Supervised Hyperspectral Image Classification Based Upon Adaptive Transforms

Abstract

Hyperspectral imaging systems has grown popular in the area of remote sensing for the past three decades since its introduction. Hyperspectral image classification is one of the common field of application and supervised kind of classification is commonly preferred. This is because of the fact that a hyperspectral image (HSI) yields significantly better results compared to that using a multispectral image. But, it is computationally complex due to its high dimensionality. One solution is feature extraction which is a pre-processing step in hyperspectral image analysis. Raw spectral bands from hyperspectral image are not significant enough for certain applications such as classification and so feature extraction techniques are applied prior to the hyperspectral image classification algorithm application. In order to deal with high dimensionality, several dimensionality reduction techniques has been proposed in literature for hyperspectral images. Here, we propose a chebfunction feature based dimensionality reduction approach that is done prior to the feature extraction procedure. In this thesis, usage of spatial feature extraction based on adaptive transforms such as two-dimensional Empirical Mode Decomposition (EWT2D) and two-dimensional Variational Mode Decomposition (2D-VMD) are being proposed. Classifiers based upon Support Vector Machine (SVM), Hybrid Support Vector Selection and Adaptation (HSVSA), Subspace Pursuit (SP), Orthogonal Matching Pursuit (OMP) are being utilized for the experiments regarding supervised hyperspectral image classification task. Both EWT and VMD based spatial feature extraction procedure gives significant performance in terms of classification evaluation metrics for hyperspectral image classification task. The combination of chebfunction feature based dimensionality method and the spatial feature extraction method (either EWT or VMD) are compared with results that has been obtained using either EWT or VMD based features and raw hyperspectral image pixels. The obtained results are sig