Comparison of support vector machine-based processing chains for hyperspectral image classification

Many different approaches have been proposed in recent years for remotely sensed hyperspectral image classification. Despite the variety of techniques designed to tackle the aforementioned problem, the definition of standardized processing chains for hyperspectral image classification is a difficult objective, which may ultimately depend on the application being addressed. Generally speaking, a hyperspectral image classification chain may be defined from two perspectives: 1) the provider's viewpoint, and 2) the user's viewpoint, where the first part of the chain comprises activities such as data calibration and geo-correction aspects, while the second part of the chain comprises information extraction processes from the collected data. The modules in the second part of the chain (which constitutes our main focus in this paper) should be ideally flexible enough to be accommodated not only to different application scenarios, but also to different hyperspectral imaging instruments with varying characteristics, and spatial and spectral resolutions. In this paper, we evaluate the performance of different processing chains resulting from combinations of modules for dimensionality reduction, feature extraction/ selection, image classification, and spatial post-processing. The support vector machine (SVM) classifier is adopted as a baseline due to its ability to classify hyperspectral data sets using limited training samples. A specific classification scenario is investigated, using a reference hyperspectral data set collected by NASA's Airborne Visible Infra-Red Imaging Spectrometer (AVIRIS) over the Indian Pines region in Indiana, USA.