Enhancing the reliability of baseline C stock digital inventory of Sri Lanka through a dense sampling
Abstract. Accurate national inventories of soil organic carbon (SOC) are essential for quantifying terrestrial carbon storage and supporting land-management and climate-mitigation strategies. However, estimates are often constrained by sparse SOC measurements and global products that inadequately represent regional environmental heterogeneity. Here, we developed a high-resolution (30 x 30 m inventory of surface SOC stocks (0–30 cm) across Sri Lanka using a SOC data obtained using a spatially balanced network of 1,488 soil sampling locations, environmental covariates, and machine-learning-based digital soil mapping. Random Forest provided the best predictive performance, explaining 85 % of the training variability (Coefficient of determination (R²) = 0.85; Root mean square error (RMSE) = 1.39 kg C m-2, Lin’s concordance correlation coefficient (CCC) = 0.90) and retaining prediction accuracy when evaluated on an independent validation dataset (R² = 0.51; RMSE = 2.36 kg C m-2; CCC = 0.67). SOC stocks were highly heterogeneous, with a median of 5.4 kg C m-2 and an estimated national stock of 382.4 Tg C. The largest stocks occurred in the central highlands, whereas warmer and drier lowlands stored substantially less carbon. Shapley additive explanations (SHAP) analysis identified aridity index, mean annual precipitation, potential evapotranspiration, mean temperature of warmest and driest quarters, and solar radiation as dominant controls. These drivers operated through contrasting pathways: cooler temperatures constrained decomposition in the highlands, while moisture availability regulated carbon inputs across the lowlands. Tea and rubber plantations supported comparatively high SOC stocks. Benchmarking showed that the updated map substantially outperformed the previous Sri Lankan inventory developed in 2019 using sparse soil SOC data, SoilGrids v2.0, and GSOCmap v1.5.0. Our results demonstrate that dense, representative sampling combined with machine learning can improve national SOC inventories and provide a robust baseline for monitoring soil-carbon change and assessing climate vulnerability in heterogeneous tropical landscapes.