A competitive neural network approach for meteorological situation clustering

A complete competitive scheme is proposed in this work in order to perform a classification analysis of meteorological data in the ‘Campo de Gibraltar’ region (in the South of Spain) from 1999 to 2002. The main objectives of the study presented here have been the characterization of the meteorological conditions in the area, using a competitive neural network based on Kohonen learning rule. Standard Principal Component Analysis (PCA) and VARIMAX rotation have allowed interpreting the physical meaning of the classes obtained from the competitive scheme. Quantitative (using three quality indices) and qualitative (from the analysis of the data projection) criteria based on Fisher Discriminant Analysis were introduced to verify the results of the clustering. A randomized procedure is developed to assure the best performance of the models and to select the best model in the experiments. The different experiments developed extracted five classes, which were related to typical meteorological conditions in the area.