Climate warming will alter runoff, generating hazards for people and ecosystems. Major uncertainties in the future of the hydrologic cycle under warming are thus one of the most important challenges in devising sound adaptation strategies (1–3). A key contributor to hydrologic uncertainty is how plants will respond to higher atmospheric CO2 concentrations (2, 4, 5) (CO2 physiological effects). More water-efficient carbon uptake by plants will likely alter leaf area and canopy conductance, which together affect water fluxes from the land to the atmosphere (evapotranspiration). Because those fluxes link the land and atmosphere, CO2 physiological effects may influence precipitation patterns (6–8) as well as the surface water balance and runoff (5).
A lineage of studies has suggested that reduced transpiration due to CO2 physiological effects may leave more water on the land surface and increase runoff (5, 9, 10–18). The expectation is that this plant-induced wetting will offset some of the drying caused by warmer air temperatures over land and modulate the impact of radiatively-driven precipitation change. Yet the mechanism of this transpiration-runoff effect is poorly constrained, leading to substantial inter-model uncertainty in runoff. The physiological impacts of CO2 on hydrology are often inferred from idealized simulations using a single Earth system model (ESM) (11, 14), or from multi-model means across an ESM ensemble (15, 17, 18). However, each ESM represents the land-surface, the atmosphere, and their interactions somewhat differently, implying a diversity of potential mechanisms underpinning plant-driven hydrologic responses. Research on other climate drivers and impacts has widely leveraged the diversity of ESM responses to rising CO2 as a measure of uncertainty in ESM-projected future climates (19–21), and as a window into culpable mechanisms driving that uncertainty (22–24). Yet, such diversity is underexplored for CO2 physiological effects on future runoff (13, 16, 25), with two critical and related implications for hydroclimate science, impacts, and adaptation.
Firstly, underexplored variation across ESM simulations translates to poorly constrained uncertainty over potential runoff increases from CO2 physiological effects. For instance, large but opposing signals in key hydrologic variables across the ensemble may cancel in the ensemble mean, yielding an inaccurate impression of low change and concealing model divergence. Secondly, to the extent that such uncertainties arise from differing representations of key processes, understanding the drivers of ESM spread requires a mechanistic examination of individual model realizations. Crucially, an ensemble mean of ESM runs is not itself a physically consistent run: optimizations and budget closures that are enforced within individual ESMs may be obscured in analyses of ensemble means. Perhaps as a result, the causal path linking CO2-induced transpiration responses to runoff impacts remains obscure (5, 11, 13).
In this study, we clarify the mechanism and robustness of runoff change in a set of 12 ESM simulations that isolate plant physiological and biogeochemical responses to increased CO2 (Supplementary Table 1). We center our analysis on the biogeochemical (BGC) simulations from the Coupled Climate-Carbon Cycle Model Intercomparison Project or C4MIP (26). These BGC runs capture carbon cycle response to increasing atmospheric CO2 (1% per year to a quadrupling over preindustrial levels), while keeping CO2 constant in model’s radiation code. We focus on these BGC experiments for two reasons: firstly, to diagnose hydrologic uncertainties specifically owing to plant effects of enhanced CO2 without the confounding effects of warming and, secondly, because these experiments are increasingly used to assess future water availability. Changes in 30-year mean runoff, precipitation, and transpiration (aggregated to water-year totals and denoted by Δ) are calculated as the difference between the last and first 30 years of each simulation. We apply a modest definition of ensemble agreement for a change to qualify as ‘robust’: at least two-thirds of the models must agree on the sign of change that is at least ¼ a standard deviation in magnitude.
Through an analytic approach that diagnoses model differences in these BGC simulations, we present three key findings. Firstly, we show that plant-driven runoff increases are only robust over a small portion of global land area, despite widespread and robust declines in transpiration. An examination of surface water balance changes reveals that, within individual BGC models, increases and decreases in runoff are equally likely given reduced transpiration, with precipitation changes largely determining the direction of runoff change. Secondly, we use preindustrial control simulations to show that these BGC-only precipitation changes reflect a mix of internal climate variability and significant forced responses to CO2 physiological effects. Lastly, using a typology that interprets CO2 physiological effects on runoff in the context of precipitation changes, we demonstrate that plant-induced reductions in runoff are at least as widespread as plant-induced enhancements, even absent the runoff impacts from climate warming. Together, these results imply a heterogenous and uncertain role for plant stomatal responses to CO2 in shaping surface hydrology and future water availability.
Limited ensemble agreement on transpiration-driven runoff enhancement
The spatial pattern of ensemble agreement across hydrologic variables indicates large uncertainty in the consequences of transpiration changes for runoff under CO2 physiological effects (Fig. 1, Supplementary Figs. 1-3). Robust transpiration declines are ubiquitous in the BGC runs, with ensemble agreement on reductions over 57% of land area (Fig. 1a, Supplementary Fig. 1). Model consensus on declining transpiration indicates that the effect of stricter stomatal regulation outweighs that of increased leaf areas resulting from CO2 fertilization (Supplementary Fig. 2). However, the ensemble disagrees on how these robust physiologically-driven transpiration declines translate to runoff changes. For example, the BGC simulations disagree on the sign of runoff change over a large majority of land area (83%), including across 90% of the area with ensemble agreement on transpiration decreases (Fig. 1a, c, d). In fact, the ensemble agrees on runoff increases over only 5% of global land area, comprising minorities of the Andes (7%), East Asia (16%), Southeast Asia (17%), and central and western Africa (34%) regions (Fig. 1c-d, Supplementary Fig. 4). Whereas these regions have been identified previously as hotspots of runoff enhancement under CO2 physiological effects (11, 14, 17, 18), our analysis reveals prevailing ensemble disagreement on these runoff responses, both globally and locally within these hotspot regions.
Precipitation changes obviously shape the extent to which transpiration declines translate into runoff enhancement (10, 18, 27). However, we find widespread ensemble disagreement over both precipitation changes and their impact on runoff under CO2 physiological effects. The BGC simulations disagree on the sign of precipitation change over 70% of land areas (Fig. 1b, Supplementary Fig. 3), suggesting potentially meaningful but highly uncertain precipitation responses. Notably, the ensemble also disagrees on the location of runoff changes even where models agree on declining precipitation. For example, while the ensemble agrees on precipitation declines over 57% of the Amazon region (Fig. 1b, d), it only agrees on attendant runoff declines over 8%, and otherwise disagrees on the direction of runoff change (Fig. 1c, d). We find a similar pattern of divergent runoff responses despite robust precipitation and transpiration declines in much of northwestern Eurasia. Further, the ensemble does not agree on compensating reductions in transpiration and precipitation, evidenced by minimal agreement on low runoff change. In other regions, such as northeastern Eurasia, central North America, and southeastern Africa, models agree on declining transpiration but disagree on both precipitation and runoff responses (Fig. 1).
In contrast to BGC, ensemble agreement is more consistent across hydrologic variables for the radiative (RAD) and fully-coupled control (CTL) counterpart experiments in C4MIP (Supplementary Fig. 5). For RAD, rising CO2 is presented only to the atmosphere and not the land surface, while for CTL, CO2 increases impact all model components. In both of these experiments, the ensemble agrees on the sign of hydrologic changes over 47-66% of land area (across variables and experiments). The sign of agreement is generally collocated across the three variables for CTL and RAD (increases in northern high latitudes, most of Asia, central Africa, and southeast South America; decreases in the Mediterranean, southern Africa, and the Amazon). These results point to an ensemble consensus that radiatively-driven precipitation changes largely determine the sign of runoff changes under fully-coupled warming (28). They further highlight the comparatively large uncertainty in the hydrologic impacts of CO2 physiological effects, as reflected in BGC.
Together, these results suggest that there is very little ensemble consensus on whether and where stomatal-driven transpiration declines translate into runoff or precipitation changes under rising CO2. Runoff is robustly enhanced by reduced transpiration in small portions of some tropical regions (11, 14, 17, 18), representing a small fraction of global land area. The pattern of discordant transpiration, precipitation, and runoff changes across the BGC ensemble implies that inter-model runoff uncertainty arises not from the transpiration response itself, but rather from the mechanism by which it does (or does not) translate into runoff changes. Since long-term runoff changes approximate the difference between precipitation and ET changes (29), two causal paths may explain why reduced transpiration does not generally correspond to increased runoff. Firstly, decreases in precipitation may compensate transpiration-driven increases in runoff (6, 7, 30). Secondly, increases in other components of ET (notably, bare-soil, Es, and leaf evaporation, EL) may compensate decreases in transpiration, breaking the transpiration-runoff response (31, 32).
Influence of precipitation on runoff from CO2 physiological effects
To examine which of these causal paths explains the weak ensemble relationship between transpiration declines and runoff responses from CO2 physiological effects, we develop a typology to classify grid cells as consistent or inconsistent with a transpiration-decrease/runoff-increase hypothesis in the BGC experiments (Fig. 2, Supplementary Fig. 6). For each ESMs, we find that increases and decreases in runoff are equally widespread across land areas with declining transpiration (ensemble averages of 39 vs 36%, dark blue and light brown bars, Fig. 2a). However, the ensemble largely disagrees on where these changes occur, with robust transpiration-decrease/runoff-increase response only over 3% of mainly-tropical land area (Fig. 2b, Supplementary Fig. 6). Runoff increases and decreases are also roughly equally widespread over regions where transpiration increases (ensemble mean 24%, 8-39% across ESMs, light blue and dark brown bars, Fig. 2a). Yet there is virtually no ensemble agreement on the location of these joint changes (Fig. 2b).
As such, the BGC ensemble disagrees on the location of joint transpiration and runoff changes over 87% of land areas. Yet, individual models consistently simulate runoff increases and decreases with about equal likelihood, irrespective of the sign of transpiration changes. We find that these runoff increases or decreases from CO2 physiological effects are largely explained by precipitation changes within each simulation (Fig. 2c-f). Regardless of the sign or magnitude of the transpiration response in the BGC simulations, the principal difference in water budget between runoff-increasing and runoff-decreasing areas is the sign of the precipitation change (Fig. 2c-d). For regions in which transpiration declines, the magnitude of the decline is similar across both runoff increasing and decreasing regions, implying that the magnitude of transpiration declines does not determine either the magnitude or sign of runoff change. In contrast, precipitation increases (-5 to 85mm/yr, ensemble range) where runoff increases, and where precipitation decreases (-10 to -100mm/yr), so too does runoff. The sign of runoff change similarly conforms with the sign of precipitation change where transpiration increases, but with smaller magnitudes of water budget changes (Fig. 2e-f).
Spatial pattern correlations among hydrologic variables further clarify the relative importance of transpiration versus precipitation changes for determining runoff responses in the BGC experiment. The strength of these correlations indicates the spatial dependence of runoff responses on changes in precipitation minus evapotranspiration (ΔP-ΔET), precipitation minus transpiration (ΔP-ΔTr), precipitation, and transpiration. ΔP-ΔET, a typical diagnostic of climatological runoff response, is quite strongly correlated with prognostic runoff change (ensemble mean correlation of 0.82), reflecting surface water budget closure (29) (dark blue points, Fig. 2g). Pattern correlation between runoff change and ΔP-ΔTr is nearly as strong (ensemble mean of 0.66, light blue points, Fig. 2g), reflecting the dominant contribution of transpiration to evapotranspiration in non-barren areas (33–35). However, whereas runoff change depends nearly as strongly on precipitation change as ΔP-ΔTr (ensemble mean of 0.53), transpiration change itself is very weakly correlated with runoff change (ensemble mean of 0.23, purple and green points, Fig. 2g). This tendency is reflected across the ensemble: the runoff-precipitation correlation exceeds 0.5 in nine of twelve models while the runoff-transpiration correlation falls below 0.5 in all twelve models. The spatial variation in runoff change explained by precipitation change exceeds that explained by transpiration change by over a factor of 5 (ensemble mean R2 of 0.28 vs 0.05).
Moreover, we do not find that combined changes in bare-soil and leaf evaporation consistently explain the correspondence, or lack thereof, between transpiration and runoff in the BGC ensemble. Where transpiration decreases, evaporation change is insufficient to alter the sign of evapotranspiration change, even if it somewhat limits transpiration-driven evapotranspiration declines in some models (Fig. 2c-d, Supplementary Fig. 7). Models are generally split on the sign of evaporation change, reflecting considerable uncertainty in the responses of these evapotranspiration components to the physiological effects of high CO2 (31, 32), but their ultimate influence on evapotranspiration is moderate.
These results indicate a leading role for precipitation in generating runoff responses, even in simulations where plants alone respond to CO2 change. Transpiration-driven changes in total evapotranspiration do modulate the degree to which precipitation change translates into runoff change. For example, where transpiration declines in the BGC experiments, runoff increases more than precipitation increases, or decreases by less than precipitation decreases (5) (Fig. 2c-d). However, while transpiration reductions shift the water balance relatively towards runoff, absolute runoff declines are equally likely given a reduction in transpiration. The sign of runoff responses within individual models is largely explained by precipitation changes, but the sign and location of these precipitation responses are highly uncertain across models (Fig. 1b). Together, these results tell us that runoff uncertainties arise because of precipitation uncertainties under CO2 physiological effects. In other words, the discrepancy between transpiration and runoff changes (Fig. 1a and c) is explained by uncertainty in precipitation (Fig. 1b).
Interpreting vegetation-forced precipitation changes
The role of precipitation in determining runoff responses to CO2 physiological effects in individual models raises the question of whether precipitation change is itself an atmospheric signal forced by vegetation responses to CO2 (6–8) or simply noise arising from simulated internal variability. Addressing this question is central to interpreting the importance of stomatal effects for future water availability. Where precipitation change in the BGC simulations reflects internal variability, precipitation-induced runoff changes would reflect noise rather than a meaningful signal of vegetation responses to CO2. But where precipitation change is itself a significant forced signal, precipitation-induced runoff change actually reflects influences of plant physiological changes on precipitation via atmospheric responses. To answer this question, we estimate the significance of precipitation in the BGC simulations relative to the distribution of 30-year mean precipitation anomalies in the preindustrial control experiments for each ESM.
The precipitation changes that strongly influence runoff under CO2 physiological effects reflect a highly uncertain mix of plant-forced responses and internal variability. Across the BGC ensemble, precipitation change is significantly different from preindustrial variability over 31-57% of land area (two-tailed test, p < 0.1, ensemble mean of 44%), with drying about three times more extensive than wetting (Fig. 3a, Supplementary Fig. 8). Yet, there remains little agreement across models on where these forced changes occur. We find no agreement on the location of plant-forced precipitation increases (Fig. 3b), despite their occurrence over 14% of land on average within any given model (Fig. 3a). Similarly, the ensemble simulates forced precipitation decreases over 30% of land on average, but agrees on their location for only 7%, mainly in the Amazon and northwestern Eurasia (Fig. 3b, Fig. 1b). Plant-forced precipitation declines have been highlighted previously for the Amazon (7, 36). However, we find that model agreement on significant precipitation changes, but disagreement on sign, is even more widespread, highlighting model uncertainty in forced precipitation responses. This occurs over 11% of land area, mainly in northern Eurasia, central Africa, and south America (pink areas, Fig. 3b). Despite relatively widespread forced precipitation signals, precipitation change is not significant over a majority of land area in 10 out of 12 ESMs (Fig. 3a), with ensemble agreement over 38% of land area (Fig. 3b). Nevertheless, disagreement on the significance and/or sign of precipitation change covers a majority of land area (55%, Fig. 3b, Supplementary Fig. 8).
Although transpiration-runoff pattern correlations are generally weak in each BGC run (Fig. 2g), we find that they are strongly negatively related to transpiration-precipitation pattern correlations (Fig. 4). The transpiration-precipitation relationships are also generally weak (ensemble mean r = 0.27, vertical dotted line, Fig. 4a), likely because forced precipitation changes cover a minority of land area within most models and may not be collocated with (but occur downwind of) transpiration changes (Fig. 3). Nevertheless, the strong link between the two (r = -0.82, slope = -1.0, p = 0.001, Fig. 4a) indicates that the scope for direct transpiration influences on runoff via water savings is mechanistically limited in models with stronger precipitation responses to transpiration change. Further, runoff-precipitation dependence does not scale with either transpiration-runoff (p = 0.55, Fig. 4a, axis annotations) or transpiration-precipitation dependence (p = 0.41, Fig. 4a, axis annotations).
Thus, under CO2 physiological effects, precipitation most strongly determines runoff responses independent of transpiration-driven effect (Fig. 4a-b). Meanwhile, transpiration only meaningfully affects runoff in models with especially weak atmospheric responses to transpiration (i.e., models with low transpiration-precipitation pattern correlations, Fig. 4a-b). Moreover, we do not find that the transpiration fraction of total evapotranspiration (Tr/ET) explains the strength of relationships among transpiration, runoff, and precipitation across the ensemble (Fig. 4c-e). This Tr/ET ratio has been previously suggested as a key control on the potential impact of vegetation changes on the water cycle in ESMs (16). The model spread in global mean leaf area growth due to CO2 fertilization also does not explain these relationships, nor does it explain global mean transpiration changes or the fraction of global land area with declining transpiration (Supplementary Fig. 2). Instead, our results implicate model differences in precipitation sensitivity to CO2-driven transpiration changes as the major source of model disagreement over CO2 physiological effects on runoff.
Given their strong dependence on precipitation changes, runoff responses under CO2 physiological effects are most accurately interpreted relative to precipitation changes and their significance. To this end, we present a typology of six precipitation-runoff effects that characterizes the diverse hydrologic consequences of plant CO2 responses. For plant stomatal responses to unequivocally boost runoff in the BGC experiments, runoff must increase more than precipitation change times the low-CO2 (or baseline) runoff efficiency (ΔP*, Fig. 5, dark blue bars, Methods). This case implies plant-driven increases in runoff efficiency, either enhancing precipitation-driven wetting or dampening precipitation-driven drying. Where runoff decreases, but less so than ΔP*, plants partly limit precipitation-driven drying. This case qualifies as relative plant-driven wetting only where the precipitation-driven drying is purely stochastic (Fig. 5, light blue bars). Where precipitation declines are forced, plant responses to higher CO2 instead diminish runoff: enhanced stomatal regulation reduces precipitation and only partly alleviates the ensuing runoff impacts (Fig. 5, hatched tan bars). Where runoff increases, but less so than ΔP*, plants limit precipitation-driven wetting (Fig. 5, dark brown bars), and where runoff decreases, but more so than ΔP*, plants amplify forced or unforced precipitation-driven drying (Fig. 5, hatched and unhatched light brown bars).
We find that plant-induced runoff drying in response to CO2 physiological effects (brown outlined bars, Fig. 5) is as common across the BGC experiments as plant-induced wetting (blue outlined bars), covering 36-59% of land area (ensemble mean of 50%) and representing the dominant response in 5 out of 12 models (Fig. 5). Further, about half of this plant-induced runoff drying is attributable to forced precipitation declines (10-37% of land area, ensemble mean of 27%, hatched bars, Fig. 5). Thus, when runoff responses are accurately contextualized with precipitation changes, ESMs exhibit little preference for plant-induced wetting under CO2 physiological effects alone.