Research on optimal spatial interpolation methods for soil organic carbon across spatial scales

Abstract Accurate estimation of soil organic carbon (SOC) storage is crucial for global carbon budgeting and climate policy formulation. Spatial interpolation bridges localized measurements with regional estimates, yet method performance variability across spatial scales remains poorly quantified. This study systematically evaluates four widely-used approaches (Ordinary Kriging (OK), Inverse Distance Weighting (IDW), Stepwise Linear Regression (SLR), and Random Forest (RF)) through a multi-scale experiment encompassing national (9.6×10⁶ km²), basin (1.03×10⁵ km²), catchment (2.9 km²) and field (0.3 km²) scales in China. Key findings reveal: 1) Scale-dependent accuracy shifts with 6.06-fold variation in R² (0.16–0.97), where machine learning (RF) outperforms at national/catchment scales (R²=0.37/0.97) while SLR excels in field-scale predictions (R²=0.84); 2) Geostatistical methods (OK) show comparative advantage at basin scales (RMSE = 1.051 vs 1.052–1.116 for others); 3) National-scale carbon stock estimates vary 23% (36.74–45.20 Pg C) across methods, highlighting methodological sensitivity. Spatial analysis demonstrates machine learning (RF) and regression approaches (SLR) better resolve SOC heterogeneity than geostatistical methods. We propose a hierarchical decision matrix informed by relative improvement metrics (RI = 0.08–78.05%), offering spatially-adaptive methodological recommendations to enhance precision in carbon accounting frameworks. These scale-explicit insights advance optimization strategies for SOC mapping in climate governance applications.