A Comparison of PCA and GA Selected Features for Cloud Field Classification

In this work a back propagation neural network (BPNN) is used for the segmentation of Meteosat images covering the Iberian Peninsula. The images are segmented in the classes land (L), sea (S), fog (F), low clouds (CL), middle clouds (CM), high clouds (CH) and clouds with vertical growth (CV). The classification is performed from an initial set of several statistical textural features based on the gray level co-occurrence matrix (GLCM) proposed by Welch [1]. This initial set of features is made up of 144 parameters and to reduce its dimensionality two methods for feature selection have been studied and compared. The first one includes genetic algorithms (GA) and the second is based on principal component analysis (PCA). These methods are conceptually very different. While GA interacts with the neural network in the selection process, PCA only depends on the values of the initial set of features.