Modeling the spatial distribution of mineral deposits using neural networks

Abstract

In the context of Geographic Information Systems (GIS), mineral potential mapping can be seen as a process whereby a set of input maps, each representing a distinct geoscientific variable, are combined to produce a single map which ranks areas according to their potential to host deposits of a particular type. However, unlike many spatial classification tasks, the objective is not to assign a pixel to some class, but to assign to each pixel a continuous value representing its mineralization potential. Logistic regression provides one method of modeling conditional probabilities on data in which the target values are binary, e.g., deposit present/deposit absent; however, logistic regression limits the output to be a logistic transformation of the inputs and hence may not be able to capture complex relationships among the input variables. This paper reports on the application of multilayer perceptrons (MLPs) to mapping reef gold mineralization potential in the Castlemaine region of Victoria, Australia. We show that the ability of MLPs to predict the presence of hold-out test deposits is significantly better than that for logistic regression. We also show that a Bayesian learning approach can be used to automatically select an appropriate value for the weight regularization coefficient without the need of using a cross-validation procedure. This means that all available data can be used for weight optimization, thus eliminating any sensitivity of the final map to the particular training/validation set partition used. This paper was presented at the 2004 Research Modeling Association World Conference on Natural Resource Modeling in Melbourne, Australia. © 2007 Rocky Mountain Mathematics Consortium.