Image Processing Reference
similar in their spectral responses. Additionally, the multispectral imaging sensors
have a reasonably wide bandwidth. The pixel magnitude, therefore, is an averaged
value of the response function over the corresponding spectral range which may not
precisely reflect the spectral information. As the spectral bands in MS images are
well-separated, the spectral information is not very dense and uniformly sampled.
Thus, we do not have a set of contiguous and dense sampled spectral bands. In order
to overcome these shortcomings of the multispectral data, hyperspectral imaging
systems have been developed.
A hyperspectral imaging sensor acquires the reflected radiation in the form of a
large number of (nominally over 50) narrow and contiguous bands. Hyperspectral
system contains an array of sensors which samples the electromagnetic spectrum
ranging from the visible to the near-infrared region typically covering wavelengths
m. The nominal bandwidth of each of the element of the sensor array
is 10 nm. The hyperspectral (HS) image provides a densely sampled and almost con-
tinuous spectral signature over the given wavelengths. This is illustrated in Fig. 1.2 c.
Each narrow rectangle in this spectrum (that appears like a thin line in Fig. 1.2 c) rep-
resents elements of the sensor which are almost contiguous. The hyperspectral image
is, thus, a dense and uniformly sampled version of the continuous spectral response.
This characteristic of high spectral resolution makes differentiation of various mate-
rials on the earth possible. The dense and contiguous nature of the data provides
a very accurate and more complete information about the scene. Several materials
radiate the incident light in a higher proportion only in a very narrow spectral range.
Such objects appear prominent and observable only in these related spectral bands;
while these objects could often be dominated by the other materials in their vicinity
at all other wavelengths. Hyperspectral images provide an almost continuous sam-
pling of the spectral data, and thus essentially capture even minor variations in the
scene reflectance. This characteristic imparts tremendous advantages as far as the
data classification is concerned. That is, it enables the user to visualize the scene at
any given wavelength.
The set of hyperspectral bands together is also referred to as the hyperspectral data
cube. The front face of this data cube provides a two dimensional, spatially sampled
information, and the third dimension provides spectrally sampled information along
the depth of data cube. Each band provides a narrow-band image of the field as
seen by the sensor. Along the wavelength dimension, each pixel provides a spectral
signature characterized by the materials at that location.
The main technical content presented in this monograph does not require the
knowledge of the image acquisition systems, We shall provide brief information to
the readers about some of the commonly discussed hyperspectral sensor systems. The
airborne sensor systems are attached to the aircraft that flies over the terrain under
investigation. The sensor elements scan a line across the ground in the direction per-
pendicular to that of the flight. An array of scan-line pixels constitutes the X -axis
data of the hyperspectral image, while the data along the Y -axis gets generated due
to the forward motion of the aircraft. The imaging system splits the incoming spec-
trum at every pixel into a large number of channels that are contiguous and narrow-
band. This aforementioned process produces hyperspectral data with a fine spectral