Tight Integration of Decision Forests into Geostatistical Modelling

This presentation will develop a novel approach to geostatistical modelling based on Conditional Random Fields. Instead of starting with a full multigaussian prior, the conditional distribution are estimated at each target location where results are required. This step does not require the specification of a full Random Function. The conditional distributions are estimated with a Decision Forest approach which is known to converge to the true conditional distribution in quite general conditions. A method for simulating realizations by correlated sampling from the distributions is shown. The advantages of this new approach are that it provides good quality estimates of uncertainty, allows the use of many secondary variables and does not require strong models of stationarity either for the target variable or for the relationship between target and secondary variables. The results of the method are compared with the classic algorithm.