Hyperspectral Image Band Selection Using Pooling

Hyperspectral images contain hundreds of spectral bands. These bands contain abundant information and more often redundant information. This article presents an unsupervised band selection method to choose most significant spectral image bands from hyperspectral datacube which maximize the relavance and minimize redundancy. The outcome of the research can be used for two purposes. One of them is to fabricate multispectral sensor which is more effective in identifying the subjects of interest. The other purpose is to enhance computational efficiency in classification of the hyperspectral image by processing only the selected spectral bands. Proposed method is based on spectral pooling of hyperspectral data cube. Since spectral pooling methods cause information loss which reduces the discrimination of material classes. This article suggests a novel approach which can minimize the loss while improving material classes discrimination. The proposed technique uses min-max pooling together with reflectance intensity gradient of neighboring pixels of hyperspectral image in spectral data axis. Experiment results of classifying terrain hyperspectral datasets are presented to validate the consistency of image classification which contains several material classes. A comparative analysis has been conducted between max pooling, principle component analysis and the proposed min-max pooling.