Soil organic carbon modelling with Multi-type environmental variables using machine learning

Soil organic carbon (SOC) spatial variability at Keren subzone, Eritrea, was modelled with good accuracies. Partial least squares model with R2 = 0.90 and RMSE = 0.08 gave the highest accuracy followed by Cubist and gradient boosting models, respectively. Land use was the most important variable for SOC prediction. Thus, the study concludes that these models have high accuracy to be used for soil fertility and productivity improvements management planning.