Unsupervised Band Selection Using Block-Diagonal Sparsity for Hyperspectral Image Classification

In order to alleviate the negative effect of curse of dimensionality, band selection is a crucial step for hyperspectral image (HSI) processing. In this letter, we propose a novel unsupervised band selection approach to reduce the dimensionality for hyperspectral imagery. In order to obtain the most representative bands, the correlation matrix computed from the original HSI is used to describe the correlation characteristics among bands, while the block-diagonal structure is measured to segment all bands into a series of subspace. After applying the spectral clustering algorithm, the optimal combination of band is finally selected. To verify the effectiveness and superiority of the proposed band selection method, experiments have been conducted on three widely used real-world hyperspectral data. The results have shown that the proposed method outperforms other methods in HSI classification application.

I. Introduction

The advances in remote sensing techniques allow sensors to acquire hundreds of spectral bands at the same time, which means that a broad range of spectrum with contiguous and narrow bands could be covered for each pixel in the hyperspectral image (HSI). However, the large volumes of data would entail inevitable costs in storage and computation, which makes HSI difficult to use in real scenarios, even though rich spectral information could provide detailed description. Therefore, it is necessary to perform dimensionality reduction on the HSI data, while ensuring that the key information is maintained.