Spatial prediction of soil carbon with deterministic and covariate-integrated geostatistical models

Soil carbon is an indicator of soil quality and it varies across landscapes. A study was conducted to assess and predict the spatial variability of total soil carbon (TC) across a heterogeneous tropical landscape comprising wetlands, midland laterites, and uplands in Kerala, India. The region, characterized by diverse physiographic and agroecological zones, encompasses various land uses, including forests, plantations, and paddy fields. TC content ranged from 0.34% to 4.46%, with higher values in forested and plantation areas and lower levels in intensively cultivated paddy fields, particularly in the Palakkad plains. The influence of biophysical drivers i.e., elevation, land surface temperature (LST), and vegetation greenness (NDVI), on TC distribution was evaluated using multiple linear regression (adjusted R² = 60.08%), highlighting significant negative correlations of LST and elevation with TC. NDVI, though positively related, was not a significant predictor, likely due to temporal variability in vegetation cover. Geostatistical analyses revealed moderate spatial autocorrelation, with the exponential model in Ordinary Kriging (OK) capturing short-range spatial structure. While deterministic methods like Inverse Distance Weighting (IDW) performed reasonably well, Empirical Bayesian Kriging Regression Prediction (EBKRP) incorporating LST and elevation achieved the best prediction accuracy (RMSE = 0.48, Lin's CCC = 0.87). Spatial predictions from EBKRP captured key landscape gradients, emphasizing the role of topography, microclimate, and land use in carbon sequestration. These findings emphasize the importance of integrating environmental covariates for robust TC mapping and suggest that EBKRP is well-suited for operational carbon assessments in diverse agroecosystems.