Improving hyperspectral data classification by fusing band and wavelet features

Abstract

The discrete wavelet transform (DWT) provides a multiresolution decomposition of hyperspectral data. Wavelet features of each level are downsampled from the band features. Fine-scale and large-scale information from hyperspectral signals can be separated and this method might provide specific discriminant capability compared to using band features alone. This article proposes using a combination of band and wavelet features (BWFs) in the stacked support vector machine (SSVM), where each feature set is solved independently by level-0 support vector machines (SVMs), and level-1 SVMs are used to correct the errors of level-0 SVMs and obtain the final classification result. The effectiveness of the proposed method was examined using two benchmark hyperspectral data sets collected over forest and urban areas, respectively. For both data sets, the proposed method significantly outperformed SVMs using band features, wavelet energy features (WEFs), wavelet concatenated features (WFs concatenated), and both BWFs and the SSVM using only WFs.