Performance of clustering for the decision of stationarity; A case study with a nickel laterite deposit

Abstract

The decision of stationarity is a fundamental prerequisite to geostatistical estimation and uncertainty characterization of natural resources. Great effort is given to delineate relevant spatial-statistical populations for modeling. Recently developed spatial clustering methodologies claim improvements over traditional methodologies for stationary decision making. However, these novel methods are generally unproven with respect to the geostatistical algorithms that generate the models upon which important and costly decisions are based. In this study, a new method to generate the prior parameter uncertainty for the decision of stationarity is established, and the effects of considering different (and uncertain) decisions of stationarity are explored through a K-Fold cross validation study applied to a nickel laterite dataset with complex multivariate relationships.Results show that, in terms of univariate regression metrics, geostatistical models generated with clustering-based modeling domains are competitive or better than those generated with the traditional merged-lithological domains. However, models generated from the clustering-based stationary domains are superior in terms of multivariate feature reproduction when compared to the merged-lithology domains. We also demonstrate a geostatistical modeling workflow that incorporates uncertainty associated with stationary domaining, with only a minimal extension to established techniques for geostatistical modeling with parameter uncertainty.