Page 236 - Digital Analysis of Remotely Sensed Imagery

200 Cha pte r F i v e

stitching. It cumulates in the mosaicked image. For instance, the
mosaicking of the third component image with the mosaic of the first
two images is subject to the inaccuracy of either of them. Therefore,
the accuracy of the generated mosaic degrades very quickly as the
number of images added to the mosaic increases. The final mosaic is
thus imprecise if the component images contain nonlinear distor-
tions that cannot be removed through rotation and scaling. There-
fore, this uncontrolled method is not recommended for mosaicking
remotely sensed images.
By comparison, it is much easier to produce a controlled mosaic
from georeferenced images in an image processing system that
allows the geometric information of an image to be preserved, such
as ERDAS Imagine. In this environment, an empty mosaic is created
first. Afterward, all component images are dumped into it. Since
they have been georeferenced to the same ground coordinate sys-
tem, the machine recognizes their spatial position in the mosaic
automatically according to their geographic coordinates. Thus, the
component images do not need to overlap each other. If they do
overlap, the overlapped portion is either trimmed or untrimmed
from the resultant mosaic. To produce an untrimmed mosaic, sev-
eral options are available to specify the output pixel values in the
overlapped portion, such as averaging, minimum, or maximum. In
the resultant controlled mosaic, geometric distortions are noncumu-
lative. Instead, they are restricted to individual images. Thus, the
accuracy of controlled mosaics is the same as the accuracy of indi-
vidual georeferenced component images. However, spatial discrep-
ancy in the position of the same features in two adjacent images
cannot be reconciled manually during mosaicking, no matter how
large it is.
Both controlled and uncontrolled mosaicking face the same
issue of radiometric inconsistency across multiple images. Prior to
mosaicking it is possible to unify the radiometric properties of all
component images through some kind of image processing. This
task becomes much easier if the component images are black and
white. Their radiometry can be matched closely by unifying the his-
togram of both images, or making them have the same mean and
standard deviation (see Sec. 6.1.6 for more information). However,
the images will not resemble each other radiometrically (Fig. 5.26)
because the ground features covered vary in their proportion. It is
also rare that the mosaic will have a uniform tone. The task of unify-
ing image radiometry is much more challenging with color images
as color has three dimensions of hue, saturation, and brightness, as
against tone of a black-and-white photograph. Unless the radiome-
try of all images can be unified to an acceptable level, it is recom-
mended that the mosaicked image not be used for any quantitative
analyses.