Recent Advances in Spectral–Spatial Hyperspectral Image Classification
Abstract : Imaging spectroscopy, also known as hyperspectral imaging, has been transformed in the last four decades from being a sparse research tool into a commodity product available to a broad user community. Particularly, in the last 10 years, a significant number of new techniques have been introduced in the domain of hyperspectral image classification. Most of these techniques are characterized by their capacity to take into account both the spatial and spectral characteristics of the hyperspectral data, as opposed to classic techniques for hyperspectral classification that perform in pixel-by-pixel fashion. Spectral–spatial hyperspectral image classification techniques can achieve better performance than their pixel-wise counterparts, as they can combine the rich spectral information contained in the data with spatial-contextual information. In this talk, we provide a comprehensive overview of recent developments in spectral–spatial techniques for hyperspectral image classification in a unified context. The idea of spatial dependency system is first introduced, which involves pixel dependency and label dependency. Resulting from this concept, we categorize available approaches into fixed, adaptive, and global. Then, existing spectral–spatial methods are grouped into four categories according to the fusion stages in which spatial information becomes effective, i.e., preprocessing-based, integrated, postprocessing-based, and hybrid techniques. Finally, typical methodologies are outlined. The talk concludes with a detailed comparison of representative spectral–spatial classification methods using hyperspectral images collected by several instruments, in the context of different applications.