Bayesian Networks are known for providing a compact and simple representation of probabilistic information, allowing the creation of models which associate a large number of variables. Generative Classifiers are a special kind of Bayesian Networks in which there exist two different types of nodes: the classifier node, and the rest, which are the attributes. In this work different generative classifiers have been used to develop a probabilistic model for the simulation of high resolution spatially distributed weather sequences for binary events. The main difference between discriminative and generative classifiers is that discriminative classifiers (Analogues, Weather Typing, SOM or SVM) directly model the posterior, and generative classifiers first model the joint distribution to get the posterior, taking into account the relationships among the different attributes given some kind of available evidence. Generative Classifiers provide us a simple method for generating stochastic weather in a spatially consistent manner. Instead of simulating values independently for each variable (as in standard weather generator methods) we simulate spatial realizations taking into account the constraints imposed by the dependencies in a graph. For the sake of simplicity, we illustrate the application of simulation algorithms in the case of discrete variables (precipitation), though a similar scheme is applicable to the continuous case.
The study has been carried out over a small area with 42 locations in the north Atlantic face of the Iberian Peninsula, and the precipitation variable based on extreme and normal events was used. The quality measurement is based on the Relative Operating Characteristics (ROC) area. ROC is used to evaluate the different classifiers: naive, augmented and generic, for the three meteorological paradigms: forecasting, nowcasting and climatic. Finally we have shown that the spatial dependencies are useless in the case of pure forecasting, that is, given an estimation of the atmospheric state (i.e. numerical weather prediction). However, these dependencies become very important in the nowcasting and climatic paradigms, that is, when some kind of evidence, other than the atmospheric state, is considered.