Fig. 2: Decorrelation of land-cover type diversity from other potential drivers of productivity.

To avoid a statistical confounding of landscape diversity, here measured as land-cover type richness (LCR), with other potential drivers of productivity, we used a stochastic subsampling technique that minimized correlations between these (Methods). The panels on the left show, in the full set of landscape plots, and for each block (Fig. 1), the correlations of LCR with the altitude, the north gradient of the slope, and the fraction of a landscape plot covered with a particular land-cover (only forest shown here as an example). With such a quasi-experimental study design and by probabilistic subsampling we obtained a dataset where the correlations of the landscape’s properties with LCR were minimized. Histograms at the right of the panels show the distribution of each variable.