Widespread and complex drought effects on vegetation physiology inferred from space

Soil moisture droughts are increasing in duration and intensity globally and impact vegetation functioning, elevating risks of carbon starvation and hydraulic failure that can cause plant mortality. Because terrestrial vegetation regulates carbon and water fluxes, plant drought responses feed back to climate and may exacerbate warming. Vegetation functioning reflects both structural (e.g., leaf area) and physiological (e.g., stomatal conductance, photosynthetic capacity) components, which can respond differently to stress. Physiological responses typically occur faster than structural adjustments, so diagnosing drought impacts solely via structural changes can underestimate functional responses. Advances in satellite remote sensing enable new opportunities: SIF as a proxy for photosynthesis (here used as relative SIF, SIFrel), LST converted to ET via a simplified surface energy balance model, and VOD ratio (midday/midnight X-band VOD) as an indicator of canopy water content dynamics and stomatal regulation. The study aims to synergistically use SIFrel, ET, and VOD ratio to quantify overall and physiological vegetation responses to drought globally, disentangle structural versus physiological effects using machine learning, and interpret mechanisms with process-based SCOPE simulations.

The paper situates its work within prior findings that satellite greenness indices and LAI have been widely used to track structural responses to drought, while physiological parameters (e.g., maximum carboxylation rate, stomatal conductance) have been assessed mainly at site level and are only implicitly represented in global observations. Previous studies show early drought-induced reductions in stomatal conductance and maximum photosynthetic rate preceding structural declines. Remote sensing advances include TROPOMI-derived SIF with reduced cloud-induced biases; SIFrel helps filter solar irradiance and angular effects. LST is tightly linked to ET, though global ET cannot be directly observed, motivating model-based inference. Microwave VOD provides sensitivity to vegetation water content and diurnal hydraulics; X-band VOD ratio captures stomatal water-saving strategies. Prior research also documents contrasting responses in wet versus dry regions, with energy limitation and atmospheric dryness modulating photosynthesis, greenness, and ET. The study builds on and integrates these strands to disentangle physiology from structure at large scales.

Data and study design: Global analysis at 0.25° spatial and 8-daily temporal resolution from March 2018 to October 2021 using concurrent datasets. Vegetation function proxies: SIFrel from TROPOMI SIF normalized by near-infrared reflected radiance; ET estimated from MODIS LST via a simplified surface energy balance model (SSEB) using ERA5-Land meteorology; VOD ratio computed as midday/midnight X-band VOD (AMSR2 LPDR v2) to reflect daytime versus nighttime canopy water content dynamics. Structure proxies: MODIS LAI (MOD15A2H) and NIRv (MCD43C4). Hydro-meteorological drivers: ERA5-Land (air temperature, incoming shortwave and longwave radiation, surface pressure, atmospheric vapor pressure, wind speed, precipitation, and 1 m soil moisture aggregated from layers). Aridity index defined as mean net radiation divided by precipitation (2018–2021), with higher values indicating drier climates. Preprocessing: All data aggregated to 8-daily using a 16-day moving window with 8-day overlap; apply quality filters (e.g., cloud fraction >0.5 removed for SIF; MODIS quality flags; exclude windows with >20% missing). Time series were de-seasonalized and detrended via locally-weighted smoothing to obtain anomalies. Spatial masks: remove areas with sparse vegetation (<5%) and high irrigation (>10%); classify vegetation by tree vs grass+shrub dominance using ESA CCI land cover fractions (threshold tree/(grass+shrub)=0.5). Drought detection: Focus on growing season (T>5°C and mean seasonal SIF>0.2 mW m−2 sr−1 nm−1). Use 40-year ERA5-Land monthly soil moisture minima (1982–2021) to identify grid cells with severe droughts in 2018–2021; the drought peak per grid cell is the lowest 8-daily soil moisture in 2018–2021. Analyze anomalies from 3 months before to 3 months after peaks. Define drought development and recovery durations as time steps to zero/positive soil moisture anomalies before/after peak. Disentangling physiology vs structure: For each grid cell and variable (SIFrel, ET, VOD ratio), fit two random forest models: (1) structure-only model using LAI (or alternatively NIRv) to predict the variable (captures structural component); (2) full model using LAI plus hydro-meteorological anomalies to predict the variable (captures combined structural + physiological effects; improves robustness to noise). Apply a leave-out strategy excluding drought-period time steps from training. Use out-of-bag R² to filter unreliable locations (retain grid cells with OOB R²>0). The physiological component is estimated as the difference between predictions from the full model and the structure-only model during drought periods. Robustness checks: Alternative decomposition using SHAP contributions of hydro-meteorological variables from the full RF model, and multiple linear regression variance decomposition; both yielded similar patterns. Tested different leave-out window sizes (6, 12, 24 time steps) with similar results. Tested stricter drought selection (-1.5 SD threshold) and anomaly-based drought detection; patterns were consistent though magnitudes varied. ET validation: Compare ET’ (latent heat proxy) from SSEB with eddy covariance latent heat at 73 sites (47 with sufficient data); median correlation 0.88 for growing seasons and 0.8 during drought windows. VOD ratio curation: Exclude regions where growing-season midday VOD exceeds midnight VOD to reduce bias from incomplete nocturnal refilling. Attribution analysis: Train random forest models to predict spatial patterns of physiological anomalies (SIFrel physio, ET physio, VOD ratio physio) averaged over development and recovery periods using predictors: aridity, vegetation composition (tree/(grass+shrub)), drought duration, and hydro-meteorological anomalies (radiation, precipitation, temperature, VPD, soil moisture) during development and recovery. Use cross-validated performance (R²>0.35 typically) and SHAP values to rank variable importance; corroborate with Spearman correlations. SCOPE simulations: Run SCOPE v1.73 for ~600 randomly selected drought-affected grid cells (subset from initial 1000) using hourly ERA5-Land meteorology and observed LAI, NDVI, LST; simulate photosynthesis, stomatal conductance (Gs), light use efficiency (LUE), water use efficiency (WUE), and physiology-driven SIFrel (difference between dark-adapted and light-adapted SIF normalized by reflected radiance at 740 nm). Aggregate to 8-daily anomalies and compare with observations during drought phases to interpret mechanisms.

• Across global drought events, vegetation functional declines are largely driven by downregulation of physiology (stomatal conductance and light use efficiency), with the strongest effects in water-limited (transitional to semi-arid) regions. - Physiological anomalies account for the majority of total functional anomalies: 60–97% across SIFrel, ET, and VOD ratio, depending on aridity class and drought phase. - Contrasting wet vs dry region behavior: In dry regions, LAI, SIF, VOD, and ET anomalies decrease during drought development; in wet regions, LAI and NIRv often increase pre-peak (energy-limited systems under sunnier conditions), while physiological downregulation still occurs, producing decoupling between structure (positive) and photosynthetic physiology (negative) especially evident for SIFrel. - SIFrel and ET show pronounced negative physiological anomalies near drought peaks; VOD ratio exhibits positive anomalies (daytime VOD relatively higher than nighttime), indicating stomatal closure and water-saving strategies under high VPD and soil dryness. - Midday vs midnight VOD: In dry regions, midnight VOD anomalies become more negative than midday, yielding positive VOD ratio anomalies through most of the drought period; this pattern is weaker/absent in wet regions. - ET anomalies are generally positive before drought peaks in wet regions and negative in dry regions; ET estimates agree with eddy covariance data across climates and covers, with slightly higher accuracy in drier regions. - Timing: Physiological anomalies emerge about one month before drought peaks in dry regions, earlier than in wetter regions; strongest downregulation occurs around peak in sub-humid and semi-arid areas. - Drivers of spatial variability during drought development: Aridity is the dominant control; vegetation composition (lower tree cover, higher grass/shrub fraction) is associated with stronger physiological downregulation. Hydro-meteorological anomalies modulate specific variables (radiation for SIFrel, precipitation for ET, VPD for VOD ratio). Drought development duration strongly affects SIFrel physiology. - During recovery, concurrent soil moisture and VPD anomalies are primary controls of physiological recovery, with hydro-meteorological variables generally more important than climatology/vegetation characteristics; VOD ratio recovery remains mainly controlled by aridity. - SCOPE simulations reproduce observed patterns: strong decreases in SIFrel physiology and LUE in sub-humid/semi-arid regions; decreases in stomatal conductance in transitional/dry regions; comparatively smaller decreases in WUE in drier regions. Simulated recovery is faster (model lacks explicit soil moisture stress legacy), but development/peak patterns align with observations, supporting the mechanistic interpretation that downregulated stomatal conductance and light use efficiency drive the observed signals.

The study addresses the challenge of separating structural and physiological responses of vegetation to drought at large scales by fusing multiple satellite data streams and applying machine learning to disentangle components. The findings demonstrate that observed declines in photosynthesis (SIFrel) and evapotranspiration are predominantly due to physiological downregulation, particularly in drier climates and in shrub/grass-dominated systems. In wet regions, physiological decreases coincide with structural greening under drought development, explaining discrepancies between structural indicators and functional performance. The cross-consistency among SIFrel, ET, and VOD ratio physiology, together with process-based SCOPE simulations, underscores stomatal regulation and reduced light use efficiency as key mechanisms. These insights clarify how ecosystem physiology modulates land–atmosphere carbon and water fluxes during drought, with implications for amplifying or damping climate extremes via land–climate feedbacks. The ability to isolate physiological signals provides a pathway to improve Earth system models through better parameterization of photosynthetic capacity, stomatal behavior, and representation of drought stress pathways.

By integrating SIFrel, ET inferred from LST, and VOD ratio with a machine learning framework, the study quantifies and isolates physiological vegetation responses to drought globally. It shows that physiological downregulation dominates functional declines, strongest in transitional to semi-arid regions and in grass/shrub systems, while wet regions can exhibit structural greening despite physiological stress. SCOPE simulations corroborate the mechanisms, linking observed physiology to reductions in stomatal conductance and light use efficiency. These results enhance understanding of ecosystem responses and land–climate feedbacks under drought and support improving Earth system models via refined physiological parameterizations and inclusion of appropriate drought stress mechanisms. Potential future research directions include extending analyses to longer time series and additional sensors, explicitly incorporating soil moisture stress legacy and groundwater access in models, refining structural proxies to reduce collinearity with meteorology, and enhancing global observations of deep water access and plant hydraulic traits.

• Structural simplification: LAI (and alternatively NIRv) is assumed to capture all relevant structural changes; any physiological regulation reflected in LAI within 8-day steps will be treated as structural, potentially underestimating the physiological component, while LAI uncertainties may overestimate physiology. - ET attribution: The physiological component of ET may still contain direct meteorological influences (e.g., atmospheric demand, aerodynamic conditions) that are not exclusively physiological. - Model performance and filtering: Random forest prediction of anomalies is inherently challenging; grid cells with OOB R²≤0 are excluded, potentially biasing spatial coverage. - VOD ratio assumptions: Assumes near-equilibrium between pre-dawn leaf and root-zone potentials; regions with incomplete nocturnal refilling may bias VOD ratio; mitigated by excluding areas with midday>midnight VOD on average. - Temporal and spatial coverage: Analysis limited to March 2018–October 2021 and 0.25° resolution may miss finer-scale or longer-term dynamics. - SCOPE limitations: Simulations do not explicitly include soil moisture stress and drought legacy, leading to faster modeled recovery compared to observations. - Decomposition sensitivity: Variance decomposition methods can underestimate structural components due to collinearity and sensitivity to the number of predictors, though main results are robust across alternative methods and parameter choices. - Potential observational and reanalysis uncertainties: Noise and retrieval errors in SIF, LST, VOD, LAI, and uncertainties in ERA5-Land meteorology and soil moisture can affect estimates.