Extended Data Fig. 6: Uncertainty of resilience change after land cover changes and fires.

Same as Fig. 4 in the main text, except that results are from an alternative version of the Bayesian dynamic linear model with intervention at the time of climate-driven land cover change (LCC) or fire (Methods). (a) The fraction of resilience lower than the baseline (F(R-)) between five years before and after LCC, grouped by LCC types, that is evergreen forest loss (EF.loss), deciduous forest loss (DF.loss), shrub loss (SHB.loss), deciduous forest gain (DF.gain), evergreen forest gain (EF.gain), herbaceous gain (HB.gain) and shrub gain (SHB.gain). An F(R-) > 0.5 indicates that most pixels in this group experienced reduced resilience. (b) The latitudinal variation of F(R-) grouped by a bin size of 0.75° for each LCC type. (c) F(R-) before and after fires grouped by the pre-fire land cover types and (d) the corresponding latitudinal variation, similar to (a) and (b). The colours represent five years before (-5) and after (+5) the land cover change or fire, and so forth. F(R-) was calculated by comparing resilience (posterior mean) in the target year to the temporal baseline, that is, the resilience at the same location averaged between 2003 to five years before changes. The bar height in (a) and (c) is the mean F(R-) across 100 sets of bootstrapping pixels for each group (n = 10,000 for each set). The thick black vertical line shows the standard deviation, suggesting robust estimates across sampled pixels. The lower/upper end of the thin grey vertical line is F(R-) quantified by comparing the upper/lower boundary of resilience (posterior mean plus/minus posterior standard deviation) to the abovementioned baseline, indicating large posterior range of resilience estimates post changes. The lines and shaded bands in (b) and (d) show the mean and standard deviation of F(R-) from 100 bootstrap resampling.