ing
‘am
hen
uld
de.
for
lay
ges
‘ent
for
ass
am
of
ple,
the
lore
be
fa
(a)
sed
tion
lum
lori
Or
e of
the
are,
and
ver,
ated
lure,
ated
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
GIS/remote sensing environment. The procedure consist of five
steps involving cither GIS based or image context based
masking (Figure 3).
3.1 GIS Based Enhancement
First, all image and GIS layers have to be registered to a
common coordinate system, e.g., UTM. GIS layers should be
displayed as polygons or raster boundaries overlaid on the
remotely sensed image to check for inaccuracies or geometrical
and semantical inconsistencies of the data sources (see Figure
2).
In a second step, GIS information to be used as feature masks
for local enhancement is analyzed and merged into meaningful
classes. If, for example, vegetation classes are to be evaluated,
all non-vegetation classes can be recoded into one mask. Other
GIS masks that can be used for local image enhancement may
separate water from land or built-up areas from open fields. The
GIS layers can be overlaid on the image data for visual
inspection. With this, editing can be performed if required
(Figure 4).
Figure 4: Selected and recoded GIS classes for ‘Water/Ocean’,
‘Water/River’, ‘Beach’, and ‘Open Field/Roads/Built-Up’
overlaid on multispectral Ikonos data
The third step is the creation of separate image layers that are
based on the selected feature classes. After recoding, the GIS
layers form 0/1 input masks (0 = area outside the feature class,
| = area inside the feature class) to segment the image into
independent layers (Figure 5). Each spectral band of the image
is multiplied with the individual GIS masks to form separate
multispectral image layers with have the original nonzero pixel
values only inside the selected GIS masks. The last image to be
created contains the complement mask to all selected feature
classes. Using this procedure, it is assured that for each pixel
location only one of the image layers contains the original
image value. All the others will have a zero value at this
location. Using the ‘union’ operator, a simple overlay of all
image separates recreates the original image. Figure 5 shows
the separate image layers for 5 different feature classes.
Uu
In a fourth step, each layer is processed separately. This step
does include the selection of an appropriate enhancement
algorithm and the choice of suitable bands for display or
printing purposes. In our study, we worked with 4-band remote
sensing data. This step, however, will become more important if
it involves hyperspectral images. For example, water
information is usually displayed with a higher lever of detail if
the blue band is included. Narrow band widths will make it
possible to select spectral bands that depict physical phenomena
such as turbidity or sediment content. Vegetation, on the other
hand, is displayed best in standard false color infrared band
combination due to the high reflectance in the near infrared
domain.
Figure 5. Separate image layers for the selected GIS classes
‘Water/Ocean’, ‘Beach’, ‘Water/River’, ‘Open Field’ and the
complementary class (mostly vegetation)
The user can interactively be involved in this process or can
leave the display and contrast enhancement to the default
options. The default display options are established based on
user experience and standard image processing literature. For
water classes, the standard bands to be displayed are near
infrared, green and blue (or for Ikonos and Quickbird bands 4,
2. 1). For all other areas, the standard display is near infrared,
red, green (or bands 4, 3, 2, respectively). For image
enhancement, we selected a contrast stretch based on 42.56.
This means that the digital numbers (DNs) for each band are
stretched so that the values [1 - 2.56, u + 2.50] are mapped to
[0,255] (ju being the mean value of the input image). Values
outside the selected range are mapped to 0 and 255,
respectively. This contrast stretch usually produces better visual
results than the histogram equalization process with often too
saturated areas of less discernible level of detail.
The last step involves merging of the separate image layers into
a single image file using standard GIS overlay procedures. As
the image masks do not overlap, the procedure is based on a
simple union process. Figure 6 shows the result of the GIS
based local image enhancement process compared to the
standard full image enhancement option. The GIS layers
selected from the database were ‘Water/Ocean’, ‘Water/River’,
‘Beach’, ‘Open Ficld/Roads/Built-Up’, and ‘Vegetation’. The
GIS based enhanced image shows more detail in all parts of the
study area. There are almost no areas that are too bright or too
dark to convey any information as is the case in the globally
enhanced image which represents a compromise over the
different spectral reflectance distribution for the image (Figure
7).
