Bayesian spatial prediction of soil organic carbon stocks in eastern DRC using INLA-SPDE and environmental covariates

Abstract

Soil organic carbon (SOC) plays a critical role in climate mitigation and agricultural sustainability, yet its spatial distribution in the eastern Democratic Republic of the Congo (DRC) remains poorly quantified. This study employs a Bayesian spatial modeling framework, Integrated Nested Laplace Approximation with Stochastic Partial Differential Equations (INLA-SPDE), to predict SOC stocks across Kalehe and Kabare territories, integrating 177 field observations with environmental covariates (soil properties, topography, and vegetation indices). The INLA-SPDE approach was chosen for its ability to handle sparse datasets effectively while providing robust uncertainty quantification, a key advantage for regions with limited observational data. Key drivers of SOC variability included soil pH, sand,clay content, bulk density, elevation, and vegetation indices (Normalized Difference Vegetation Index(NDVI), Soil-Adjusted Vegetation index (SAVI)). The INLA-SPDE model outperformed the global SoilGrids250m dataset, achieving a significantly higher correlation with observed data ($r=0.49\text{vs}0.045,p<0.001$). Higher SOC stocks were predicted in forested southern regions ($105.11\pm 11.36$ MgC ha⁻¹), while data-sparse northern areas exhibited greater uncertainty, with a posterior standard deviation of up to $32.68\pm 5.46$ MgC ha⁻¹, (vs. the spatial field’s global standard deviation averaged to 12.74 MgC ha⁻¹ at the 97.5% quantile). Posterior distributions revealed significant spatial heterogeneity, linked to land use and observational density. Our results underscore the importance of localized SOC mapping for informed land management and climate resilience strategies in tropical Africa, demonstrating the INLA-SPDE framework’s superior predictive accuracy and interpretability in data-scarce environments.