An Unsupervised Clustering Approach for the Geostatistical Domaining of Univariate Data
Ortiz, Julian M.
The first formal step in geostatistical workflows consists of establishing domains and making the stationarity assumption for each domain. The validity of the stationarity assumption relies on samples within each domain to be statistically and spatially continuous. Domains are usually established by grouping data with similar categories into domains, for example a few geological units with matching properties could be grouped together into a single domain. However, in the case of a univariate data set with no categorical information, the current domaining methods are limited to either purely spatial clustering such as K-means clustering or purely statistical methods such as grade domaining. Both purely spatial and statistical domaining methods are unsuitable for establishing adequate domains due to the stationarity assumption validity. A practical workflow has been developed which balances statistical and spatial clustering for univariate data scattered in three-dimensional (3D) space. The established methodology also provides a technique for determining the optimal number of domains by splitting the data set into a training and validation subset to evaluate the weighted fit for different sets of domains. The unsupervised clustering workflow is applied to an exploratory drill hole data set from an undisclosed copper porphyry deposit which is currently an active open pit mine. The methodology is tested for the geostatistical domaining of ore grades in the context of resource estimation using the data set which is comprised of soluble copper (Cus) ore grade. The established workflow was successful and resulted in the discovery of three domains with an 8%/36%/57% split to be the optimal number of domains for the exploratory Cus data set. Although the example presented here is ore grade domaining for resource estimation, the workflow can be applied to any univariate data set scattered in 3D space.