Research report: Vegetation increasingly controlled by atmospheric dryness rather than soil moisture

A new analysis of global satellite records suggests that assumption is quietly failing. Across most of the planet, soil moisture still dominates year-to-year vegetation variability, but its grip is weakening while the influence of atmospheric dryness, measured as vapour pressure deficit, is strengthening. The shift is asymmetric: more ecosystems are moving from soil-moisture control towards atmospheric control than the reverse, and continued warming is expected to intensify the pattern.

The team set out to answer a deceptively simple question that has limited predictions of ecosystem–climate feedbacks under warming: when plants respond to water stress, is it the water available at their roots or the drying power of the air that matters more, and how has that balance changed over time?

Vapour pressure deficit, or VPD, is the gap between how much water vapour the air could hold at saturation and how much it actually holds. Higher values indicate greater atmospheric water demand, a potential driver of plant water stress even when soil moisture is adequate. Because warmer air can hold more moisture, rising temperatures tend to widen this gap, pulling harder on the water inside leaves and forcing stomata to narrow. Sustained stomatal limitation reduces carbon assimilation, which downregulates leaf production and the canopy leaf area index, and chronic hydraulic stress can lead to leaf shedding, branch dieback and tree mortality. All of these responses alter canopy greenness, which is exactly what satellite vegetation indices detect.

To disentangle the two drivers, the researchers combined multi-source satellite observations, reanalysis datasets and machine learning in an integrated analytical framework. The primary vegetation signal was the normalized difference vegetation index, or NDVI, drawn from two complementary products, GIMMS-3G+ and PKU GIMMS, both providing global coverage from 1982 to 2022. The GIMMS-3G+ product is based on corrected and calibrated measurements from the Advanced Very High Resolution Radiometer and accounts for deleterious effects including calibration loss, orbital drift and volcanic eruptions, reporting maximum NDVI values within twice-monthly compositing periods. The PKU dataset, used for robustness, applies machine learning with Landsat samples and AVHRR-MODIS consolidation to eliminate orbital drift and sensor degradation effects. The long record is essential for detecting temporal changes in driver importance, although the authors note that NDVI saturation in dense ecosystems may limit detection of some structural and functional changes.

Soil moisture and atmospheric variables came from the ERA5-Land hourly reanalysis dataset. Root-zone soil moisture was calculated as a depth-weighted mean of the first three soil layers, spanning 0 to 100 centimetres, while vapour pressure deficit was computed from hourly air temperature and dew point data using standard saturation vapour pressure equations. Shortwave solar radiation and air temperature were also extracted so that their contributions could be separated from the water variables. All datasets were harmonized to a monthly temporal resolution and a 0.5-degree spatial grid, and the analysis focused on growing-season months, defined as those with climatological monthly air temperature exceeding 5 degrees Celsius.

Isolating the year-to-year signal required careful processing. The team applied seasonal-trend decomposition based on Loess, known as STL, to every time series, splitting each into a seasonal cycle, a long-term trend and a residual component. The residuals, representing interannual anomalies, became the focus of the analysis, while the trend component captures longer-term effects such as carbon dioxide fertilization. Grid-cell-specific vegetation response models were then built using automated machine learning, specifically the Fast and Lightweight AutoML algorithm, which searches for optimal models and hyperparameters for each grid cell. Each model was trained on data from a grid cell and its eight surrounding neighbours to expand the training sample, with 80 percent of the data used for training and 20 percent for testing. Grid cells whose final models achieved a testing R-squared below 0.2 were removed as a quality control, and performance was generally robust across global vegetated areas, with a mean testing R-squared of 0.54 for the GIMMS NDVI analysis.

The interpretability step came from explainable artificial intelligence. Using SHAP, a game-theoretic approach that attributes model output to input features by calculating Shapley values, the researchers quantified how each environmental driver contributed to predicted vegetation greenness variation. Relative importance was computed as the mean absolute SHAP value for each driver, normalized by the sum across all four drivers, and a dominance metric was defined as the difference between the relative importance of VPD and that of soil moisture. Positive values indicate VPD-dominated systems and negative values indicate soil-moisture-dominated ecosystems, with the magnitude reflecting the strength of dominance. Statistical significance of dominance and of temporal changes was assessed at each grid cell using two-sided t-tests on absolute SHAP values. The AutoML search was deliberately constrained to two tree-based ensemble methods compatible with exact SHAP computation, and robustness checks on 400 globally representative grid cells across six model configurations showed strong agreement, with correlations of 0.96 to 0.99, confirming that the results do not depend on algorithm choice or model complexity.

The headline finding is that soil moisture dominates vegetation variability in most regions, but that this dominance is eroding. Dividing the 1982 to 2022 record into two epochs, pre-2001 and post-2001, the researchers computed the relative change in each driver’s importance. VPD influence has strengthened over time while soil moisture influence has weakened, producing an asymmetric transition in which more soil-moisture-dominated ecosystems are shifting towards increasing VPD control than the other way around. The authors attribute this transition to changes in the mean and variability of soil moisture and VPD, which alter plant water-stress responses. Continued warming will probably intensify the pattern, they conclude, as rising VPD and more frequent extreme events strengthen atmospheric constraints on vegetation.

To explain why these shifts occur where they do, the team ran a second attribution analysis, using the dominance metrics and their temporal changes as targets and water-related environmental conditions as predictors: climatological means of soil moisture and VPD, their interannual variabilities expressed as coefficients of variation, and, for the temporal analyses, the changes in those means and variabilities between the two epochs. Sensitivity analyses that progressively added temperature and radiation variables confirmed that water-related variables remained the primary predictors, with consistent SHAP dependence patterns, while energy-related variables showed non-monotonic or inconsistent patterns. The water-focused models achieved testing R-squared values of 0.55 to 0.66 across the four target metrics. Regional patterns were aggregated using the 43 vegetated IPCC climate reference regions, and vegetation types were classified from the MODIS Land Cover Type product into nine functional groups, from needleleaf and broadleaf forests through shrublands, savannas, grasslands and croplands, with water bodies, wetlands, urban areas, snow and ice, and barren lands excluded.

The study carries caveats worth weighing. The dominant land cover class at 0.5-degree resolution may mask within-grid variation, the path-dependent SHAP approach reflects how the trained model partitions importance among correlated predictors, and how acclimation may have modulated driver sensitivity over time was not separately decomposed. Even so, the message is stark. The air above the world’s forests, grasslands and croplands is becoming a stronger governor of plant life than the water beneath them, and if warming continues on its current trajectory, the atmosphere’s claim on vegetation will only tighten. All datasets and code supporting the analysis are publicly available, the authors report, offering other researchers a transparent basis for probing an increasingly atmospheric future for the planet’s green cover.

Subject of Research: Temporal shifts in the relative influence of soil moisture and atmospheric dryness on global vegetation variability

Article Title: Vegetation increasingly controlled by atmospheric dryness rather than soil moisture

Article References: Li, L., Fang, Y., Mao, J., Xu, S., Zheng, Z., Li, W., Shi, M., Koven, C. D., Chambers, J., Leung, L. R., & McDowell, N. (2026). Vegetation increasingly controlled by atmospheric dryness rather than soil moisture. Nature Geoscience. https://doi.org/10.1038/s41561-026-02107-w

Image Credits: AI Generated

DOI: 10.1038/s41561-026-02107-w

Keywords: vapour pressure deficit, soil moisture, vegetation variability, NDVI, satellite observations, machine learning, SHAP, hydroclimatic drivers, climate warming, ecosystem–climate feedbacks, water stress, Nature Geoscience