Band selection in hyperspectral imagery using sparse support vector machines

Abstract

ABSTRACT In this paper we propose an ` 1 -norm penalized sparse support vector machine (SSVM) as an e mbedded approachto the hyperspectral imagery band selection problem. SSVMs exhibit a model structure that includes a clearlyidenti able gap between zero and non-zero weights that perm its important bands to be de nitively selected inconjunction with the classi cation problem. The SSVM Algor ithm is trained using bootstrap aggregating toobtain a sample of SSVM models to reduce variability in the ba nd selection process. This preliminary sampleapproach for band selection is followed by a secondary band s election which involves retraining the SSVM tofurther reduce the set of bands retained. We propose and comp are three adaptations of the SSVM band selectionalgorithm for the multiclass problem. Two extensions of the SSVM Algorithm are based on pairwise bandselection between classes. Their performance is validated by using one-against-one (OAO) SSVMs. The thirdproposed method is a combination of the lter band selection method WaLuMI in sequence with the (OAO)SSVM embedded band selection algorithm. We illustrate the p erfomance of these methods on the AVIRISIndian Pines data set and compare the results to other techni ques in the literature. Additionally we illustratethe SSVM Algorithm on the Long-Wavelength Infrared (LWIR) d ata set consisting of hyperspectral videos ofchemical plumes.Keywords: Band selection, classi cation, sparse support vector mach ines, sparsity, bootstrap aggregating,hyperspectral imagery