Continental United States may lose 1.8 petagrams of soil organic carbon under climate change by 2100

Aims High-resolution information on soils' vulnerability to climate-induced soil organic carbon (SOC) loss can enable environmental scientists, land managers, and policy makers to develop targeted mitigation strategies. This study aims to estimate baseline and decadal changes in continental US surface SOC stocks under future emission scenarios. Location Continental United States. Time period 2014-2100. Methods We used recent SOC field observations (n = 6,213 sites), environmental factors (n = 32), and an ensemble machine learning (ML) approach to estimate baseline SOC stocks in surface soils across the continental United States at 100-m spatial resolution, and decadal changes under the projected climate scenarios of Coupled Model Intercomparison Project Phase Six (CMIP6) earth system models (ESMs). Results Baseline SOC projections from ML approaches captured more than 50% of variability in SOC observations, whereas ESMs represented only 6-16% of observed SOC variability. ML estimates showed a mean total loss of 1.8 Pg C from US surface soils under the high-emission scenario by 2100, whereas ESMs showed no significant change in SOC stocks with wide variation among ESMs. Both ML and ESM predictions agree on the direction of SOC change (net emissions or sequestration) across 46-51% of continental US land area. These differences are attributable to the high-resolution site-specific data used in the ML models compared to the relatively coarse grid represented in CMIP6 ESMs. Main conclusions Our high-resolution estimates of baseline SOC stocks, identification of key environmental controllers, and projection of SOC changes from US land cover types under future climate scenarios suggest the need for high-resolution simulations of SOC in ESMs to represent the heterogeneity of SOC. We found that the SOC change is sensitive to key soil related factors (e.g. soil drainage and soil order) that have not been historically considered as input parameters in ESMs, because currently more than 95% variability in the SOC of CMIP6 ESMs is controlled by net primary productivity, temperature, and precipitation. Using additional environmental factors to estimate the baseline SOC stocks and predict the future trajectory of SOC change can provide more accurate results.

Automated summary

This study estimates baseline and decadal changes in SOC stocks in continental US surface soils using high-resolution data and machine learning methods. The results suggest that machine learning methods outperform earth system models in predicting SOC changes, mainly due to the use of more detailed data.