A novel artificial intelligence approach for regolith geochemical grade prediction using multivariate adaptive regression splines

Abstract

The necessity for applying a potent analytical regolith geochemical grade estimator is driven by the reality of mineral exploration. This is because many exploration geologists rely upon the classical geostatistical method of Kriging which oftentimes do not produce accurate predictions due to the complexity of interactions between geological features and spatial variables. In this study, a novel non-linear data-driven approach known as Multivariate Adaptive Regression Spline (MARS) is proposed as an effective predictive tool to unravel regolith geochemical complexities. The proposed MARS approach was used to predict regolith geochemical grade from a thick regolith cover in the Tarkwaian paleo-placer of the South-Western Ashanti belt in Ghana. Out of the 891 samples, the data was partitioned into 70% training (model development) and 30% testing (model validation). The proposed MARS result was compared with three other artificial intelligence techniques (i.e., radial basis function neural network, backpropagation neural network and generalised regression neural network) and kriging geostatistical technique. Based on the test results, MARS had the highest correlation coefficient ( R = 0.9675) and the least statistical error metrics (RMSE = 0.7791, MAE = 0.6014, and ρ = 0.0472). The implementation of the MARS approach in regolith geochemical grade estimation domain has yielded outstanding and promising results. The MARS superiority was evident in its calibration strength, prediction accuracy, robust interaction of variables and overcoming the black box system of ANN. Thus, the proposed MARS approach could be an excellent tool in regolith geochemical grade estimation workflow when fully integrated with exploration tasks.