Fusion of Weighted Mean Reconstruction and SVMCK for Hyperspectral Image Classification

Abstract

Conventional classifiers treat hyperspectral image (HSI) as a list of spectral measurements without considering the spatial relationship between pixels. Therefore, these methods discard the information associated with spatial correlations among distinct pixels in the image, and they are sensitive to the noise pixels in HSI. To address these drawbacks, a spatial-spectral combined classification method termed weighted mean reconstruction-based support vector machine with composite kernels (WMR-SVMCK) is proposed for HSI classification. At first, WMR-SVMCK tries to reconstruct pixels by means of spatial neighbor pixels with reconstruction weights to fuse spatial and spectral information. The reconstructed pixels include the spatial-spectral combined features that enhance the separability of the samples from different classes and reduce the influence of noise pixels in HSI. Then, the proposed method utilizes SVMCK classifier to classify the reconstructed pixels. WMR-SVMCK effectively extracts the discriminating spatial-spectral features and enhances the robustness by reconstructing each pixel with spatial neighbor pixels, which significantly improves the classification performance. Experiments on two real hyperspectral data sets (Indian Pines and PaviaU) are performed to demonstrate the effectiveness of the proposed WMR-SVMCK method.