by Didier Renard » Thu Aug 03, 2023 5:55 pm
Hi Jeffrey
Sorry for this late answer. And sorry for this NEGATIVE answer.
Let us recall the main principle of the sequential (conditional) simulation... In turn, we consider a new target. We calculate the estimation at the target site, based on whatever information is available at the current time (i.e. the initial hard data + the values at the target sites processed before the current one). This estimation, to which we add a random term (obtained sampled from a normal distribution but tuned to the variance of the estimation error), produce the conditional simulation at the current target site.
This new simulated value will be part of the available information when processing the next target site.
In RGeostats (and still in future gstlearn), we have not been willing to add the Sequential simulation technique. Essentially for two reasons:
- a theoretical one: Unless we always consider ALL the previously simulated targets in the kriging procedure, the results of such a procedure is unsafe (unless we use specific models and pattern of information where screen effect is valid). Therefore the available information grows with the rank of the target site. This quickly leads to an issue as the kriging system becomes huge and inverting it becomes intractable. Then people usually introduce a problem reduction by introducing a *neighborhood*. The impact of this neighborhood on the quality of the result has not been seriouly analyzed (for my point of view).
- a practical one: at each sequence, the information comes on one hand from the initial data, and from the already simulated targets on the other hand. A proper implementation would require handling a double neighborhood ... which is not available in RGeostats (nor in gstlearn).
Sorry.