High resolution mapping of soil organic carbon in northern Nigerian agricultural lands using comparative machine learning with uncertainty analysis

Abstract Soil organic carbon (SOC) is a crucial component in soil quality, health, and ecosystem function. Knowledge about its distribution is paramount for sustainable agriculture, food security, soil stabilityand carbon sequestration monitoring, especially in areas like northern Nigeria, where the soils are highly heterogeneous. This study applied machine learning algorithms of random forest, extreme gradient boosting, and categorical boosting to predict and map the spatial distribution of SOC in northern Nigeria at 30 m resolution. We used 181 soil samples collected from the 0–20 cm top layer across agricultural lands; 12 environmental covariates were selected from 22 covariates to fit the models using ten-fold cross-validation. Quantile random forest was used to estimate the associated SOC uncertainties. The result revealed that Random Forest achieved higher prediction accuracy (R2 = 0.43, RMSE = 0.45%, MAE = 0.33%), outperforming XGBoost (R2 = 0.39, RMSE = 0.54%, MAE = 0.37%) and CatBoost (R2 = 0.37, RMSE = 0.49%, MAE = 0.38%). Random forest further indicates that Landsat 9 Band 5 (NIR) is the most critical predictor variable, followed by annual precipitation, elevation, and mean temperature; the predicted SOC map shows a significant variation of SOC, with higher concentration in the southern area and lower in the northern part. The uncertainty estimation shows a broader range of uncertainties in the highly elevated areas. The study suggests that further observations and denser sampling should be carried out, especially in highly elevated areas, to capture more spatial variability.