ABSTRACT

In the hyperspectral case, the remote sensors capture digital images in hundreds of narrow spectral bands spanning the visible to infrared spectrum [15]. Pixels in hyperspectral imagery (HSI) are represented by vectors whose entries correspond to the spectral bands. Different materials usually re¤ect electromagnetic energy differently at speci†c wavelengths. This enables discrimination of materials based on the spectral characteristics. HSI has found many applications in various †elds such as military [16-18], agriculture [19,20], and mineralogy [21]. Target detection and classi†cation are two of the most important applications of HSI. A number of algorithms have also been proposed for target detection in HSI based on statistical hypothesis testing techniques [16]. Among these approaches, spectral matched †lters [22,23], matched subspace detectors [24], and adaptive subspace detectors [25] have been widely used to detect targets of interests. Support vector machines [26,27] have been a powerful tool to solve supervised classi†cation problems for high-dimensional data and have shown a good performance for hyperspectral classi†cation [28,29]. Variations of the SVM-based algorithms have also been proposed to improve the classi†cation accuracy. These variations include semisupervised learning which exploits both labeled and unlabeled samples [30], postprocessing of the individually labeled samples based on certain decision rules [31,32], and incorporating spatial information directly in the SVM kernels [33,34]. More recent HSI classi†cation techniques can be found in Refs. [35-42].