Multispectral & Hyperspectral Remote Sensing Image Extraction

Multispectral sensors such as the Landsat series, SPOT, IKONOS, and QuickBird acquire anywhere from three to ten simultaneous bands of information across a scene.  Each of these bands cover a relatively broad spectral range of electromagnetic radiation observation.  Hyperspectral remote sensing, as its name implies, is generally composed of a greater number of spectral bands which observe a more precise (narrow) spectral threshold.  The MODIS sensor uses 36 bands, while NASA's AVIRIS sensor captures as many as 224 bands.

John R. Jensen continues to describe in his book, Introductory Digital Image processing: A Remote Sensing Perspective, two methods of multi- and hyperspectral image acquisition methods: the whiskbroom system, and the linear and area array technique.  In both methods, instantaneous field-of-view (IFOV) radiant flux – observed reflectance from the Earth's surface – is passed through a spectrometer in order to disperse, or separate the light into separate bands; ranging from blue to near infrared (NIR) and IR wavelengths, which are passed onto a spectrometer where it is dispersed and focused onto a array of detectors which digitally record the field of view.  The whiskbroom method uses a rotating mirror to reflect and direct radiant flux through the spectrometer to a linear array detector which individually measures the value of radiation of each band that has been separated.  This technique is best suited for capturing broad spectral ranges and is utilized by multispectral sensors such as Landsat MMS and SPOT, etc.  These sensors undersample the observed radiant flux by making only a few measurements from wide spectral bands; as wide as several hundred nanometers which may cover more than one color of the spectrum simultaneously.  Although the use of a dispersing element is similarly used in each method to separate incoming light into individual bands, the alternative method of does not use a scanning mirror; thus allowing a longer amount of time for a detector to record the incoming radiant flux of a given area.  This extra duration of detection yields improved geometric and radiometric accuracy.

These methods have dramatic implications on the type of information that is produced.  The varying techniques influence spectral, temporal, radiometric, and even spatial resolution of a produced image.  Various types of investigation require unique parameters in the type of data that will be used in analysis.  Having an understanding of the different types of information that are produced by multi- and hyperspectral imagery with a respect to ground conditions that are being observed will ensure the most accurate results are obtained during image analysis and exploration.

The process of extracting information from multispectral and hyperspectral imagery are largely similar, however there are a few preliminary steps one must complete before analyzing hyperspectral datasets.  Various forms of calibration – radiometric, geometric, etc. – are common between these two types of imagery, however the larger volume of highly specific spectral bands associated with hyperspectral imagery permit the construction of "spectra" that closely resemble the quality of spectral signatures captured by spectroradiometers in laboratories.  Further, initial image quality assessment of hyperspectral can be a much more tedious undertaking, although it is occasionally possible to use hyperspectral images with poor quality for atmospheric correction.  Due to this more time consuming and redundant task of visual examination, many image processing packages have animation functions that provide a more efficient means of inspecting up to hundreds of images in a single session.

Overall, the dimensionality is the most distinguishing characteristic between these two types of remote sensing techniques.  The low data dimensionality of multispectral imagery is significantly more accessible and less difficult to work with in basic research due to the low number of spectral bands that make up an image.  The higher number of bands/images involved with hyper- and ultraspectral imagery generate numerous obstacles, ranging from data storage to processing abilities.  Numerous images of highly specific spectral bands – often representing bandwidths of just 10 nm each – produce certain amounts redundant information.  Statistical analysis aids in identifying, and removing or transforming data in order to reduce the dimensionality of the overall hyperspectral dataset, improving the efficiency of exploration and analysis.  Although there are indeed variations that exist with regard to image processing techniques between these two types of remote sensing data, the causes again are largely due to dimensionality.  An analyst must decide whether few, spectrally broad images or numerous and specific images will best suit his or her analysis.