Clustering of climate data in Yugoslavia by using the SOM neural network

Abstract

The climate data are In the form of spatial-temporal fields. The most popular method for analyzing such signals is the empirical orthogonal functions (EOF) method. The method is based on the eigenvectors of the spatial cross-covariance matrix of a meteorological field. The EOF method, being linear, is optimal for feature extraction if the data are well characterized by a set of orthogonal structures or functions. Since the dynamics of climate are nonlinear the EOF may become inefficient. Several nonlinear methods for analyzing such fields are known. Here, the nonlinear analysis by using a neural network of the self-organizing map (SOM) structure is applied on the precipitation and the temperature data observed in the region of Yugoslavia.