Image Processing Reference
In-Depth Information
monograph is primarily into remote sensing applications. Therefore, our illustrations
are based on the remote sensing images.
As explained above, the analysis and applications of hyperspectral data have been
explored in various areas, and they have been proved to be highly effective. However,
the data analysis imposes several challenges:
1. A hyperspectral image contains nearly 200-250 bands. For an airborne sensor,
the dimensions of the images are governed by the scanline width and the time of
flight. The field of view (FOV) along with the altitude covers a large terrain on
the earth in case of a spaceborne sensor. Storage of such data requires terrabytes
of memory.
2. The computational costs of data processing are proportional to the size of the
data. Due to high volume of the hyperspectral data, the cost of processing is also
very high.
3. The spectral response of a scene does not change drastically over the adjacent
bands. These bands exhibit a very high degree of spatial correlation due to the
contiguous nature of the hyperspectral sensor array. Thus, hyperspectral data
contain a high amount of redundancy.
The standard display systems are tristimulus in operation. That is, the display
systems can accept three different images as inputs, and assign them to red, green,
and blue channels to generate a color (RGB) image. Alternatively it can accept only
a single image to produce a grayscale image output. The hyperspectral data contain
a far more number of bands than those can be displayed on a standard tristimulus
display. The appearance of the objects is based on their spectral response. Certain
objects in the scene appear prominent in some bands for which their constituent
materials exhibit a large reflectance, while they can practically disappear in some
other bands due to a very poor value of reflectance. Therefore, an observation of
merely 1 or 2 bands does not provide a complete information about the data content.
One has to go through all 200
bands to understand the contents of the image.
Also, in order to understand the spatial relations among two or more objects that
are prominent in different spectral ranges, the observer needs to manually process
the different bands so that the spatial alignment of features across them can be well
understood. To address the aforementioned associated problems and to exploit this
rich source of data, an efficient technique for visualization of hyperspectral image
could prove to be a primary and important processing step. If a human analyst is
able to quickly observe the contents of the scene, s/he can initiate the necessary
data-specific image processing algorithms. One can save a significant amount of
computation by choosing a set of appropriate processing algorithms once the data
contents are roughly known. Visualization of hyperspectral image is thus, very useful
preprocessing task for most applications. This visualization can be accomplished by
means of fusion of bands of hyperspectral image which is a process of combining
data from multiple sensors. Here, the goal of fusion is to form a single image that
captures most of the information from the constituent bands of the data. Image fusion
is a vast area of research spanning fields from computational photography to medical
Search WWH ::

Custom Search