Comparative Analysis Of Neural-Network And Traditional Meta-Modelings For Placer Gold Oregrade Spatial Variability Estimation

Traditional geostatistical methods have been used in ore reserve estimation for decades. Neural Networks (NNs) have provided an alternative approach to data analysis and ore reserve estimation. Unlike the traditional geostatistical methods, neural network uses a highly non-linear complex algorithm to do the estimation. Although Neural Networks have been demonstrated to be attractive alternatives as with every new techniques, there are questions which remain unanswered. This paper presents a comparative analysis with different types of metamodeling approaches -- Triangular, Kriging, Spline, and Neural Network. The neural-network is organized in a multiple-layers and Back-propagation model is used to handle the non-linearity and hidden slabs for smoothing the predicted results. With multiple dimensional and very noisy drill hole sample data, the ore grade is predicted and the overall performance is validated by the analysis of R-squared (R2), Root-Mean-Squared (RMS), and contours of ore grade spatial variability. The ore grade and tonnage of estimated outputs are also compared with various cut-off grades.