Supervised Feature Reduction Based on a Mutual Information Measure for Hyperspectral Image Classification

Md. Ali Hossain
School of Engineering and Information Technology, University

*Mark Pickering
School of Engineering and Information Technology, University College, The University of New South Wales Australian Defence Force Academy

*Xiuping Jia
School of Engineering and Information Technology, University College, The University of New South Wales Australian Defence Force Academy

     Full text: Not available
     Last modified: July 29, 2011

Abstract
Reducing the dimensionality of high dimensional data without losing its essential information is an important task in information processing. For hyperspectral data, Principal Component Analysis (PCA) is a common dimensionality reduction method which removes the correlation between bands by finding the lower dimensional projections of the original data which have maximum variance. However, since PCA does not take class label information into account, it does not always provide the best set of reduced features for class discrimination. Our proposed approach for dimensionality reduction is based on a measure of Mutual Information (MI). Our new algorithm, called MI-PCA, uses a mutual information measure to find those principal components which are spatially most similar to all the target classes. Our MI-PCA approach selects features which contain the most information about the output class structures and are also uncorrelated. We provide both theoretical and experimental analysis of this method in the paper. Experiments were conducted using hyperspectral data with 191 bands, covering the Washington DC area. The results show that 2 features selected out of 191 using MI-PCA provides 98% and 93% classification accuracy for training and test data respectively, with a Support Vector Machines (SVM) classifier. Whereas, to obtain a similar accuracy, the first 5 principle components are needed for a PCA-only approach. We also compared our approach with measuring the mutual information between the original bands and the class labels. However this approach performed poorly compared to our algorithm as the best features selected were all highly correlated with each other and could only provide discrimination between a small subset of the classes.