Dimension Reduction of Multispectral Data using Canonical Analysis

Remotely Sensed Images are composite images consisting of large number of spectral bands, from electromagnetic spectrum. Analysis and Implementation of such images is much complex processing and takes lot of time. Therefore, dimension of these images must be reduced before any complex operation is performed. Selecting bands, which have higher capability to discriminate between classes, is a process of reducing number of bands with minimum loss of information [1]. In this paper, Canonical Analysis (CA) is used for band selection based on its discriminating power for classification of various classes. CA is based on Fisher’s Linear Discriminant Analysis which maximizes the distance of pixels between classes and simultaneously minimizes the distance between pixels in the same class [5]. It computes eigenvalues and eigenvectors of each band for all the classes. Based on these values, loading factor matrix is computed and the band with highest discriminating power is given highest priority. Band with less priority are not selected leading to reduction of size of the image. Results show that spectral bands 1, 3, 5 are selected using Canonical Analysis whereas bands 4, 3, 2 are selected using Principal Component Analysis from the same LANDSAT image.