A comparative study of the performance of single neural network vs. Adaboost algorithm based combination of multiple neural networks for mineral resource estimation

Nome placer offshore gold grade estimation is a challenging problem in light of extreme ore heterogeneity. However, this type of phenomenon is quite common in placer ore grade variation. Further, the characterization of spatial ore grade variability and subsequent ore grade modelling of the deposit becomes problematic due to the unavailability of adequate samples. Being located in the Arctic, the sampling is cumbersome and expensive. As a result, ore samples are taken at wide intervals. Therefore, any estimate of ore reserve might be associated with a low level of confidence. Recognizing this inevitability, ore reserve estimation should be carried out using a suitable technique Geostatistics and neural networks are the two most popular techniques used for ore grade estimation. The efficacy of one method over the other has not been clearly established. In some applications, the superiority of neural networks over the ordinary kriging technique has been demonstrated (Samanta et al. 2003 b). On the other hand, Samanta et al. (2003 a), showed no improvement of neural network techniques over traditional geostatistical techniques. Use of the neural network for ore grade estimation is still an open area of research. Therefore, the emphasis of this study is put on the detailed investigation of the neural network in placer gold grade estimation. For this purpose, not only was a single neural network model used, but also the possibility of improved neural network performance using an ensemble of neural network models through the Adaboost algorithm was examined.