Comparative assessment of machine learning algorithms and Ordinary Kriging for predictive mapping of soil clay content in a tropical Guinea Savanna

Abstract

Clay content is an important determinant of soil physical and chemical behaviour, influencing water retention, nutrient availability, and agricultural productivity. Accurate spatial prediction of clay distribution is therefore important for digital soil mapping and sustainable land management. This study compared Random Forest (RF), Cubist, Gradient Boosting Machine (GBM), and Ordinary Kriging (OK) for predictive mapping of clay content in the tropical Guinea Savanna of Kaduna State, Nigeria. Legacy soil profile data collected between 2005 and 2024 (n = 156) were integrated with terrain, climatic, and Sentinel-2-derived environmental covariates generated in Google Earth Engine. Model performance was evaluated using R2, root mean square error (RMSE), mean absolute error (MAE), and computational time, while SHAP analysis was used to assess predictor importance. Clay content varied substantially (2.3–71.0%), with the ML models predicting higher values mainly along drainage corridors and lower landscape positions. Mean annual precipitation, elevation, Topographic Position Index, EVI, and slope showed the highest mean SHAP importance, although their relative contributions varied among models. Cubist achieved the highest validation accuracy (R2 = 0.67, RMSE = 8.32%, MAE = 6.37%) and shortest execution time (0.14 s), followed by RF (R2 = 0.56) and GBM (R2 = 0.45), while OK recorded R2 = 0.08 and RMSE = 12.44%. Under the configurations tested, environmental-covariate-driven ML models provided more accurate predictions than OK. Cubist was the best-performing model among the evaluated configurations, demonstrating the potential of ML-based digital soil mapping for clay characterization and soil resource management in data-scarce tropical environments.