Abstract:

Hyperspectral Data has a large volume compared to panchromatic and RGB data. This large volume can lead to processing, storage, and transmission problems. Therefore, it is crucial to decrease the size of the hyperspectral data for practical applications. Feature selection can be used in order to get rid of large data size problems. In this paper, a preband selection framework is presented to reduce the data size and to reduce the complexity of a well-known band selection method in hyperspectral imagery: Sequential Forward Selection (SFS). The proposed pre-band selection method is based on “dominant sets”. Clustering performance of each spectral band is evaluated, and a reduced set of spectral bands is formed based on the clustering performances. SFS is applied to this reduced hyperspectral data. The aim of the study is to reduce the computational complexity of SFS by applying a dominant set based pre band selection method. Besides reducing the computational complexity of SFS method, results on Pavia and Indian Pines datasets show that the proposed pre-feature selection method performs slightly better than the state-of-the-art feature selection methods in terms of classification accuracy.

Keywords: