We first tested significance of differences of the pools, fluxesτ _e,C, τ _e,N andτ _e,P among different forest types using one-way ANOVA with the software R v.4.0.5 (www.r-project.org/). Before the analysis, we verified whether data fit the normal distribution and then logarithmic transformed the non-normal data. Original or transformed data according with normal distribution was test by LSD test (Williams & Abdi 2010), while data inconsistent with normal distribution after transformation was tested by Waerden test (Van der Waerden 1952). We then conducted linear or nonlinear regression betweenτ _e,C, τ _e,N andτ _e,P and other variables.

To quantify the contributions of 30 variables related to climate, vegetation, soil and terrain (see Table S3) to the variances of the estimated τ _e,C, τ _e,N andτ _e,P, we used variation partition method in R (“vegan” package, (Jari Oksanen et al. 2020)). We analysed contributions from the direct effect by each of the four groups of variables and eleven interactions among the four groups of variables.

To identify the most dominant variable on the variations ofτ _e,C, τ _e,N andτ _e,P , we used correlation analysis. Based on the correlation analysis, we identified T _min as the most important variable for τ _e,C,τ _e,N and τ _e,P. To identify possible threshold in the dependence of τ _e,C,τ _e,N and τ _e,P onT _min, we applied segmented regression (Muggeo 2003). We tested the significance of the change in the regression slope at the breakpoints (“Different-in-slope”) using Davies test (Davies 1987). The analysis was also done using “segmented” R-package (Muggeo 2021). Linear regression was also used to quantify the dependence ofτ _e,C, τ _e,N orτ _e,P on T _min.