A parallel network of modified 1-NN classifiers and R-NN classifiers is described and compared with a standard k-NN classifier. All the component classifiers decide between two classes only. The number of all possible pairs of classes determines the number of the component classifiers. The global decision is formed by voting of all the component classifiers. Each of the component classifiers operates as follows. For each class i a certain area A(i) is constructed in such a way that area A(i) covers all training samples from the class i and possibly a small number of training samples from other classes. In the classification phase, if a sample lies outside of all areas A(i), then the classification is refused. When it belongs only to one of the areas Ai, then the classification is performed by the 1-NN rule. Samples that lie in an overlapping area of some A(i) are classified by the k-NN rule. Such a classification rule, in this paper called a combined (I-NN, k-NN) rule, is used by all component classifiers. Two feature selection sessions are recommended for each of the component classifiers: one to minimize the size of the overlapping areas and another to minimize the error rate for the k-NN rule. The aim of this work is to create a classifier with improved performance compared to the standard k-NN rule. It is shown that the replacement of the k-NN rule by the combined (1-NN, k-NN) rule reduces computing time required for classification while the parallelization of the classifier structure decreases the error rate. The effectiveness of the proposed approach was verified on a real data set of 5 classes, 15 features and 8839 samples which was derived from a couple of multisensorial remote-sensing images. (C) 1998 Elsevier Science B.V.
A parallel network of modified 1-NN and k-NN classifiers - application to remote-sensing image classification
ROLI, FABIO
1998-01-01
Abstract
A parallel network of modified 1-NN classifiers and R-NN classifiers is described and compared with a standard k-NN classifier. All the component classifiers decide between two classes only. The number of all possible pairs of classes determines the number of the component classifiers. The global decision is formed by voting of all the component classifiers. Each of the component classifiers operates as follows. For each class i a certain area A(i) is constructed in such a way that area A(i) covers all training samples from the class i and possibly a small number of training samples from other classes. In the classification phase, if a sample lies outside of all areas A(i), then the classification is refused. When it belongs only to one of the areas Ai, then the classification is performed by the 1-NN rule. Samples that lie in an overlapping area of some A(i) are classified by the k-NN rule. Such a classification rule, in this paper called a combined (I-NN, k-NN) rule, is used by all component classifiers. Two feature selection sessions are recommended for each of the component classifiers: one to minimize the size of the overlapping areas and another to minimize the error rate for the k-NN rule. The aim of this work is to create a classifier with improved performance compared to the standard k-NN rule. It is shown that the replacement of the k-NN rule by the combined (1-NN, k-NN) rule reduces computing time required for classification while the parallelization of the classifier structure decreases the error rate. The effectiveness of the proposed approach was verified on a real data set of 5 classes, 15 features and 8839 samples which was derived from a couple of multisensorial remote-sensing images. (C) 1998 Elsevier Science B.V.
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
