Incorporated Decision-maker-based Multiobjective Band Selection for Pixel Classification of Hyperspectral Images

Abstract

Hyperspectral images (HIs) are characterized by a higher spectral resolution than other images and have applications in various fields, to wit, medicine, agriculture, mining, among others. Segmentation can be obtained from the pixel classification and it is a powerful tool for object identification. Notwithstanding, the problems of the curse of dimensionality and the demand for computational resources occur due to the number of bands. Techniques that reduce dimensionality, such as genetic algorithms, are promising, but they cannot assure a balance between conflicting objectives such as improving classification and reducing the number of bands. Multiobjective band selection can be applied to search for tradeoff solutions that have this balance. Therefore, in this manuscript, we propose a novel method called Incorporated Decision-Marker-based multiobjective band selection (IDMMoBS) that tries to find tradeoff solutions using spectral and spatial information. In the experiments, the IDMMoBS reduced the number of bands between 85.4 and 85.8 percent of the total and it outperformed the majority of other methods compared in this criterion. For the pixel classification, the IDMMoBS presented better results than all compared cases taking into account all evaluated metrics using SVM classifier. Accordingly, the IDMMoBS is suitable for band selection.

Keywords

• remote sensing
• hyperspectral imaging
• image segmentation
• image classification
• evolutionary computation