The paper presents decision-oriented mapping of pollution using hybrid
models based on statistical learning theory (support vector regression
or SVR) and spatial statistics (geostatistics). Adaptive and robust
SVR approach is used to model non-linear large scale trends in the
region and geostatistical models -- spatial predictions and spatial
simulations -- are used to prepare decisionoriented maps: prediction
maps along with maps of error variance and equiprobable digital models
of the pollution based on conditional stochastic simulations. The
quality of the proposed approach is tested with the validation data
set not used for the model development. Real data on soil contamination
by Chernobyl radionuclides in Russia is used as a case study.