Direct vegetation response to recent CO2 rise shows limited effect on global streamflow

Streamflow is a vital component of freshwater supply, supporting ecosystems, agriculture, and human use. Detecting and explaining long-term changes in streamflow is essential for water resources management. Over recent decades, substantial climate variability is expected to have altered streamflow. Elevated atmospheric CO2 (eCO2) influences the water cycle and land surface processes both directly—through plant physiological and structural changes—and indirectly—via impacts on radiation and temperature that alter precipitation and potential evapotranspiration. Direct vegetation responses to eCO2 can have opposing effects: reduced stomatal conductance and reduced soil evaporation from increased leaf area can increase streamflow, whereas increased transpiration and intercepted evaporation from greening can decrease it. The net effect remains uncertain. Prior studies have reported conflicting conclusions regarding the magnitude and sign of eCO2 effects on runoff, and model-based attributions often lack observational support. Most earlier analyses focused on periods before 2010 or short records. Given a 21.8% rise in CO2 from 1981 to 2020 and pronounced warming, this study aims to quantify how vegetation’s direct response to eCO2 has influenced streamflow at catchment and global scales over the past four decades, using large observed streamflow datasets and two complementary modeling frameworks focused on direct regulation pathways.

Previous research presents conflicting views on eCO2 impacts on streamflow. Some studies attribute increased runoff to plant water-saving effects from stomatal closure under higher CO2, implying a positive eCO2 contribution to streamflow. Others find that observed streamflow trends are primarily driven by climate variability/change and land-use change rather than direct eCO2 effects. Global modeling studies (e.g., Gedney et al. vs. Piao et al.) have produced inconsistent attributions over the last century, highlighting substantial model uncertainty and a lack of robust observational constraints. Many historical assessments focus on the second half of the 20th century or end by 2010, or rely on short experimental periods (1–3 decades), limiting inference for recent decades. Indirect eCO2 effects on climate (radiation, temperature, precipitation, and potential evapotranspiration) are often included in climate change assessments, but disentangling direct vegetation effects from these indirect pathways remains challenging and contributes to low confidence in previous attributions.

Data and study design: The authors compiled annual streamflow data from >20,000 catchments worldwide and selected 1,116 unimpacted small-to-medium catchments for catchment-scale attribution, plus 44 large basins for global model constraint. Selection criteria for the 1,116 catchments included: area <100,000 km²; land-use/vegetation type changes <5% (HILDA+); no reservoir regulation (reservoir capacity divided by multi-year average streamflow = 0); irrigated area <5% (GMIA); consistent dominant vegetation type over all 40 years; and ≥30 years of continuous observations. Additional screening removed catchments with abrupt runoff coefficient shifts, yielding similar results in a 550-catchment subset. Climate classifications used Köppen–Geiger maps. Reservoir datasets included Basin ATLAS, GRanD v1.3, and GDAT. Precipitation came from MSWEP v2.8; potential evapotranspiration (ETp) from MSWX with FAO Penman–Monteith (Yang) formulation; LAI from GIMMS3g V4_1.

Trend analysis: Trends were estimated by Sen’s slope; significance via Mann–Kendall test (α=0.05).

Catchment-scale fully differential method: Observed streamflow Qobs is modeled as a function of hydroclimatic variables. A fully differential (increment-based) standardized multiple regression relates annual increments of streamflow to increments in three drivers: annual precipitation (P), annual potential evapotranspiration (ETp), and annual atmospheric CO2. The regression yields standardized coefficients kP, kETp, kCO2. Factor-specific streamflow changes are dQobs,X = kX × Trend(Q). Absolute and real relative contributions are computed as ΔQobs,X divided by the sum of absolute contributions or by the algebraic sum, respectively. Model fit was evaluated by R²; 75% of catchments had R²>0.55 and 50% had R²>0.7. Multicollinearity was assessed via VIF (<5 indicating no collinearity). Observation error impacts were evaluated with Monte Carlo simulations; even 10% streamflow error induced <0.5% eCO2-contribution uncertainty for 50% of catchments and <0.6% for 75%.

Global ecological models and experiment setup: Fourteen process-based global ecosystem models from TRENDY phase 11 (S3 scenario with time-varying eCO2, climate, and land-use change) provided gridded streamflow, ET, and precipitation at 0.5° resolution (resampled as needed). Using TRENDY control experiments (S0–S2), streamflow trend components were partitioned into: CO2-driven ΔQCO2 = Trend(QS1−QS0), climate-driven ΔQCU = Trend(QS2−QS1), combined ΔQCO2,CU = Trend(QS2−QS0), and interaction ΔQinteraction = ΔQCO2,CU − ΔQCO2 − ΔQCU.

Observation-constrained modeling: To reduce uncertainty, four observation-constrained models were constructed per continent using the 44 large basins: two value-constrained (maximize Nash–Sutcliffe efficiency: Best-VAS-SM single model, Best-VAS-EM ensemble) and two trend-constrained (minimize RMSE of trends: Best-TAS-SM single model, Best-TAS-EM ensemble). A Strategic Random Search algorithm optimized ensemble weights/parameters. Modelled trends for large basins were validated against observations.

Attribution with regularized optimal fingerprinting (ROF): Attribution used the ROF method with streamflow responses to external forcings (climate change, eCO2, land-use change) derived from TRENDY scenarios and internal variability ε estimated from pre-industrial control (CTL) simulations of 47 CMIP6 ESMs. Internal variability covariances were constructed from multiple random 80-year blocks (split into two 40-year blocks) and repeated 100 times; temporal aggregation used non-overlapping 3-, 4-, and 5-year means, yielding 300 realizations. A factor is attributable if the 90% confidence interval of its scaling factor β is >0 and includes 1. Probabilistic attribution Px = (Ax/N)×100% quantifies how often attribution criteria are met across realizations.

Uncertainty assessment: The spread in global trends across the 14 TRENDY models (−0.20 to 0.40 mm yr−2) was quantified; observation constraints reduced the standard deviation by 62%. Alternative CTL selections for ε tested sensitivity of attribution outcomes. Additional analyses examined robustness across deserts-excluded domains, vegetation phenology (growing seasons only), catchment size effects, and ET partitioning uncertainties.

• Catchment-scale attribution (1981–2020, 1,116 unimpacted catchments):

  • Precipitation is the dominant driver of streamflow change, contributing >70% to the overall absolute relative contribution.
  • Potential evapotranspiration contributes <20% and is predominantly negative.
  • eCO2 has a very small effect: absolute relative contribution <8% overall, with median real relative contribution ~0 and no clear sign at the population level.
  • Spatial heterogeneity exists: eCO2 contributions tend to be positive in southeastern North America and negative in eastern Oceania.

• Global-scale trends and attribution:

  • Global area-weighted streamflow anomalies show a small, non-significant increase: 0.09 ± 0.05 mm yr−2 over 1981–2020; excluding deserts yields 0.13 ± 0.05 mm yr−2, also non-significant.
  • More than 80% of global grid cells show insignificant increases; <5% show significant changes.
  • ROF attribution robustly attributes global streamflow changes to climate change (scaling factor >0 and includes 1 in both single- and multi-factor cases).
  • eCO2 and land-use change cannot be reliably attributed to global streamflow changes, especially in the multi-factor setting where their scaling factors fluctuate widely.

• Uncertainty reduction via observation constraints:

  • Observation-constrained models reduce the standard deviation of global trend estimates by 62% compared to the 14-model ensemble.
  • Under varying internal variability datasets, unconstrained models can yield ambiguous attribution to eCO2 or even fail to attribute to climate; observation-constrained results consistently attribute to climate change but not to eCO2.

• Additional insights:

  • Growing-season-only analyses produce results consistent with full-year analyses, reinforcing robustness.
  • In global models, opposing mechanisms—stomatal closure increasing streamflow vs. greening increasing water use—likely offset each other, resulting in limited net eCO2 impacts.
  • Interaction effects between eCO2 and climate on streamflow are of the same order as direct eCO2 effects, and both are much smaller than climate-driven effects.

The findings show that despite a 21.8% rise in atmospheric CO2 since 1981, vegetation’s direct physiological and structural responses have had limited net impact on streamflow over the past four decades. At the catchment scale, precipitation overwhelmingly governs streamflow variability, with ETp secondary and eCO2 contributions near zero on median. Spatial patterns of eCO2 effects align with differences in vegetation density and climate regimes, but the magnitudes remain small relative to climatic drivers. At the global scale, streamflow trends are generally insignificant and can be attributed to climate change, while neither eCO2 nor land-use change can be robustly linked to the observed changes when considered alongside climate. The limited direct eCO2 effect likely reflects offsetting mechanisms: increased water-use efficiency from stomatal closure tending to raise streamflow versus CO2-driven greening and leaf area expansion enhancing transpiration and interception, reducing streamflow. Observation-constrained models markedly reduce uncertainty in trend and attribution estimates and consistently reinforce the primacy of climate forcing. Robustness checks (e.g., growing season analysis, desert exclusion) corroborate the main conclusions. The work highlights the challenge of resolving complex vegetation–climate feedbacks and evapotranspiration partitioning in global models, which contributes to spread in unconstrained projections. Overall, the results clarify that recent global streamflow changes cannot be explained by direct vegetation responses to eCO2 alone and instead reflect dominant climatic influences.

Using a large dataset of observed streamflow and two complementary approaches—a fully differential catchment-scale analysis and a global regularized optimal fingerprinting framework constrained by observations—the study demonstrates that the direct vegetation response to elevated CO2 has had a limited effect on streamflow from 1981 to 2020. Precipitation changes dominate streamflow variability, with ETp secondary and eCO2 contributions small and spatially variable, averaging near zero. Globally, streamflow trends are weak and attributable to climate change, with no robust attribution to eCO2 or land-use change. Observation-constrained modeling substantially reduces uncertainty relative to unconstrained multi-model ensembles. Future research should better quantify vegetation–climate feedbacks under eCO2, improve ET partitioning and observational constraints, integrate constraints on both interannual variability and trends, and expand high-quality observations to reduce attribution uncertainty across regions and scales.

• The frameworks isolate only direct vegetation regulation by eCO2 and cannot resolve indirect effects via climate (e.g., precipitation responses to vegetation feedbacks under eCO2).
• Catchment screening cannot fully eliminate anthropogenic influences, especially in larger basins; uncertainty in the fully differential method increases with catchment size.
• Potential multicollinearity and structural model errors are mitigated but not eliminated; results depend on data quality for streamflow, precipitation, and ETp.
• Streamflow observation errors exist (typically 3–6%); Monte Carlo tests suggest small impacts on eCO2 contribution estimates but some uncertainty remains.
• Global attribution depends on the representation of internal variability ε from CTL datasets; outcomes are sensitive to CTL selection, though ensemble experiments reduce this sensitivity.
• TRENDY models exhibit notable uncertainty in ET partitioning and trend magnitudes; interaction terms and process representations remain imperfect.
• Land-use change effects are included at the global scale but not explicitly isolated in the catchment-scale fully differential framework.