Extended Data Fig. 8: Resilience trend calculated using a moving window approach exhibits large uncertainties.

(a) The vegetation resilience trend calculated by replicating the method as described in ref. ⁴¹ using the EVI data same as in this work during 2000–2019 across the ABoVE core domain. Key steps include deseasonalizing and detrending the 16-day EVI time series, calculating the long-term lag-1 autocorrelation of EVI (TAC), using long-term average forest density, background climate, climate variability and autocorrelation in climate as predictors (X) of a Random Forest regression model to model TAC, calculating annual lag-1 autocorrelation of EVI (TAC^t) and predictors (X^t) within a 3-year moving window, calculating the impact of autocorrelation in climate on the lag-1 autocorrelation of EVI (TAC^t|X_ac) using the Random Forest regression model (RF(X^t)-RF(X_-ac^t, X_ac²⁰⁰⁰)), factoring out the impact of autocorrelation in climate from the annual lag-1 autocorrelation of EVI (TAC^t-TAC^t|X_ac), and computing the linear trend of the resulting enhanced annual autocorrelation of EVI. (b)-(c), The 5th percentile (b) and 95th percentile (c) of vegetation resilience trend from 100 times of pairwise bootstrapping the raw annual time series of lag-1 autocorrelation of EVI (TAC^t) with replacement. (d)-(e),The 5th percentile (d) and 95th percentile (e) of vegetation resilience trend from sampling the impact of autocorrelation in climate on the autocorrelation of EVI (TAC^t|X_ac) 100 times from a Gaussian distribution centered around the prediction of the random forest regression model with a standard deviation calculated from the regression residuals. The difference between (b) and (c) indicates a large uncertainty of resilience trend arising from estimates of annual lag-1 autocorrelation, likely caused by a large fraction of missing data in high latitude regions. The difference between (d) and (e) indicates a large uncertainty due to a much lower explicative power of the Random Forest regression model in this region (30%) compared to that in the original global scale study (87%).