O2S2 for the Geodata Deluge

We illustrate a fewrecent ideas of Object Oriented Spatial Statistics (O2S2), focusing on the problem of kriging prediction in situations where a global second order stationarity assumption for the random field generating the data is not justifiable or the space domain of the field is complex.

By localizing the analysis through the Random Domain Decomposition algorithm, we build ensembles of local predictors eventually aggregated in an ultimate one. The localization allowed by the algorithm is also effective for dealing with data which are mildly non-Euclidean and can be locally linearized, as it happens for data embedded in a Riemannian manifold.