Extended Data Fig. 8: Scale-dependence of Southern Amazon forest responses to drought, showing that detected response patterns are largely invariant across different scales of analysis.

(a) At 0.4 degree (40-km) scale (across the Southern Amazon. all three droughts): Climate-adjusted EVI responses (standardized anomalies from MODIS) vs. water-table depths (indexed by HAND) for observations (solid points ±95% CI and solid regression line) and for unified multi-drought GAM predictions (model of Supplementary Table 1a, shaded bands and dashed regression line slopes) for the 2005 (green, slope = −0.019 ± 0.001 SE m⁻¹), 2010 (purple, slope = −0.020 ± 0.002 SE m⁻¹), and 2015 (blue, slope = −0.028 ± 0.002 SE m⁻¹) droughts (with N = 1,384, 1,673, and 1,837 0.4° pixels for 2005, 2010, and 2015 droughts, respectively); (b) At 1-km scale (across the Southern Amazon, all three droughts), as in (a): climate-adjusted EVI responses vs. HAND for observations (solid points and regression line) and corresponding GAM (with the same Supplementary Table 1a model now fit at 1 km scale, revealing autocorrelation in observations causing too-narrow confidence bands, and slight model underpredictions of the extremes of the 2005 greenup and the 2010 browdown, but maintaining the similar negative dependence on HAND across all droughts); (c) At 30 to 180 m scales (for a forest region around Manaus, 2015-2016 drought only): Delta EVI, the fraction change in EVI due to the drought = (after-drought EVI (July 2016) - pre-drought EVI (August 2015))/pre-drought EVI (Landsat OLI8, at 30 m resolution) vs. water-table depths (indexed by HAND) for Landsat observations (solid points ±95% CI and solid regression line) at native (30 m) and aggregated to 90 and 180-m scales (with N = 105,359, 11,901, and 2,999 pixels for 30-m, 90-m, and 180-m scales, respectively). Also shown in the bottom of each panel is the distribution of water-table depth (HAND proxy) at each scale. Aggregations to larger (coarser) scales induce an apparent regression towards the mean in the water-table depth distributions (as more extreme water-table depths at finer scales become diluted by averaging to large scales), while similar dilution of extremes in EVI response (not shown) preserves the overall relation between EVI responses and watertable depth (especially evident in the Landsat analysis where the slopes through data aggregated at different scales do not detectably differ).