Atmospheric dryness reduces photosynthesis along a large range of soil water deficits

The study addresses how soil water content (SWC) and atmospheric dryness (vapor pressure deficit, VPD) independently and jointly control terrestrial gross primary production (GPP), despite their strong covariation. Droughts threaten ecosystems and carbon uptake, but prior work has reported conflicting conclusions about whether SWC or VPD dominates GPP reductions. Plants close stomata under high VPD to limit water loss, and low SWC further constrains stomatal conductance and hydraulic supply. The research aims to (1) identify under which SWC and VPD conditions GPP is most negatively affected; (2) test whether reductions in GPP from stomatal closure with decreasing SWC are compensated by increases in photosynthetic biochemical capacity; and (3) evaluate whether Earth System Models (ESMs) capture the relative influence of VPD and SWC on GPP. The 2018 European mega-drought provides a natural experiment for extreme conditions, and a global analysis with nonlinear machine learning (ANN) is used to separate sensitivities to VPD and SWC while accounting for radiation and temperature effects.

The paper synthesizes prior findings showing both SWC and VPD can limit GPP and canopy conductance, but with disagreement on their relative importance. It highlights that VPD and SWC typically covary via land–atmosphere feedbacks, complicating attribution. Leaf-level and theoretical studies indicate stomatal conductance declines strongly with increasing VPD to prevent hydraulic failure, and SWC exerts a threshold-like influence on plant water potential. Previous global assessments using SIF suggested predominant SWC control but may have underplayed VPD due to sensor overpass timing and the weaker sensitivity of SIF to stomatal regulation, as well as neglect of VPD–radiation coupling. The study builds on these insights to disentangle drivers at ecosystem scales using flux tower observations and nonlinear models.

Data sources: Half-hourly or hourly eddy-covariance measurements from ICOS and FLUXNET2015 were used, including GPP (nighttime partitioning reference product), air temperature, VPD, SWC (layered where available), and incoming shortwave radiation. Standardized quality control, gap filling, energy balance correction, and flux partitioning were applied by data providers. Only measured and good-quality gap-filled data (QC 0 or 1) were used. Cropland and wetland sites were excluded. 67 sites with at least 300 growing-season days were used for relative effects; 59 sites met performance criteria for ANN analyses. European 2018 drought analysis: 15 ICOS sites (2014–2018) spanning major ecosystems (excluding croplands and wetlands) were analyzed. Daytime data (07:00–19:00) were aggregated to daily values. Summer (JJA) relative changes (ΔX) of GPP, SWC, VPD were computed against the 2014–2018 summer average. Daily z-scores for GPP, SWC, VPD were regressed using multiple linear regression: GPP = β1·SWC + β2·VPD + β3·(SWC×VPD) + β4·Temperature + β5·Radiation + intercept + error. Slopes were compared including vs excluding 2018 to assess nonlinearity. Global analysis and filtering: Analyses focused on growing-season days with temperature > 15 °C, VPD > 0.5 kPa, and incoming shortwave radiation > 250 W m−2 to emphasize conditions where SWC/VPD influences are most relevant. Variables were standardized (z-scores). SWC and VPD were binned into 10×10 percentile bins per site. ANN and sensitivity estimation: Feed-forward ANNs (one hidden layer, 10 nodes; robustness checked for 4–20 nodes) were trained per site using daily predictors: temperature, VPD, SWC, and radiation, with response variables GPP (and separately Gc, Amax, Vcmax, iWUE). Data split: 60% training, 20% validation, 20% testing; optimization via Levenberg–Marquardt, up to 1000 epochs. Performance assessed via correlation and RMSE; most sites had r > 0.7. Sensitivities were computed by perturbing one predictor by +1 standard deviation (given normalized inputs) while holding others unchanged, and calculating the fractional change in predicted response; repeated five times and median taken. Median sensitivities per SWC/VPD bin were aggregated across sites; significance tested via t-tests (p < 0.05). Deriving physiological variables: Canopy conductance (Gc) was computed by inverting the Penman–Monteith equation using fluxes and meteorology; aerodynamic resistance used wind speed, measurement height, and canopy height inferred following surface-layer theory. Maximum photosynthetic assimilation rate (Amax) was derived from daytime light response curve fits to non-gap-filled CO2 flux (Fc) using REddyProc, with VPD limitation above 1 kPa. Internal CO2 (ci) was estimated via Fick’s law using GPP and resistances (rc = 1.6/Gc), then Vcmax was computed from the Farquhar model, with temperature normalization to 25 °C via Arrhenius functions. iWUE was defined as GPP/Gc. Sensitivity and uncertainty analyses were repeated for deep SWC layers (where available), by plant functional type, peak growing season, and deseasonalized anomalies. Uncertainty analyses: Partitioning method uncertainty was assessed by comparing nighttime vs daytime GPP products; NEE processing uncertainty via quartile GPP products (GPP_NT_VUT_25/75). For Gc, both LE and energy-balance-corrected LE.CORR were tested. Standard errors of sensitivities per bin and relative uncertainties were computed. Bins at extreme SWC/VPD percentiles had fewer data and larger uncertainties. Comparison with CMIP6 ESMs: Five daily-output ESMs (ACCESS-ESM1-5, CMCC-CM2-SR5, IPSL-CM6A-LR, NorESM2-LM, NorESM2-MM) provided GPP, temperature, radiation, surface soil moisture, and VPD (computed). Site-extracted historical simulations (1995–2014) were analyzed identically (z-scores, ANN, binning). Modeled minus observed sensitivities were mapped to assess biases. Reproduction of Liu et al. approach: The difference in GPP (and radiation-normalized GPPi) between high vs low VPD at fixed SWC bins yielded ΔGPP(VPD|SWC); between high vs low SWC at fixed VPD bins yielded ΔGPP(SWC|VPD), to assess relative roles and the effect of radiation coupling.

• 2018 European drought: Summer mean SWC was 25 (±5)% lower and VPD 22 (±4)% higher than 2014–2018 averages across 15 sites, with GPP reduced by 15 (±5)%. SWC and VPD anomalies covaried strongly.
• Linear sensitivity shifts with extremes: Standardized partial regression slopes of GPP vs SWC were −0.22 (−0.14 to −0.31, 95% CI) for 2014–2018, but weakened to −0.14 (−0.06 to 0.23) excluding 2018, indicating nonlinear SWC effects; VPD slopes were consistently negative and similar whether including 2018 (−0.45, −0.32 to −0.58) or not (−0.49, −0.37 to −0.60).
• Nonlinear sensitivities from ANNs (global): • GPP vs SWC: Negative sensitivity strengthens as SWC decreases and becomes significant below ~70th SWC percentile. At high SWC (>70th percentile), decreasing SWC can increase GPP (positive sensitivity), indicating beneficial effects of moderate soil drying under wet conditions. • GPP vs VPD: Negative sensitivity to increasing VPD persists across the full SWC range, with larger magnitude at lower VPD bins.
• Physiological underpinnings: Canopy conductance (Gc) decreases with increasing VPD across all SWC levels and decreases with declining SWC mainly under dry soils. Amax shows negative sensitivity to VPD but positive sensitivity to decreasing SWC when SWC is high; Vcmax sensitivities mirror Amax and GPP. These indicate compensation of partial stomatal closure by increased biochemical capacity under moderately wet soils.
• Deep soil effects: Using deeper SWC layers accentuated negative GPP sensitivities to both SWC decreases under dry soils and VPD increases under wet soils, implying deeper-layer droughts can cause larger GPP reductions.
• Robustness to processing: GPP sensitivity patterns were consistent between nighttime and daytime partitioning and across NEE processing quartiles; differences mostly within ±0.1. Gc sensitivity patterns were robust to LE vs LE.CORR choices.
• Plant functional type differences: Grasslands/savannas exhibit more negative GPP sensitivity to decreasing SWC than forests (DBF, ENF), likely due to shallower access to soil moisture. Forests show positive GPP and Amax sensitivities to decreasing SWC over wider SWC ranges (intermediate to high), suggesting differing SWC thresholds for biochemical upregulation.
• iWUE increases with soil drying: iWUE (GPP/Gc) shows generally positive sensitivity to decreasing SWC, more positive under high VPD; sensitivity to decreasing VPD is more negative under low VPD.
• Relative roles of VPD vs SWC: VPD generally dominates dryness stress on GPP except under low SWC conditions (<30th percentile), where SWC limitations can exceed VPD effects. Along VPD gradients, VPD effects remain more negative than SWC effects, though SWC effects intensify with higher VPD.
• Reconciling SIF-based studies: When radiation effects are removed (using GPP normalized by radiation), high VPD limitations exceed low SWC limitations across sites; neglecting radiation coupling can lead to an overemphasis on SWC effects in prior SIF-based analyses.
• ESM evaluation: All five CMIP6 ESMs reproduced the general negative GPP sensitivity to VPD. Three models (with Community Land Model land components) captured the sign change in SWC sensitivity (negative at low SWC, positive at high SWC). However, all ESMs underestimated the magnitude of negative VPD sensitivity (0.19 ± 0.12 underestimation) and the positive sensitivity to decreasing SWC at high SWC (−0.20 ± 0.07 modeled-minus-observed), implying underestimation of future GPP reductions from rising VPD and missing biochemical compensation under moderate soil drying.

The study disentangles the intertwined effects of soil and atmospheric dryness on photosynthesis, showing that VPD exerts a consistent negative control on GPP across soil moisture states due to rapid stomatal responses to atmospheric demand and leaf water potential. In contrast, SWC impacts are threshold-like: under wet to moderately moist soils, slight drying can enhance biochemical capacity (Amax, Vcmax) and maintain or increase GPP despite modest stomatal limitations, whereas under dry soils, further drying sharply reduces GPP. These findings reconcile conflicting literature by demonstrating that the relative dominance of VPD vs SWC depends on the background soil moisture state and on accounting for radiation–VPD coupling. They reveal an emergent compensation mechanism at ecosystem scale consistent with optimality theory: trade-offs between stomatal conductance and carboxylation capacity allow maintenance of photosynthesis under moderate stress, increasing iWUE. Differences among plant functional types suggest varied SWC thresholds for biochemical upregulation and drought resistance strategies (e.g., deeper rooting in forests). The systematic underestimation by ESMs of VPD impacts and wet-soil biochemical compensation indicates potential biases in projecting future land carbon sink responses as warming increases VPD and SWC trajectories diverge regionally. Incorporating plant hydraulics and dynamic adjustments of carboxylation capacity is crucial to improve model realism, especially under extreme events.

This work demonstrates that atmospheric dryness (higher VPD) consistently reduces photosynthesis across a broad range of soil moisture states, whereas soil drying reduces GPP mainly when soils are already dry; under wet soils, moderate drying can enhance photosynthetic capacity and sustain or even increase GPP. Physiological diagnostics (Gc, Amax, Vcmax) reveal a compensatory increase in carboxylation capacity under moderate soil drying that offsets partial stomatal closure, increasing iWUE. VPD generally dominates dryness stress on GPP except under low SWC. Reconciliation with previous SIF-based studies emphasizes the need to account for VPD–radiation coupling. State-of-the-art ESMs capture qualitative patterns but underestimate the magnitude of sensitivities, particularly the negative VPD effect and the positive SWC effect at high SWC, implying potential underestimation of future GPP declines with rising VPD. Future research should: (1) incorporate plant hydraulics and dynamic photosynthetic capacity adjustments into ESMs; (2) improve representation of deep soil moisture and rooting strategies; (3) expand observational constraints across biomes and extremes; and (4) further decouple co-varying drivers (radiation, temperature, VPD, SWC) to refine attribution under climate change.

• Data scope and representativeness: Analyses used 67 sites globally (59 for ANN) excluding croplands and wetlands, potentially limiting generalizability to managed or wetland ecosystems. Some bins (extreme SWC/VPD percentiles) had fewer data, increasing uncertainty.
• Processing and methodological uncertainties: Although robust to partitioning and NEE processing variants, GPP estimates inherit uncertainties from flux partitioning and gap-filling. Gc depends on latent heat flux (with energy balance correction assumptions) and aerodynamic parameterizations. Amax estimation via light response curves assumes stable parameters within short windows and a fixed VPD threshold (1 kPa) for downregulation. Vcmax derivation relies on inferred ci from ecosystem-scale resistances and may accumulate measurement errors.
• ANN approach: While designed to avoid overfitting and with good performance at most sites, ANNs are empirical and sensitivities are local to the training data distributions; 8 sites failed performance criteria. Results depend on predictor selection and standardization.
• Temporal and environmental filtering: Analyses focus on growing-season, high-light, warm days (T>15 °C, VPD>0.5 kPa, radiation>250 W m−2), which limits inference to such conditions and may exclude important dynamics under cooler or low-radiation conditions.
• Model intercomparison: Only five CMIP6 ESMs with daily GPP outputs were included; conclusions about model biases may not generalize to all ESMs. Soil moisture used from models is surface layer; deep soil processes may be underrepresented.
• Attribution complexity: Despite efforts to account for radiation and temperature, residual co-variation and site-specific factors (species composition, management history, nutrient status) may influence sensitivities.