Clustering and Mapping Spatial-Temporal Datasets Using SOM Neural Networks

Abstract

Large datasets can be analyzed through different linear and nonlinear methods. Most frequently used linear method is Principal Component Analysis (PCA) known also as EOF (Empirical Orthogonal Function) analysis, permitting both clustering and visualizing high-dimensional data items. However, many problems are nonlinear in nature, so, for analyzing such a problems some nonlinear methods will be more appropriate. The SOM (Self-Organizing Map) neural network is very promising tool for clustering and mapping spatial-temporal datasets describing nonlinear phenomena. The SOM network is applied on the precipitation and temperature data observed in the region of Serbia and Montenegro during 48 years period (1951-1998) and the zonal maps of homogeneous geographical units are derived. These maps are compared with those recently derived via EOF analysis. Significant similarity of results derived from the two methods confirms high efficiency of the SOM network in analyzing spatial-temporal fields. Moreover, the SOM neural network is more appropriate in analyzing climate data since both climate data and the SOM analyzing method are nonlinear in nature.