Spectral Subset Detection for Hyperspectral Image Classification

Abstract

Listen

Translate article

Remote sensing hyperspectral image (HSI) is one of the emerging technologies for the possible recognition and specific identification of necessary minerals, vegetation, synthetic materials, and ground surfaces. However, HSI contains an enormous amount of information of the scene which is becoming an important field of research, but the ground object identification is a most challenging task for such a high dimensional datacube. Moreover, the spectral bands of HSI are highly correlated in both spatial and spectral contents and suffer from dimensionality problems. As a result, dimensionality reduction is the important preprocessing tasks for the classification. This paper proposed a dimension reduction approach that includes both feature reduction and feature selection for finding a relevant subset of features for efficient HSI classification. Minimum Noise Fraction (MNF) is used as a feature extraction method. Feature selection is done using an entropy-based method, cross cumulative residual entropy (CCRE). To estimate the efficiency of detected subsets, kernel support vector machine classifier is employed on two real hyperspectral image data set. The experimental results indicate a noticeable improvement with respect to classification accuracies. The proposed technique shows 96.72% and 99.75% classification accuracy on image 1 and image 2 respectively which is higher than the standard methods studied.