Page 125 - Digital Analysis of Remotely Sensed Imagery

96 Cha pte r T h ree

3.4 Data Compression
Compression of remotely sensed data is becoming an increasingly
important issue in digital image processing in light of the emergence of
hyperspatial and hyperspectral resolution data that are measured easily
in hundreds of megabytes and more. These bit-mapped images require
an incredible amount of storage space. For instance, a very small 16-bit
Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) scene of 512
by 512 at 224 spectral bands requires over 117Mb of space to store.
Processing of this small scene requires correspondingly large swap and
temporary spaces for intermediate results. On the other hand, data
redundancy in the form of repeated occurrence of the same pixel values
(e.g., extensive distribution of the same cover on the Earth’s surface) is
rife in some satellite images. Data redundancy is effectively solved by
reducing the data volume via data compression, defined as “the process
of reducing the amount of data required to represent a given quantity
of information” by Gonzalez and Woods (2002). Data compression not
only reduces the amount of data that have to be stored and transferred,
but also speeds up the processing, thus saving time and cost. Data
compression techniques fall into two broad categories, those that do
not result in any loss of information (i.e., error free) and those that result
in partial loss of information. Error-free compression, also known as
lossless compression, is essential when the compressed image data
have to be restored to their original state without any loss of information.
Typically, a compression ratio, defined as the ratio of the number of
information carrying units in the compressed data to that of the raw
data, of 2 to 10 can be expected. There are a number of error-free
compression techniques, including variable-length coding, run-length
coding, and lossless predictive coding.

3.4.1 Variable-Length Coding
The simplest approach toward data reduction is to reduce coding
redundancy. One way of achieving coding reduction is to assign the
shortest codes to the most probable sequence of pixel values in the
input data or the result of a gray level mapping operation (e.g., pixel
difference, run lengths, and so on) after a variable-length code is
constructed. A good example of variable-length coding is Huffman
coding. As the most popular technique, Huffman coding produces
the smallest possible number of codes from the same source than
other coding methods. It involves three steps:
• First, all possible pixel values in the input image are identified
with their probabilities of occurrence calculated. These
probabilities are then ordered in the descending order. The
two lowest probability values are combined recursively to
form a “compound” value that replaces them in the next
round of probability calculation. This process is iterated until
only two probabilities are left.