Page 130 - Digital Analysis of Remotely Sensed Imagery

Storage of Remotely Sensed Data 101

components, such as horizontal, vertical, and diagonal coefficients with
zero mean and Laplachian-like distributions. Most of the important
visual information is projected into a few coefficients. They are then
quantized and coded using one of the lossless coding methods
mentioned above. Those coefficients that carry little visual information
are either quantized at a coarse level or discarded altogether. Con-
sequently, the raw image cannot be restored via decoding that is accom-
plished by inverting the encoding process without the quantization step.
Wavelet coding differs from transform coding in that the input
image does not have to be divided into subimages because wavelet
transforms are both computationally efficient and inherently local.
The level of computation intensity is affected by the specific form of
wavelets. There are several in use, the most common being the
Daubecies wavelets and biorthogonal wavelets. The latter is more
computationally intensive than the former, but can achieve a higher
compression ratio. The level of computation intensity is also affected
by the number of transform decomposition levels.

3.4.5 JPEG and JPEG 2000
The JPEG compression is usually implemented in several sequential
steps (Smith, 2004):
• First, the image is divided into subimages of 8 8 pixels from
left to right and from top to bottom, each to be compressed
independently. A subimage initially represented with 64 bytes
is reduced to much fewer bytes by subtracting the quantity of
n-1
n
2 , 2 being the maximum pixel value. The difference is then
transformed with the discrete cosine transform (DCT). DCT
is the best among various standards in terms of ease of
implementation and the achievable compression ratio. Block-
based DCT techniques are characterized by lossy compression
that has become the norm. Thus, it has been widely used to
store satellite data in several remote sensing systems at
present, even though other more efficient and flexible
compression techniques have become popular. Each of the
64 spectra produced by the 8 8 subimage has the amplitude
of a basis function. Each spectrum is compressed by reducing
the number of bits and eliminating some of the components
in a step controlled by a quantization table.
• Next, the modified spectrum is converted from an 8 8 array
into a linear sequence, at the end of which all of the high
frequency components are merged. This groups the zeros
from the eliminated components into long runs. These runs
of zeros are compressed using run-length encoding.
• Finally, the compressed file is formed by encoding the sequence
with either Huffman or arithmetic encoding.