Comparison of geostatistical and machine learning models for predicting geochemical concentration of iron: case of the Nkout iron deposit (south Cameroon)

This paper investigated a comparative study between geostatistical methods and machine learning techniques in order to predict geochemical concentration of iron (Fe) in the Nkout iron deposit (South Cameroon). To this end, geostatistical methods and machine learning techniques are used. The geostatistical methods used included statistical analysis, variogram analysis, ordinary kriging (OK) and Turning Bands Simulations (TBS). The machine learning techniques employed included k-Nearest Neighbour (k-NN) and Random Forest (RF). Cross-validation is applied to determine the best method. The prediction maps of iron content from a drilling campaign carried out in southern Cameroon are then produced using each method. The results obtained show that machine learning methods give better results in predicting content. The Random Forest offers better validation than ordinary kriging, turning bands simulations and k-NN. Ordinary kriging and turning bands simulations maximized the base minimum from 2.18% to 16.37% and 18.33%, respectively. This study identifies the Random Forest as a good first choice algorithm for the prediction of geochemical concentration of Fe. This technique is simple to formulate, efficient from a computational point of view, very stable with regard to the variations of the values of the parameters of the prediction model. •Determination of the limits of geostatistical prediction and simulation models.•Use of machine learning in solving geoscience problems.•Construction of geostatistical prediction (ordinary kriging and turning bands simulation) and machine learning models (k-NN and Random Forest) for iron concentration.•Choice of optimal prediction model.