Search

In this paper, the usefulness of artificial neural networks (ANNs) as a suitable tool for the study of the medium and long-term climatic variability is examined. A method for classifying the inherent variability of climate data, as represented by the rainfall regime, is investigated. The rainfall recorded at a climatological station in Cyprus over a long time period has been used in this paper as the input for various ANN and cluster analysis models. The analysed rainfall data cover the time span 1917-1995. Using these values, two different procedures were followed for structuring the input vectors for training the ANN models: (a) each 1-year subset consisting of the 12 monthly elements, and (b) each 2-year subset consisting of the 24 monthly elements. Several ANN models with a varying number of output nodes have been trained, using an unsupervised learning paradigm, namely, the Kohonen's self-organizing feature maps algorithm. For both the 1- and 2-year subsets, 16 classes were empirically considered as the optimum for computing the prototype classes of weather variability for this meteorological parameter. The classification established by using the ANN methodology is subsequently compared with the classification generated by using cluster analysis, based on the agglomerative hierarchical clustering algorithm. To validate the classification results, the rainfall distributions for the more recent years 1996, and 1998 were utilized. The respective I- and 2-year distributions for these years were assigned to particular classes for both the ANN and cluster analysis procedures. Compared with cluster analysis, the ANN models were more capable of detecting even minor characteristics in the rainfall waveshapes investigated, and they also performed a more realistic categorization of the available data. It is suggested that the proposed ANN methodology can be applied to more climatological parameters, and with longer cycles. Copyright © 2001 Royal Meteorological Society. 21 11 1401 1414 Cited By :59