Temporal adjustment approach for high-resolution continental scale modeling of soil organic carbon

Open-source legacy data available for training soil organic carbon (SOC) models are limited and not uniformly distributed in space or time. While some process-based models predict SOC changes, most of the large-scale data-driven SOC modeling efforts overlook temporal shifts. Accounting for the expected temporal drift allows us to increase the accuracy of dataset available for machine learning models. Here we present an approach for creating proximity-based distance matrices using the legacy data available in contiguous US (CONUS) and generating spatially resolved temporal shift projections that adjust observations to the target date. The approach was evaluated by comparing SOC observations projected to two reference years, SOC₁₉₈₀ and SOC₂₀₂₀ and without temporal adjustment (SOC_no-adj). Stocks of SOC projections showed significant differences between SOC_no-adj and SOC₂₀₂₀. Baseline estimate of SOC stocks in CONUS croplands (top 1 m) were higher based on SOC_no-adj (14.49 Pg C) compared to SOC₂₀₂₀ (13.29 Pg C), for pasture lands 15.49 Pg (SOC_no-adj) and 14.22 Pg C (SOC₂₀₂₀), for forest lands at 39.52 Pg C (SOC_no-adj) and 40.83 Pg C (SOC₂₀₂₀). The study results confirmed the validity of our methodology, and its capability to enhance SOC stock projections effectively with temporal adjustments. Potential users of this study's outcomes include many stakeholders involved in carbon incentive programs, including farmers, scientists, policy makers, and industry partners.