Extended Data Fig. 2: Random 10-fold cross-validation (RCV) of the spatial root-mass fraction models and spatial autocorrelation of model residuals.
From: The global distribution and environmental drivers of aboveground versus belowground plant biomass
a–c, Heat plots showing the relationships between predicted and observed RMFs in forests (a), shrublands (b), and grasslands (c) based on RCV. Solid lines indicate fitted relationships based on ordinary least squares regression [coefficient of determination values relative to the 1:1 line (equation 2) shown in the bottom right corner], dashed diagonal lines indicate a 1:1 relationship between observed and predicted points. d–f, The standard errors of the observed (black) and predicted (grey) mean values of root mass fractions decrease with increasing sample size. The operation was repeated with 1,000 random seeds for the observed and predicted mean values, and the calculated standard errors of the mean are shown. Note, ‘sample size’ in D–F refers to the number of pixels, and thus denotes square kilometres. g–i, Semivariograms illustrating spatial autocorrelation of model residuals in forests (g), shrublands (h) and grasslands (i). Semivariances of residuals were computed based on random 10-fold cross validation (blue) and spatial leave-one-out cross-validation (LOO-CV) with buffer radii of 150km (dark green), 250km (green) and 500km (light green). Dashed vertical lines indicate the buffer radii of the final validation model reported throughout the text.
