Band Reducing SVM Classification Method in Hyperspectral Image Processing

Abstract

In this study, a method based on image processing and machine learning is proposed for classification in hyperspectral images. The proposed method is tested on Indian Pines and KSC (Kennedy Space Center) datasets. As a preprocessing step, normalization, median filter and mean filter were applied to all bands in the hyperspectral data set. After the pre-processing, new bands are obtained by averaging the 5, 25 and 125 bands in the dataset. The obtained bands are combined and features extracted. The multi-band dataset is transformed into a single-band feature matrix. Classification is made by adding class labels to the feature matrix. SVM (Support Vector Machines) Linear, SVM Quadratic and SVM Cubic methods are used for classification using MATLAB Classification Learner Toolbox. For all the two data sets, 99% accuracy was obtained with the SVM classification algorithm.