Signal and Image Processing in Hyperspectral Remote Sensing

I n recent years, it has become clear that hyperspectral imaging has formed a core area within the geoscience and remote sensing community. Armed wi th advanced optical sensing technology, hyperspectral imaging offers high spectral resolution—a hyperspectral image can contain more than 200 spectral channels (rather than a few channels as in multispectral images), covering visible and near-infrared wavelengths at a resolution of about 10 nm. The result, on one hand, is significant expansion in data sizes. A captured scene can easily take 100 MB, or more. On the other hand, the vastly increased spectral information content available in hyperspectral images (or large spectral degrees of freedom in signal processing languages) creates a unique opportunity that may have previously been seen as impossible in multispectral remote sensing. We can detect difficult targets, for example, those appearing at a subpixel level. We can perform image classification with greatly improved accuracy. We can also identify underlying materials in a captured scene without prior information of the materials to be encountered, by carrying out blind unmixing. There are many other exciting advances contributed by researchers in hyperspectral remote sensing, and their great effort has resulted in an enormous number of applications, such as surveillance, reconnaissance, environment monitoring, land-cover mapping, and mineral identification, just to name a few. Hyperspectral imaging is also a key technique for planetary exploration, astrophysics, and nonremote sensing problems such as food inspection and forensics. There has been much growth in research activities related to hyperspectral imaging lately. Figure 1 shows a report on the number of publications and citations in the “hyperspectral” topic. The results were obtained by searching the Science Citation Index (SCI)-Expanded database of the ISI Web of Science with the topic “hyperspectral” from 1994 to September 2013. A sharp rise with both the publications and citations counts can be observed from 2010 to 2013. While major research activities on hyperspectral remote sensing are in the geoscience and remote sensing community, hyperspectral remote sensing is also an area that contains many interesting and important signal processing problems. In fact, this area has attracted growing attention and contributions from different communities, such as signal processing, image processing, machine learning, and optimization— and this is what motivates us to organize this special issue. IEEE Signal Processing Magazine published a special issue on signal processing for hyperspectral image exploitation in 2002, which was particularly relevant at the time. After more than ten years, we believe that now would be an appropriate time to consider another special issue on this topic, chronicling recent advances, challenges, and opportunities. Also, this issue has a unique theme—to provide a balanced collection of tutorial-style articles that introduce prominent and frontier signal processing topics in hyperspectral remote sensing and demonstrate the insight and uniqueness of signal processing techniques established in those topics. We also intend to take this opportunity to bridge the gap between remote sensing and signal processing by showing readers a [FIg1] The number of published papers having the keyword “hyperspectral” and the corresponding citations. Data is obtained from the SCI-Expanded database, ISI Web of Science. (a) Published items in each year. (b) Citations in each year. (b) (a) 0 2,000 4,000 6,000 8,000 10,000 12,000 14,000 900 800