Comparing machine learning-derived global estimates of soil respiration and its components with those from terrestrial ecosystem models

The CO2 efflux from soil (soil respiration (SR)) is one of the largest fluxes in the global carbon (C) cycle and its response to climate change could strongly influence future atmospheric CO2 concentrations. Still, a large divergence of global SR estimates and its autotrophic (AR) and heterotrophic (HR) components exists among process based terrestrial ecosystem models. Therefore, alternatively derived global benchmark values are warranted for constraining the various ecosystem model output. In this study, we developed models based on the global soil respiration database (version 5.0), using the random forest (RF) method to generate the global benchmark distribution of total SR and its components. Benchmark values were then compared with the output of ten different global terrestrial ecosystem models. Our observationally derived global mean annual benchmark rates were 85.5 ± 40.4 (SD) Pg C yr-1 for SR, 50.3 ± 25.0 (SD) Pg C yr-1 for HR and 35.2 Pg C yr-1 for AR during 1982-2012, respectively. Evaluating against the observations, the RF models showed better performance in both of SR and HR simulations than all investigated terrestrial ecosystem models. Large divergences in simulating SR and its components were observed among the terrestrial ecosystem models. The estimated global SR and HR by the ecosystem models ranged from 61.4 to 91.7 Pg C yr-1 and 39.8 to 61.7 Pg C yr-1, respectively. The most discrepancy lays in the estimation of AR, the difference (12.0-42.3 Pg C yr-1) of estimates among the ecosystem models was up to 3.5 times. The contribution of AR to SR highly varied among the ecosystem models ranging from 18% to 48%, which differed with the estimate by RF (41%). This study generated global SR and its components (HR and AR) fluxes, which are useful benchmarks to constrain the performance of terrestrial ecosystem models.