Mapping Long-Term Soil Organic Carbon Stocks Across the Conterminous United States with Machine Learning.

Soil organic carbon (SOC) is a critical component of the global carbon cycle, but existing estimates of SOC stocks across the contiguous United States (CONUS) vary widely for the top 30 cm, hindering efforts to assess carbon sequestration potential and inform climate mitigation strategies. We compiled and harmonized SOC measurements from five major databases spanning 1984-2022 across diverse environmental conditions. Using a parsimonious set of environmental covariates, we compared the performance of linear regression and three machine learning algorithms for predicting SOC stocks. Random Forest demonstrated the best validation performance (R2 = 0.47, root mean squared error (RMSE) = 0.95 kg/m2) and was used to generate nationally consistent annual 30 m SOC stock maps from 1990-2022. Our analysis estimates total CONUS SOC stocks at 60.4 Pg C, with agricultural soils, forested soils, and shrubland and grassland soils representing the three dominant pools. Comparison of predicted and observed annual SOC across U.S. Department of Agriculture Economic Research Service (USDA ERS) farm resource regions showed that annual means calculated from long-term observations were more variable than model predictions. This pattern indicates that apparent annual SOC trends are strongly influenced by the spatial and temporal distribution of available observations. These results demonstrate that the maps are most useful for evaluating broad spatial patterns, regional and land cover differences, and the limits of interpreting annual SOC change from heterogeneous long-term observations. This framework provides a complementary resource to existing SOC products for carbon accounting, land use planning, biofuel life cycle assessment, and prioritizing future SOC monitoring.