Mapping soil organic carbon with Soil Health Card observations and multisensor satellite data in Odisha, India

Soil organic carbon contributes to soil fertility and terrestrial carbon storage, making its spatial distribution important for soil assessment. We used 13,284 organic carbon measurements from India’s Soil Health Card programme and 17 optical, radar, terrain and climate covariates to map organic carbon in the upper 15 cm across Odisha, India, using a random forest on a nominal 30 m grid. A reserved 20% subset gave R² = 0.501 and a root mean square error of 0.137 percentage points. In separate five-fold assessments, R² declined from 0.468 with random folds to 0.346 when observations within 30 m were grouped, and to −0.103 with geographic-cluster folds. Coverage of nominal 90% prediction intervals declined from 90.14% under random validation to 81.96% under geographic-cluster validation. The area of applicability, defined using covariate dissimilarity, included 97.6% of mapped cropland and 68.1% of the full extent. Cropland within this domain covered 5.32 million ha and yielded upper-15-cm stock estimates of 58.83 and 77.66 Tg C under two Walkley–Black recovery assumptions. The results distinguish covariate representation from geographic predictive performance when interpreting soil organic carbon maps and stock estimates.