A fundamental concept of geostatistics that is often overlooked or misunderstood in resource estimation during the early project stage is the relationship between the global grade tonnage curve and the quality of the local block by block estimate.
Kriging, with a properly optimised set of search parameters, will always give the best local block estimate with the data available. That means misclassification errors resulting from local block selection compared to reality are minimised.
The trade-off is that the less data you have, the higher the so-called smoothing effect on the local blocks becomes. Consequently, the overall tonnages and grades estimated at higher cut-offs can still be significantly different from the actual tonnages
and grades at those cut-offs. Grades at higher cut-offs are typically underestimated and tonnages overestimated.
Reducing the smoothing by introducing increased variability into the block estimation in an attempt to match the actual block variance is an alternative. This produces local block estimates that are worse than those of optimised kriging. Misclassification of ore and waste would be more extensive if mining were to take place with selection based on those blocks. However, on a global grade and tonnage basis, the overall nonselective grades and tonnages at higher cut-offs are closer to reality, if the smoothing is reduced by some appropriate method.
It is therefore critical to understand the purpose of early stage resource estimating and to understand any differences between the best local block estimate and the theoretical global grade tonnage curves.
At the early stage of a project the accuracy of the local block estimate is often not critical but the overall tonnage and grade at a specific cut-off are critical. The differences between the best local estimate and the theoretical global estimate for an early stage project are most pronounced with commodities with highly skewed distributions, such as gold.
In early stage projects, techniques for obtaining a global estimate that is closer to reality, compared to optimal kriging, include global change of support, sub-optimal kriging (reduced sample search), uniform conditioning and simulation.
