Reduced resilience of terrestrial ecosystems locally is not reflected on a global scale

The study addresses how global climate change affects the resilience of terrestrial vegetation from local to global scales. Resilience is considered in the ecological sense of a system’s capacity to tolerate changing drivers or disturbances without shifting to an alternative state. Direct measurement of resilience is difficult, so the authors use statistical early warning indicators (EWIs) based on critical slowing down. Prior work has emphasized climate variability or mean state effects on ecosystem functioning and resilience, but the relative importance of temperature vs precipitation and mean state vs variability remains underexplored. Moreover, local changes in resilience may not scale to global patterns due to spatial synchrony/asynchrony among regions. The authors aim to quantify trends in resilience over the last 35 years, identify climatic drivers and their forms, and examine whether spatial synchrony mediates the scaling from local to global resilience.

The paper reviews two main resilience concepts: ecological resilience (persistence without regime shift) and engineering resilience (rate of return). It focuses on ecological resilience given potential multistability in ecosystems. EWIs of critical slowing down, such as increasing autocorrelation and variance, have been used to infer approaching regime shifts, though with known caveats (false/no alarms, data length, variable choice, stochasticity). Climate change is a dominant driver of terrestrial ecosystem structure and function; previous studies have mapped local resilience/vulnerability and have linked ecosystem responses to climate variability or mean state, often in isolation. Spatial synchrony in population or ecosystem fluctuations can amplify variability at regional scales, raising concern over synchronized regime shifts. Conversely, spatial asynchrony can buffer variability at larger scales. This context motivates analyzing resilience across scales and the role of spatial synchrony, alongside disentangling temperature vs precipitation and mean vs variability effects.

Data: Monthly NDVI from GIMMS NDVI3g (8 km, 0.083°, July 1981–Dec 2015; 414 images; ~1,822,400 pixels per image). Preprocessing included maximum value compositing, removal of water/ice (>40% null), exclusion of barren sites, and linear interpolation for remaining missing values (<40%). Eight WWF terrestrial biomes were analyzed: Tropical/Subtropical Moist Broadleaf Forests, Temperate Broadleaf and Mixed Forests, Boreal Forests/Taiga, Tropical/Subtropical Grasslands/Savannas/Shrublands, Temperate Grasslands/Savannas/Shrublands, Montane Grasslands/Shrublands, Tundra, Deserts/Xeric Shrublands. Climate drivers used CRU TS v4.03 monthly temperature and precipitation, rescaled to NDVI resolution via bilinear interpolation. Resilience indicators: Seven EWIs were computed in sliding windows on linearly detrended and seasonally handled monthly NDVI: ACF1, AR1, SD, skewness, kurtosis, return rate, and density ratio. Window size was 36 months (37 for lag-based indicators) to minimize seasonality effects, yielding EWI series length 379. To form a composite early warning indicator (CEWI), four indicators (ACF1, SD, skewness, kurtosis) were selected via variance inflation factor backward selection. Each EWI was standardized (z-score) and summed per time step to obtain CEWI. Trend assessment: Kendall’s tau between CEWI and time quantified resilience trends per pixel (local scale) and for NDVI averaged across the world or within biomes (global/biome scales). Positive tau indicates increasing CEWI (decreasing resilience). Drivers: Using 36-month windows, four climatic series were constructed per pixel: mean temperature, mean precipitation, CV of temperature, CV of precipitation. Variation partitioning quantified the relative importance of temperature vs precipitation and mean state vs variability to CEWI changes; relative importance indices were derived. Multiple linear regression estimated directions and strengths of each climatic factor’s effect and contributions to CEWI change. Scaling and spatial asynchrony: To explain differences between local and global trends, spatial asynchrony phi was computed using a covariance-matrix-based framework relating local instability (CV_L), global instability (CV_G), and phi = 1 - (CV_L^2 / CV_G^2), applied globally and by biome over time. Standardized multiple regression related vegetation asynchrony to climatic means, variability, and climate asynchrony in temperature and precipitation. Analyses were conducted in R using rgdal and earlywarnings packages. Source data and code are publicly available.

• Local-scale resilience decline: CEWI increased (indicating decreased resilience) in 64.5% of vegetated land pixels. Increases were strongest at high northern latitudes, with notable hotspots in Northeastern America, Eastern Siberia, Europe, and Australia. Across all biomes, median Kendall’s tau of CEWI was significantly > 0 (p < 0.001), with tundra showing the largest values, followed by deserts/xeric shrublands, temperate grasslands, and montane grasslands.
• Climatic drivers: Temperature changes explained more variation in CEWI than precipitation in most biomes (p < 0.001), except boreal forests/taiga and temperate grasslands where precipitation was comparatively more important. The influence of the climatic mean state (mean temperature and mean precipitation) exceeded that of climatic variability (CVs) in driving resilience changes. Biomes most sensitive to mean-state changes included tropical/subtropical moist broadleaf forests, montane grasslands, and deserts/xeric shrublands.
• Directional concordance: Over half of terrestrial areas (>50%) showed CEWI trends aligned with climate mean trends. Specifically, 55.8% of vegetated regions exhibited both warming and increasing CEWI (reduced resilience).
• Global/biome-scale resilience: At the global scale, CEWI decreased significantly over time (r = -0.137, p < 0.001), indicating no evidence of reduced resilience globally despite widespread local declines. Similar decreasing CEWI trends (resilience not declining) were observed for tropical/subtropical moist broadleaf forests and montane grasslands (both p < 0.001). In contrast, temperate broadleaf/mixed forests, temperate grasslands, and tundra exhibited positive CEWI trends at biome scale.
• Spatial asynchrony: Global CV_G (instability across the world) decreased while local CV_L increased over recent decades, consistent with opposite CEWI trends at local vs global scales. Spatial asynchrony of NDVI increased globally and in several biomes (tropical/subtropical moist broadleaf forests, temperate grasslands, montane grasslands), buffering global variability and offsetting local resilience losses. Temperate broadleaf/mixed forests and tundra showed decreased asynchrony.
• Climate influence on asynchrony: Spatial asynchrony of temperature decreased over time and was positively correlated with vegetation asynchrony. Standardized regressions indicated that changes in mean temperature were the dominant driver increasing vegetation asynchrony at global and several biome scales. Table 1 reports significant positive standardized coefficients for temperature mean and variability in multiple biomes and globally (e.g., World: TEM mean 0.959***, TEM CV 0.270***).

The results demonstrate scale-dependent responses of terrestrial ecosystem resilience to climate change. Locally, increasing CEWI and positive Kendall’s tau across most vegetated pixels indicate reduced resilience, particularly in high-latitude tundra, deserts/xeric shrublands, temperate and montane grasslands—regions known to be climate-sensitive. Temperature generally exerts a stronger control than precipitation, and changes in mean climate state dominate over variability, pointing to gradual warming and altered hydroclimate as primary pressures on resilience. When aggregated globally, however, CEWI decreased, indicating that global-scale resilience did not decline over the study period. This apparent paradox is explained by rising spatial asynchrony in vegetation dynamics: asynchronous fluctuations among regions generate compensatory dynamics that stabilize aggregate productivity, reducing global variability despite local instability increases. The analysis links vegetation asynchrony to climate drivers, especially mean temperature, implying spatially heterogeneous ecosystem responses to warming that increase asynchrony. This spatial insurance effect aligns with theory on stability scaling and has parallels in global crop production stabilization. Biome-specific nuances emerge: boreal forests/taiga and temperate grasslands are more precipitation-sensitive; tropical forests may be buffered by high diversity; and shrublands, often in transitional states, are more vulnerable. The findings underscore the need to consider spatial scale and synchrony in resilience assessments and conservation planning, as synchronized regional regime shifts could still pose global risks despite current asynchrony-driven buffering.

The study provides a high-resolution, multi-decade assessment of terrestrial vegetation resilience using composite early warning indicators from satellite NDVI. It reveals widespread local declines in resilience driven primarily by temperature and mean-state climate changes, while finding no concurrent global-scale resilience decline due to increasing spatial asynchrony that buffers global variability. These insights clarify the scale dependence of resilience and highlight spatial asynchrony as a key mechanism stabilizing the Earth’s vegetated systems at large scales. Management implications include prioritizing vulnerable regions (e.g., tundra, deserts/xeric shrublands, temperate and montane grasslands) and focusing on factors affecting mean climate state and water availability. Future research should investigate thresholds and positive feedback loops underlying regime shifts, interactions among multiple anthropogenic pressures (e.g., land-use change, biotic homogenization), and how globalization and species invasions may reduce beta diversity, diminish spatial asynchrony, and erode global resilience.

• Early warning indicators are indirect proxies of resilience and may produce false or missed alarms depending on data length, observation/process errors, variable choice, and stochastic properties. They do not always imply bistability or forecast exact tipping points.
• NDVI saturation and sensitivity differences across vegetation densities motivated use of rank correlations, but residual measurement biases may remain.
• The composite indicator selection, sliding window size, and detrending choices influence results, though standard practices were followed.
• Climate drivers were limited to temperature and precipitation means and variability; other factors (e.g., CO2 fertilization, radiation, soil moisture, disturbances, land-use change) and human pressures were not explicitly modeled.
• The study does not define thresholds for planetary-scale shifts; conclusions pertain to trends over the historical period analyzed (1981–2015).