Digital soil mapping in support of voluntary carbon market programs in agricultural land

Voluntary carbon market (VCM) programs in agriculture depend on accurate measurements of soil organic carbon (SOC) that can be deployed at scale efficiently, but barriers are preventing widespread adoption. To overcome these challenges, we developed a digital soil mapping (DSM) framework driven by machine-learning and numerous spatial covariates, including long-term climate proxies, short-term climate and weather-related variables, topographic and edaphic measurements, and remote sensing time-series summaries. We show that the model can predict SOC content in the top 30 cm of soil using 5,230 measurements of SOC in agricultural land within 47 states in the contiguous United States (CONUS). Model predictions closely matched independent measured values. The intercept and slope of the cross-validated relationship at the agricultural field level were −0.179 and 1.095. The coefficient of determination was R2 = 0.811, and the RMSE was 0.041. In contrast, comparison of independent field measurements to four publicly available SOC data products using 165 fields that contained 3,285 in-situ soil samples showed poor ability of existing public SOC maps to reproduce measured values, underscoring the importance of quantification technologies developed specifically for agricultural land and with recent soil measurements. Three prior SOC data products underestimated SOC content at small values and overestimated it at large ones, while one underestimated SOC content at all values examined. Analysis of feature importance showed that time series summaries from Sentinel-2 are the strongest predictors, followed by temperature variables and features related to surface hydrology. These findings underscore the value of geographically representative training and validation data for quantifying SOC content in agricultural land and demonstrate that feature engineering can increase the sensitivity of SOC quantification to optical remote sensing summaries. Data-driven algorithms can generate accurate estimates of field-level SOC content in agricultural land in CONUS that overcome barriers to scale in the VCM.