Page 257 - Digital Analysis of Remotely Sensed Imagery

Image Enhancement 219

It must be pointed out that the two images will not look alike
even if their histogram has been perfectly matched unless they are
multitemporal images of the same geographic area that has not expe-
rienced much radiometric variation in the interim. This effectively
guarantees that the proportion of different land covers in the area
remains little changed. Since different objects have different spectral
properties, water pixels tend to have a smaller value than their land
counterpart. A perfect matching is almost impossible to achieve in
this case. Besides, even the second best match still causes noticeable
disparities in radiometry between the two images.
The above discussion applies to black-and-white imagery of a
single band. It is more difficult to achieve a perfect match for color
images as they involve more parameters. Apart from hue, saturation,
and brightness, both need to be matched as well. So the matched
color image will resemble the master image less closely than if it were
black and white.

6.3 Spatial Filtering
Spatial filtering is a window-based image processing technique for
altering the input pixel value based on its own value and the value of
the pixels surrounding it. It requires the use of a spatial mask known as
a spatial filter. Filtering is carried out to achieve several functions, such
as image smoothing and feature enhancement within a neighborhood.

6.3.1 Neighborhood and Connectivity
In a raster image, neighborhood refers to a defined window inside
which all the pixels surrounding the one in question are considered
its neighbors. The appearance of a neighborhood depends upon the
definition of connectivity. There are two types of connectivity, four
and eight. In the former case, four neighboring pixels above, below,
to the left and to the right of pixel (r, c) are regarded as the neighbors
(Fig. 6.11a). In the eight-connection situation, all the pixels immedi-
ately adjoining the pixel under consideration at (r, c) are regarded as
its neighbors, including those to the upper left, upper right, lower
left, and lower right (Fig. 6.11b). This neighborhood is defined by a
window size of 3 × 3 pixels. Other larger neighborhood sizes (e.g., 7 × 7
and 9 × 9) are also commonly used in spatial filtering.

6.3.2 Kernels and Convolution
Spatial filtering requires a convolution kernel, also called “template”
by some authors. A kernel is a matrix of values whose size governs
the sphere of influence of neighboring pixels in spatial filtering. Com-
mon kernel sizes are odd numbers between 3 and 9. The larger the
kernel size, the more computation is involved, the more the output
pixel value is subject to that of its neighboring pixels. Elements in the