Emerging signals of declining forest resilience under climate change

The study investigates how global forest resilience has evolved under recent climate change. Resilience is defined as the capacity to withstand and recover from perturbations and avoid state shifts. Forests are critical for carbon uptake and ecosystem services but face increasing natural and anthropogenic disturbances. Theory predicts that as systems approach tipping points, resilience declines and recovery slows (critical slowing down, CSD), detectable via increased temporal autocorrelation (TAC) in ecosystem state variables. Despite its importance for conservation and climate mitigation planning, the temporal evolution of forest resilience at global scales remains poorly quantified due to observational and methodological challenges. The authors aim to quantify recent trends in forest resilience globally, identify their drivers, and assess links to abrupt declines in forest productivity.

Prior work established generic indicators of ecological resilience and early warning signals for critical transitions based on increases in temporal autocorrelation and variance as systems approach thresholds. Studies mapped static patterns of forest resilience and ecosystem sensitivity to climate variability, and documented increasing disturbances and tree mortality under climate change across regions. However, dynamic, global-scale assessments of changing resilience have been limited by short observational records, strong seasonality, varying autocorrelation in climate drivers, and noise. Recent advances in Earth observation time series and robust vegetation indices (e.g., kNDVI) offer opportunities to monitor time-varying resilience. The study builds on this literature by applying CSD indicators to satellite observations in a machine-learning framework to analyze spatiotemporal dynamics and drivers of forest resilience.

• Data: Global satellite-based kernel NDVI (kNDVI) from MODIS at 0.05° spatial resolution for 2000–2020, used as a proxy for forest ecosystem state/productivity.
• Resilience metric: Computed lag-1 temporal autocorrelation (TAC) of kNDVI as a critical slowing down indicator. Two uses: (1) long-term TAC from the full 2000–2020 series; (2) annual TAC estimated with 3-year rolling windows to capture temporal dynamics. The temporal trend in TAC (denoted STAC) indicates changes in resilience (increasing TAC implies declining resilience).
• Driver modeling: Trained a Random Forest regression to relate long-term TAC to environmental predictors (forest and climate metrics). Model achieved high explanatory power (R^2 = 0.87). Performed factorial simulations with the RF model to separate contributions of changing background climate, climate variability, and forest structure (density/greening), and to filter confounding autocorrelation originating from climate drivers.
• Trend assessment: Analyzed STAC globally and by biome (tropical, arid, temperate, boreal). Also compared TAC between two independent decades (2000–2010 vs 2011–2020) to corroborate trends.
• Management effects: Compared long-term TAC and STAC between managed and intact forests within similar background climates to isolate the effect of management versus climate drivers.
• Abrupt declines (ADs): To link resilience loss to ecosystem shifts, identified ADs in intact forests as negative anomalies of growing-season kNDVI exceeding 1–6 standard deviations below an undisturbed mean. Assessed the probability of AD conditioned on increases in TAC and examined spatial patterns (with emphasis on tropical and boreal regions). Retrieved TAC in the year preceding AD (TAC_AD) to infer threshold values and ecosystem tolerance (difference between TAC_AD and average TAC).

• Global trends: Widespread and significant increases in TAC (declines in resilience) in tropical, temperate, and arid forests. Mean STAC rates: tropical 1.63×10^-3 yr^-1, temperate 1.43×10^-3 yr^-1, arid 1.26×10^-3 yr^-1. Boreal forests show heterogeneous local patterns but on average decreasing TAC (increasing resilience) at −1.54×10^-3 yr^-1, with notable TAC declines in Eastern Canada and European Russia.
• Decadal comparison: From 2000–2010 to 2011–2020, 53% of global forested areas show increased TAC; 56–63% of tropical, arid, and temperate forests show increased TAC, while 56% of boreal forests show decreased TAC.
• Drivers: Greening (increased forest density), likely from CO2 fertilization and warming, enhanced resilience especially in cold and temperate climates. However, intensified water limitations and increased climate variability/extremes, particularly in tropical, arid, and temperate regions, outweighed greening benefits, yielding net resilience declines. In boreal regions, positive effects of warming and CO2 fertilization generally exceeded climate stress over the study period.
• Management: Intact forests exhibit lower long-term TAC (higher resilience) than managed forests (means 0.13 vs 0.21; distributions significantly different). Temporal trends (STAC) are similar between managed and intact forests: comparable fractions with positive STAC (decreasing resilience)—72% managed vs 66% intact—indicating large-scale climate controls dominate ongoing changes.
• Abrupt declines (ADs): In intact forests globally, the probability of AD conditional on increasing TAC exceeds 0.5 and rises with AD severity, indicating that declining resilience is linked to upsurges in negative productivity anomalies. This relationship is strongest in boreal regions (central Russia, western Canada), consistent with local drifts toward resilience thresholds potentially culminating in insect outbreaks and other disturbances. In tropical forests, ADs are not statistically associated with increases in TAC, consistent with rapid, intense disturbances (fires, droughts) triggering ADs irrespective of long-term CSD trends.
• Critical threshold status: About 23% of intact, undisturbed forests—accounting for approximately 3.32 Pg C of gross primary productivity—have already crossed critical resilience thresholds and are undergoing further resilience degradation.

The analysis demonstrates that many forests, especially in tropical, temperate, and arid biomes, exhibit increasing temporal autocorrelation in productivity, signaling reduced resilience consistent with critical slowing down theory. These trends are driven by heightened water stress and climate variability that counteract beneficial greening effects attributed to CO2 fertilization and warming. Boreal forests presently show, on average, increasing resilience, likely due to temperature limitation being alleviated and CO2 fertilization benefits, although localized declines and future water constraints could reverse this pattern. Management strongly influences baseline resilience (intact > managed), but contemporary resilience trends are similar across management classes, implicating large-scale climate as the dominant driver. The statistical linkage between resilience decline and abrupt drops in productivity, particularly in boreal regions, supports the notion that slow drifting toward thresholds increases the likelihood of sudden ecosystem deterioration. These findings underscore the need to incorporate resilience dynamics into forest management, conservation, and climate mitigation/adaptation strategies.

By integrating satellite kNDVI with machine learning and critical slowing down indicators, the study provides a global, time-resolved assessment of forest resilience from 2000–2020. It reveals widespread resilience declines in tropical, temperate, and arid forests, contrasting with regionally variable but on-average increasing resilience in boreal forests. Greening has bolstered resilience in colder climates, but intensifying water limitations and climate variability have led to net declines elsewhere. Intact forests are more resilient than managed ones, yet both exhibit similar temporal declines driven by climate. Declining resilience is statistically associated with abrupt productivity drops, and nearly a quarter of intact forests have surpassed critical thresholds. These insights should inform land-based mitigation and adaptation planning and support the application of resilience thinking to forest management under rapid climate change. Future work should continue refining early-warning metrics, extend observational records, and integrate resilience monitoring into policy and management frameworks.

• Methodological constraints inherent to CSD detection at large scales include limited time-series length, dominant seasonal cycles in ecosystem and climate signals, variations in autocorrelation of climate drivers, and stochastic noise.
• Reliance on a single satellite sensor (MODIS) and one vegetation index (kNDVI), although chosen for robustness, may still introduce sensor- or metric-specific biases.
• Attribution relies on data-driven Random Forest modeling and factorial simulations to disentangle drivers; residual confounding cannot be fully excluded.
• Abrupt decline (AD) detection is based on statistical anomalies relative to an undisturbed mean and may be influenced by disturbance type, severity thresholds, and data noise; analysis was restricted to intact forests to exclude harvest effects, potentially limiting generalizability to managed forests.
• Boreal resilience gains observed over 2000–2020 may be transient given projected increases in water stress and phenological shifts, cautioning against extrapolation.