Figure 1

Maps of potential GPP for different land cover types derived from RF predictions. (a–c) Satellite-derived present-day GPP for forests (a), grasslands (b) and croplands (c) (i.e., the training data). (d–f) Potential GPP predicted by the RF algorithm. (g) Land cover with highest potential GPP according to (d–f). (h) Global fractions of the most productive land cover type. i, Potential GPP distribution across the total suitable area. R² and RMSE values are computed on the out-of-bag testing data. The good model performance can partly be explained by the very large training data and to some degree by spatial autocorrelation²⁴ (Supplementary Discussion 2 and Figs. S3 and S4). Global area-weighted GPP means are given by the numbers at the bottom of the maps. Grid cells where no forests exist today or potential forest cover (Supplementary Fig. S5) is < 36.3% (i.e., 5th percentile of all currently forested grid cells) or which are too cold or dry for grass/crop growth are removed from (d–f) and removed from (g) if unsuitable for at least one land cover type. Dots in (i) indicate area-weighted means. Maps were created using R version 4.1.0 (https://cran.r-project.org/)²⁵.