Combining Wavelet Features of Hyperspectral Data by Stacked SVM

Discrete wavelet transform (DWT) provides a multiresolution view of hyperspectral data. This paper proposes to use stacked support vector machine (SSVM) to combine the wavelet features at different layers to improve the classification accuracy of hyperspectral data, where both global and local spectral features could be exploited. After feature extraction using DWT, the wavelet feature set of each layer is processed independently by level-0 support vector machines (SVMs). Then, the decision values of level-1 SVMs at each layer are used as inputs of level-1 SVMs. The classification result of level-1 SVMs is the final classification result. Experimented with the Washington DC Mall hyperspectral data, the results demonstrate that the proposed method can outperform the same SVM classifier with original features, the wavelet features (without fusion), and the wavelet energy features.