Non linear geostatistic approach to the daily rainfall data modelling

The environmental data modeling have often to deal with non linear problems such as: the probability estimation to overcome a given threshold and long tale distribution. The specific case studied in this paper deals with the reconstruction of daily Rainfall data from Florida. The data set is characterized by a high percentage of zero values, followed by a long tale of decreasing value ending with low frequency intense events. Both the characteristic of the dataset has to be carefully reproduced in the missing values estimation. To this purpose the non linear Geostatistic is introduced. The disjunctive Kriging is a non parametric approach based on the decomposition of the variable in indicators, according with chosen cut-off values. The indicator estimated value represents the probability of the raw regionalized variable to belong to the given interval. It also allows reproducing an anomalous probability distribution without modeling it.