is meaningful.

We define the BGWS metric as: 1BGWS=R-EtP×100, which is dimensionless and expressed in percent. Positive BGWS values indicate that a larger share of precipitation is partitioned towards runoff, whereas negative BGWS values indicate that a larger share is partitioned towards transpiration.

BGWS is therefore complementary to existing partitioning metrics. Relative to the blue water trade-off (BWT) metric used by , BGWS contrasts runoff directly with transpiration, does not include interception or storage terms, and reports the result as a normalised share of precipitation. Thus, interception-driven canopy effects are not represented explicitly by BGWS. For assessing vegetation-mediated green water use, however, transpiration as a direct ESM output is useful because it provides an immediate indicator of plant water use.

We focus on water fluxes rather than water stores because hydrological fluxes are more consistently available and more directly comparable across the selected CMIP6 ESMs. Moreover, on multi-decadal timescales (here, 30-year means), storage changes tend to be small relative to fluxes, as many catchments approach a near-steady state within this period . Neglecting storage terms therefore provides a pragmatic and interpretable framework for large-scale analysis, although this simplification remains a limitation and an avenue for future research.

Changes in BGWS, denoted as ΔBGWS, indicate shifts in blue–green water partitioning and are calculated as 2ΔBGWS=BGWSFuture-BGWSHistorical.

A positive ΔBGWS reflects an increasing partitioning towards blue water flow, whereas a negative ΔBGWS indicates an increasing partitioning towards green water flow. Although ΔBGWS does not directly measure absolute changes in all blue and green water components, it provides a clear measure of the changing trade-off between runoff and transpiration.

To complement the transpiration-based analysis, we also compute non-transpiration evaporation Ent as the difference between ET and transpiration Et, 3Ent=ET-Et.

This quantity represents non-transpiration evaporation, including soil evaporation, interception evaporation, and other non-transpiration components as represented in each dataset. We do not use CMIP6 component evaporation variables directly because key components (e.g. evspsblsoi) are unavailable for many models.

We analyse both the historical 30-year climatological means and the future-minus-historical changes in selected climatic and hydroecological variables associated with the distribution and shifts in blue–green water partitioning. The selected variables are mean precipitation (P; mmd-1), precipitation seasonality (Pseas; mmd-1), annual maximum consecutive five-day precipitation (RX5day; mm), surface soil moisture (SM_surf; mm), leaf area index (LAI; m2m-2), transpiration-based water-use efficiency (WUE; gCm-2mm-1), mean vapour pressure deficit (VPD; hPa), VPD seasonality (VPD_seas; hPa), and total cloud cover (CLT; %). Except for RX5day, which is derived from daily precipitation, all variables are based on monthly mean output. The variable selection is based on the predictor screening described in Sect. . Seasonality, RX5day, VPD, and WUE are computed as follows.

Precipitation and VPD seasonality are defined as the standard deviation of the 12 monthly climatological means within each 30-year period. For a given variable X and monthly climatology X‾m, seasonality is computed as 4Xseas=112∑m=112X‾m-X‾2, where X‾m is the climatological mean for month m and X‾ is the mean of the 12 monthly climatological values.

RX5day is derived from daily precipitation as the annual maximum consecutive five-day precipitation. For a given year j, 5RX5dayj=max⁡(RRkj), where RRkj denotes total precipitation accumulated over any five-day interval ending on day k within year j. For each 30-year period, RX5day is averaged over the annual maxima.

VPD is calculated using the Buck equation to estimate saturation vapour pressure (es), 6es=611.21×exp⁡18.678-T234.5T257.14+T, where T is near-surface air temperature in °C, and 611.21 Pa is the saturation vapour pressure at 0°C. Actual vapour pressure (ea) is obtained from specific humidity (q) and surface pressure (ps, in Pa) as 7ea=qps0.622+0.378q, and VPD is then given by 8VPD=es-ea.

VPD is converted from Pa to hPa for the analysis.

We use the ratio of gross primary productivity (GPP) to transpiration to compute WUE, a widely used proxy for the CO₂ effect on stomatal resistance : 9WUE=GPPEt, which quantifies the efficiency of carbon assimilation per unit of water consumed by vegetation through transpiration.

We construct multiple linear regression (MLR) models with Elastic Net regularisation separately for the two historical BGWS regimes to explain the spatial pattern of projected ΔBGWS across grid cells. Each grid cell is treated as one sample, and both the predictors and the response variable are based on ensemble mean changes between the future and historical periods. The analysis therefore explains spatial differences in projected ΔBGWS within each regime. It does not attribute the temporal evolution of BGWS at individual grid cells or regions.

Because neighbouring grid cells are not independent, we apply spatially blocked cross-validation. Grid cells are grouped into 10°×20° latitude–longitude blocks, and all model training, tuning, and testing are performed using these spatial blocks as groups. For each regime, we apply repeated grouped train–test splits with 20 % of the spatial blocks held out per repeat, combined with nested grouped cross-validation for Elastic Net hyperparameter tuning. To favour simpler models and reduce overfitting, the final hyperparameters are selected using a one-standard-error rule. Model performance is evaluated using out-of-sample R2 on held-out spatial blocks.

To address multicollinearity among predictors and to regularise the regression, we apply Elastic Net regularisation . The Elastic Net objective function is 10min⁡β12n∑i=1n(yi-y^i)2+αρ∑j=1p|βj|+α(1-ρ)2∑j=1pβj2, where yi and y^i are the observed and predicted values, n is the number of samples, βj are the regression coefficients, α controls the overall regularisation strength, and ρ determines the balance between L1 (Lasso) and L2 (Ridge) penalties. The L1 penalty promotes sparsity by shrinking some coefficients to exactly zero, whereas the L2 penalty shrinks coefficients towards zero without eliminating them, helping to stabilise estimates under multicollinearity. Prior to model fitting, all predictors are standardised to zero mean and unit variance within each training split, while ΔBGWS is left on its original scale.

The final predictor set was selected from a larger initial set of candidate predictors according to their contribution to predictive performance, their physical interpretability, and data availability across the full 12-model ensemble (Sect. S2). The final set includes changes in mean precipitation, precipitation seasonality, RX5day, mean VPD, VPD seasonality, total cloud cover, near-surface soil moisture, LAI, and WUE. Together, these variables represent climatic and hydroecological processes relevant to climate change impacts and allow physical interpretation of their influence on ΔBGWS. Changes in mean precipitation, precipitation seasonality, and extreme precipitation represent water supply and hydrological extremes. RX5day is used as the main extreme precipitation metric because multi-day precipitation accumulation shows slightly better spatial agreement than single-day extremes in CMIP6 and observational evaluations, making it a more robust indicator of extreme precipitation in large-scale datasets . Changes in mean vapour pressure deficit and VPD seasonality represent atmospheric dryness and plant water stress. Changes in near-surface soil moisture represent surface water availability, while changes in LAI and WUE represent vegetation state and plant physiological responses to elevated CO₂. Changes in LAI reflect both climatic and human influences, including land-use and land-cover change, although these contributions are not isolated explicitly here. Changes in total cloud cover represent an additional atmospheric control related to energy and moisture conditions.

To quantify the contribution of each predictor to model performance, we compute permutation importance scores . This model-agnostic metric is obtained by randomly permuting the values of a given predictor and quantifying the resulting change in R2. For each repeat and each predictor, predictor values are permuted 20 times and the resulting decrease in R2 is recorded. We interpret these test-set importance scores as measures of the contribution of each predictor to the spatial pattern of projected ΔBGWS within each regime. For comparison with the ensemble-mean attribution, we repeat the same blocked regression workflow for each individual ESM and report individual-model permutation importances only where mean held-out R2>0.3.

The historical BGWS pattern in the CMIP6 ensemble mean (1985–2014; Fig. a) reveals three overarching features: energy-limited, water-limited, and precipitation-seasonality regulated blue–green water partitioning. The first two features are consistent with the broader literature distinguishing energy- and water-limited land-surface or ecosystem regimes, while the third extends this logic to precipitation-seasonality effects that are particularly relevant for blue–green water partitioning . On average, the area-weighted BGWS is slightly positive, indicating that ∼3% more precipitation is partitioned towards blue water flow, with notable ensemble variability from -13% to +17% (Table S4). These contrasts stem from differences in the model representation of precipitation characteristics, land-surface hydrology, and vegetation dynamics, which contribute to well-documented CMIP6 biases in simulated land-surface water fluxes relative to reference products .

Comparing the spatial pattern of the ensemble mean BGWS with observation-based and reanalysis products indicates a reasonable large-scale representation for the historical climatology (Fig. S2). Agreement is weaker, but still qualitatively similar, for recent historical changes over the shorter evaluation period (Fig. S3). Compared with both reference datasets, the ensemble mean BGWS values are biased towards too much blue water. Decomposing BGWS into runoff (R/P) and transpiration (Et/P) ratios shows that the bias is driven primarily by underestimated Et/P (negative mean offset), while regional differences in R/P also contribute (Figs. S4 and S5). One possible structural contribution to the underestimated Et/P bias is the limited representation of lateral groundwater redistribution in many large-scale land models used in ESMs, where this process is often overlooked or strongly simplified. By underrepresenting shallow groundwater support to ET in drylands or during dry periods, this may partly favour a more blue-biased BGWS . Accordingly, we treat BGWS as a process indicator and base interpretation primarily on directional patterns, sign-robust spatial structures, and covariation with relevant predictors, while avoiding detailed interpretation in regions with disagreement against both reference datasets (stippling in Fig. a).

Energy-limited blue–green water partitioning, the first BGWS feature, is dominant across higher latitudes (>60° N) and high-altitude regions, which consistently show greater partitioning towards blue water flow (Fig. a). This pattern aligns with the Budyko framework, which links energy-limited conditions to increased runoff . In the higher latitudes, 95 % of the land area exhibits a larger blue water share, with an area-weighted mean BGWS of +36%. Here, low net radiation (light-limited) and low air temperatures (kinetic/phenology-limited) coincide with below-global-average Et/P, while near-surface soil moisture remains comparatively high (Figs. c and S6). Consequently, precipitation predominantly contributes to runoff (Fig. b). Mountain ranges exhibit similar behaviour, where orographically enhanced precipitation and topography-controlled runoff parametrisations in ESMs (e.g., SIMTOP) can contribute to larger blue water shares . Snowmelt-driven runoff likely further enhances blue water shares in cold regions , although simplified representation of snow sublimation in ESMs may also contribute to runoff overestimation in some high-altitude regions and thus reinforce locally blue-biased partitioning . Ongoing warming-related snow cover loss, by contrast, increases net radiation and ET and can reduce runoff . Energy-limited partitioning is also evident across parts of the mid-latitudes (40–60° N/S), but with much smaller positive BGWS values than in the higher latitudes, with an area-weighted mean BGWS of about +6%. This is consistent with seasonal transitions between winter energy limitation and summer water limitation . Compared with observation- and reanalysis-based products, however, the ensemble mean tends to overstate both the magnitude and the spatial extent of blue water shares in energy-limited high- and mid-latitudes (Fig. S2).

Water-limited blue–green water partitioning, the second BGWS feature, dominates in semi-arid and dry sub-humid regions such as the Eurasian Steppe or Australia’s Outback (Fig. a). These water-limited environments are marked by annual precipitation below the global mean and relatively low near-surface soil moisture (Fig. S6). Despite these hydrological constraints, these ecosystems partition a larger fraction of available water towards transpiration, increasing the green water share in line with observation-based data (Figs.  and S2). The RX5day-to-annual-precipitation ratio is elevated across many of these dryland regions (Fig. S7), indicating that rainfall is often temporally concentrated even where mean precipitation is low. Nevertheless, within the retained vegetated drylands, the annual partitioning still tends to favour transpiration over runoff. This is consistent with water-limited ecosystems maintaining relatively low runoff fractions while using a large share of the limited incoming precipitation for vegetation water use .

Precipitation-seasonality regulated blue–green water partitioning, the third BGWS feature, emerges most clearly in monsoonal and other strongly seasonal climates highlighted by the inset in Fig. a. In these regions, BGWS depends not only on how much precipitation falls annually, but also on how strongly rainfall is concentrated within the year. Regions with high precipitation seasonality, such as southern and eastern Asia and parts of the Maritime Continent, tend to show larger blue water shares where wet-season rainfall is strongly concentrated, RX5day is high, and near-surface soil moisture is comparatively high (Figs.  and S6). By contrast, humid regions with more weakly seasonal rainfall, such as much of the Amazon and central Africa, tend to show larger green water shares and higher Et/P. Thus, the contrast is not simply between wet and dry regions, but between climates with strongly concentrated wet seasons and those with more even rainfall distribution. This mechanism is also consistent with the contrast between East Asian monsoon regions, which show larger blue water shares, and subtropical regions such as the Florida Peninsula, where more evenly distributed rainfall coincides with larger green water shares.

Overall, the historical BGWS distribution is therefore consistent with a transition from energy-limited blue-water dominance in cold and high-altitude regions, to water-limited green-water dominance in semi-arid regions, and to seasonality-regulated partitioning in monsoonal and humid subtropical climates. Several regions, especially within this third feature, also show disagreement with both reference datasets or elevated ensemble spread, underscoring the challenges ESMs still face in representing humid subtropical and tropical hydroecological dynamics (; Figs. a and S8).

Figure a maps the projected BGWS changes (2071–2100 minus 1985–2014) under the regional rivalry scenario (SSP3-7.0), separated by whether grid cells historically belong to the blue or green water regime. The four change classes occupy broadly comparable fractions of the analysed land area. In the historical green water regime, 21.5 % becomes greener and 32.8 % less green; in the historical blue water regime, 21.0 % becomes bluer and 24.7 % less blue. Despite this broad balance in areal coverage, the ensemble-mean global BGWS increases slightly by +1.35 ppts (Table S4), with individual-model global mean changes ranging from -1.86 to +5.1 ppts. Regime shifts are less extensive than the strengthening or weakening of existing regimes, and occur more often from green to blue than from blue to green (7.2 % versus 2.4 % of the analysed land area; Fig. S9). Because BGWS reflects the difference between runoff and transpiration relative to precipitation, a larger blue water share arises wherever runoff changes more positively than transpiration, whereas a larger green water share arises wherever transpiration changes more positively than runoff. The spatial expression of these mechanisms varies strongly across latitude bands and climate regimes (Fig. ).

Larger blue water shares most directly occur where runoff increases more than transpiration increases. This pattern is evident in wetter parts of the historical blue water regime, for example, in northeastern Eurasia and along the East Asian monsoon belt, where enhanced precipitation coincides with increasing runoff and a strengthening of blue-water dominance (Fig. ). In these regions, both blue and green water flows increase in absolute terms, but runoff gains are larger, so partitioning shifts further towards blue water. Similar positive BGWS changes are also visible in historically blue-water dominated tropical regions such as large parts of India, where future trends are generally positive and coincide with increased wetting (Fig. ).

Larger blue water shares can also occur where runoff increases while transpiration decreases. This pattern is especially relevant in parts of the subtropics and tropics, where wetter conditions and increasing runoff coincide with declining transpiration. The major rainforests exhibit an asymmetric response in this respect: only the Amazon rainforest tends towards a greater green water share, whereas parts of central Africa and Maritime Southeast Asia shift towards greater blue water shares. This divergence coincides with precipitation and runoff increases in the latter two regions, while transpiration declines across all three rainforest regions (Fig. ). Consistent with these regional examples, 69 % of the analysed subtropical land and 77 % of the analysed tropical land are projected to experience larger blue water shares.

A third pathway to larger blue water shares occurs where both runoff and transpiration decrease, but transpiration declines more strongly than runoff. This mechanism is visible in scattered historically water-limited patches of the subtropics and mid-latitudes, for example in parts of eastern Australia and northwestern Mexico (purple in Fig. a). In these regions, the relative decline in transpiration exceeds the runoff loss, shifting partitioning towards blue water even though absolute blue water flow may also decline. This contrasts with observed CO₂-driven blue water losses where greening increased ET and reduced streamflow in Australia's sub-humid/semi-arid basins , and with vegetation-driven runoff declines in the American West . While LAI increases across many of these regions, transpiration does not always rise in parallel because higher WUE can offset rising water demands (Figs. d and S10). More generally, whether dry regions will experience a blue or green water share increase depends largely on the interplay between rising vegetation water demand and WUE-related transpiration limitations, although structural uncertainties remain in how models represent vegetation responses to rising CO₂ and the resulting water-cycle impacts .

Larger green water shares occur where transpiration increases more than runoff increases. This mechanism is apparent in parts of the higher latitudes and mid-latitudes where leaf area expands and transpiration gains exceed the concurrent runoff increase, for example in northwestern Eurasia (Fig. S10). Consistent with this pattern, 64 % of the analysed land area north of 60° N experiences a decrease in BGWS. Although a larger share of precipitation is partitioned towards green water flow, these regions concurrently experience an increase in blue water flow (Fig. c). In such cases, the shift towards green water does not necessarily imply a drying climate, but rather that transpiration becomes relatively more important than runoff.

A larger green water share can also arise where transpiration increases while runoff decreases. This pattern is particularly evident in parts of eastern Europe, where greener partitioning coincides with runoff loss (Fig. ). In such cases, vegetation gains and enhanced transpiration shift the balance towards green water while blue water availability declines in absolute terms. The same logic applies to other mid-latitude regions where warming- and CO₂-driven vegetation responses amplify transpiration despite weak or negative runoff trends.

Finally, larger green water shares occur where both runoff and transpiration decrease, but runoff declines more strongly than transpiration. This mechanism is evident in parts of southern Europe, where a greater reduction in runoff compared to transpiration results in a green water share increase (Fig. ). This mechanism is also relevant in other regions with blue water losses, as receiving a smaller or only slightly larger share of precipitation does not offset the overall decline in runoff.

The regional BGWS changes described above suggest that future blue–green water partitioning reflects competing effects of changes in mean and extreme precipitation, changes in atmospheric and soil-moisture conditions, and vegetation responses. To quantify which selected predictors best explain the spatial pattern of ΔBGWS within each historical regime, we fitted separate blocked multiple linear regression (MLR) models for the historical blue and green water regimes (Sect. ). Figure  summarises the resulting permutation importance scores for the ensemble mean together with the corresponding importance ranks from individual ESMs. The ensemble-mean ranking should be interpreted as the common large-scale signal across the 12-model ensemble field, while the individual ESM points illustrate model-to-model variability in predictor importance. Both MLR models show moderate to strong predictive skill (R2=0.51 for the historical blue water regime and R2=0.74 for the historical green water regime). This interpretation and the importance rankings are further supported by the nonlinear Random Forest sensitivity analysis (Fig. S11).

Our results demonstrate that BGWS alterations in both regimes are most sensitive to extreme five-day precipitation changes (Fig. ). RX5day is projected to increase across most of the global land area (Fig. S10), consistent with theoretical expectations that warming-driven increases in atmospheric moisture intensify precipitation extremes . Its positive regression coefficient indicates that increases in RX5day are associated with larger blue water shares. This statistical relationship is physically plausible, because more intense multi-day rainfall can quickly saturate soils, triggering saturation-excess runoff in CMIP6 land-surface schemes. The RX5day effect would likely be even larger in models that also represent infiltration-excess runoff, which only a few schemes currently include . Additionally, current coarse-resolution ESMs tend to underestimate precipitation extremes and often exhibit drizzle bias, suggesting that the hydrological influence of RX5day may be underestimated in our analysis . The individual ESMs show that ΔRX5day is an important predictor across most of the ensemble, but not consistently one of the top three predictors. This indicates that its dominance is strongest for the ensemble-mean BGWS field, whereas individual-model importance rankings are more variable. Such smoothing of model-specific noise and internal variability is a known consequence of multi-model averaging and helps isolate the common large-scale forced response, although it can also mask structural differences among models .

Our findings emphasize two key points. First, even where mean precipitation decreases regionally, stronger increases in RX5day can still shift BGWS towards larger blue water shares. This suggests higher runoff sensitivity to intense rainfall even where average runoff is decreasing. Second, where mean precipitation also increases, larger RX5day changes can further enhance blue water shares because wetter mean conditions make saturation-driven runoff during extreme precipitation events more likely. This helps explain why ΔRX5day emerges as the dominant positive predictor of ΔBGWS despite contrasting regional trends in mean precipitation.

The dominant role of ΔRX5day is consistent with studies showing strong effects of rainfall extremes on runoff ratios and catchment retention . We extend this insight from retention and runoff-based metrics to a partitioning metric that also includes transpiration, linking precipitation intensity and plant water use to the balance between runoff and transpiration. Whereas found that runoff partitioning changes are governed by precipitation changes, including mean precipitation and five-day precipitation extremes, our results identify extreme five-day precipitation as the leading spatial predictor of projected BGWS change.

The strongest negative predictor of the ensemble mean ΔBGWS field in both regimes is ΔLAI, which is strongly supported by the individual ESM results (Fig. ). Larger increases in LAI are associated with greener partitioning, indicating that vegetation expansion tends to favour transpiration relative to runoff. This provides the main counterweight to the positive RX5day effect and helps explain why blue-to-green shifts still occupy a substantial fraction of global land area despite the dominant blueward influence of extreme precipitation. Global greening is evident in the ensemble mean, with larger LAI increases in the historical blue water regime (Fig. S10). Particularly blue water-dominated higher and mid-latitudes experience extensive vegetation growth due to global warming, where energy and temperature limitation is the primary controlling factor. In addition, CO₂ fertilisation and scenario-dependent land-use changes further impact LAI . Overall, larger LAI is associated with increased transpiration water demand and may also increase canopy interception losses, thereby reducing throughfall and runoff and potentially reinforcing green-ward shifts.

Beyond the two dominant predictors, the third-ranked predictor differs between regimes. In the blue water regime, ΔSM_surf ranks third, indicating that increases in near-surface soil moisture are associated with larger blue water shares. This is consistent with the blue water regime being concentrated in already wet, energy-limited, or seasonally wet regions, where higher near-surface soil moisture can increase the likelihood of saturation-driven runoff . In the green water regime, ΔVPD_seas ranks third in the full-domain analysis and becomes the leading predictor when the attribution is restricted to grid cells where at least 8 of the 12 models agree on the sign of projected ΔBGWS (Table S3). This high-confidence subset, however, covers only a spatially restricted part of the green water regime (Fig. S1) and should therefore not be interpreted as a replacement for the regime-wide attribution. Nevertheless, it indicates that seasonal atmospheric demand is an important additional control in regions where models robustly agree on the sign of change, consistent with the green water regime being concentrated in water-limited or strongly seasonal climates where atmospheric moisture demand constrains transpiration .

The ranking beyond the leading predictors should not be over-interpreted. Differences in ensemble-mean permutation importance are relatively small for several predictors, and the individual ESMs show considerable variability in both predictor magnitude and rank (Fig. ). This indicates that the ensemble-mean ranking reflects a common large-scale signal rather than a ranking that is reproduced identically by each individual model. The ranking also depends on ensemble composition and domain choice (Sect. S2). The leading predictors should therefore be understood as acting jointly rather than in isolation. In summary, projected BGWS changes arise primarily from the interplay between precipitation intensification and vegetation expansion, while plant–atmosphere controls on transpiration and changes in near-surface wetness emerge as additional regionally relevant predictors.

The BGWS metric is a process rather than a quantity indicator. It characterises how an incremental unit of precipitation (e.g., the next millimetre of rain) is partitioned between runoff (blue water) and plant use via transpiration (green water), not how large those fluxes are in absolute terms. In this sense, BGWS addresses the question “where does the next unit of rain tend to go?” rather than “how much water is there overall?”. As our analysis excludes hyper-arid and sparsely vegetated regions, this distinction matters for impacts that depend on hydrological sensitivity and timing . It shifts the focus from one on absolute change, to one in which we try to understand controlling factors and enables another level of model evaluation that centres on comparing these controlling factors rather than just focusing on matching historical observations . Although our analysis is flow-based and excludes storage terms, persistent BGWS shifts signal sustained pressure on blue water stores (e.g., rivers, reservoirs, groundwater) versus green water stores (root-zone moisture).

BGWS changes become most meaningful when interpreted jointly with absolute changes in runoff and transpiration rather than in isolation (Fig. ). In combination with runoff changes (Fig. a), the robust sign combinations point to four distinct hydrological implications. Where ΔBGWS and ΔR are both positive, both runoff volume and the runoff share of precipitation increase, indicating stronger blue-water partitioning and greater potential for high-flow pressure during wet periods (e.g., southern Asia); where ΔBGWS is positive but ΔR is negative, runoff shrinks in absolute terms but its share of precipitation still increases. Thus, the relative contribution of runoff pathways per unit precipitation increases, indicating higher runoff sensitivity to intense rainfall. Yet, this pattern occurs only rarely in the robust map. Conversely, negative ΔBGWS with positive ΔR suggests more total runoff but a smaller share of precipitation, implying that runoff may remain elevated if water supply stays sufficient (e.g., northwestern Eurasia); where both ΔBGWS and ΔR are negative, shrinking blue water volumes and shares indicate concurrent pressure on blue-water supply (e.g., large parts of Europe). Comparable patterns are seen during European droughts, where blue water declines outpace green water responses, tightening low-flows even when vegetation maintains transpiration .

Joint interpretation of BGWS with transpiration changes adds eco-physiological and land-surface insight (Fig. b). Where ΔBGWS and ΔEt are both positive, landscapes become more runoff-dominated while plant water use rises, consistent with higher atmospheric/phenological demand and greater reliance on stored water (e.g., northeastern Eurasia); where ΔBGWS is positive and ΔEt is negative, partitioning shifts toward runoff as vegetation down-regulates (e.g., southeastern South America). The latter response reduces evaporative cooling and elevates heat risk . Conversely, where ΔBGWS is negative and ΔEt is positive, regions experience greener partitioning, which may strengthen land–atmosphere coupling and support stronger latent cooling (e.g., northeastern North America); when both ΔBGWS and ΔEt decrease, relative allocation to green water rises despite lower plant water use, which points to weaker evaporative cooling and vegetation-mediated buffering of heat and moisture stress (e.g., parts of the Amazon rainforest).

In communicating results, it is important to make uncertainty explicit. Figure  therefore only shows areas where at least eight of the twelve models agree on the sign of both ΔBGWS and the corresponding flux change. Even so, the results should be read as robust regional tendencies rather than local predictions. In regions with robust sign agreement, joint interpretation of ΔBGWS with ΔR and ΔEt provides a compact way to distinguish whether future change is more likely to intensify runoff sensitivity, strengthen transpiration demand, or simultaneously reduce water availability and vegetation cooling. Given the additional baseline biases relative to reanalysis and observation-based products (Sect. S3), a regionalised and observationally constrained analysis of ΔBGWS and co-varying hydroecological responses would still be needed before drawing site-specific management conclusions.