AL
thematic
mbining
ncluding
ispectral
istogram
stimated
> photon
umerical
control
ufficient
st, arid,
v digital
'eometry
original
itic and
the same
| affine
del was
x, y) and
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
the indices in the numerical matrix: (x, y) ^ f (row, column),
e the position of grid cells in the ideal grid (u,v) and the
corresponding indices: (u, v) = g (row’, column’).
For resampling the 20 m multispectral data intol0 m scene
elements, a file containing ground control point coordinates
was needed. This was achieved by identifying sets of features
(six were sufficient) that were common to both images using
the 20 m resolution image as "slave" and the 10 m image as a
"master". Features taken as ground control points (GCPs)
were line intersections, field boundaries and intersections of
roads.
The registration process was executed by running the
interactive ground control point selection program to generate
the control point file.
2.03.2 Resampling the SPOT data
Three methods are generally used for resampling: nearest
neighbour (NN), bilinear interpolation (BI) and cubic
convolution (CC).
The NN method was used for this study because it was
readily available. Its advantage is in copying existing values
but it has the disadvantage of producing geometric artefacts,
as does CC. BI does not introduce geometric artefacts. CC
may be acceptable for visual interpretation but not for further
spectral classification methods where the radiometric dis-
tortion will increase.
The differences between GCP locations on the slave image
and the master image were submitted to a least squares
regression analysis that interrelated both image coordinates.
The resulting residuals were less than one-half element
spacing in both row and column, which meant a good
registration was achieved. Three files were created,
representing the resampled three multispectral bands, and
were merged afterwards.
3. FEATURE ENHANCEMENT TECHNIQUES
The usefulness of digital image processing techniques has
been demonstrated many times for extracting. information
about land use, forestry and so on. To assist the work of
interpreters and to improve the visual impact of the image,
the data are usually enhanced after correction.
Image enhancement includes a large range of techniques
which usually transform a given image into another two-
dimensional representation more suitable for either machine
or human decision making (Mulder, 1986; Holderman, 1976).
Features can be divided into spatial features and spectral
features. The three classes of spatial features are
homogeneous areas, edges (boundaries) and lines. Feature
detection follows the evaluation of the "evidence" (le,
indicators of their presence) for one of these classes. If there
is sufficient evidence, a feature is said to be "detected".
Detected features can be represented in an enhanced image or
the evidence can be strengthened. Evidence is derived from a
set of feature extraction operators; for example:
a- smooth areas — averaging operator;
b- edges — gradient operator;
¢- lines — Laplacian operator.
These enhancement operations are characterized by
749
operations over a limited neighbourhood (sub-image) in the
image, using a 3 x 3 kernel filter (convolution) moved across
the image which yields a transformed image.
Every element of a sub-image is processed with its eight
neighbours in the input image by multiplying all elements by
their respective weighting coefficients and the results are
summed. The result is assigned to the central element in the
output image.
4. EDGE AND LINE ENHANCEMENT OF THE
PANCHROMATIC IMAGE
Most digital images are handicapped by noise (unwanted
variation in a signal) and blurred edges caused by the
sampling process. The aim of the experiment was the visual
improvement of the image using enhancement techniques to
make it “sharper’ and easier to interpret. Linear features, such
as roads, railways and rivers, are major components of
topographic maps. The enhancement of edges and lines is
therefore very important and an effective means of increasing
geometric detail in the image.
Blurring is an averaging or integrating operation (Rosenfeld,
1976). The image can therefore be sharpened through
differentiation operations. Laplacian operators (which are
differentiating filters) are the most useful for this sharpening
(Rosenfeld, 1976; Dawson, 1985). The Laplacian filter is an
omni-directional operator and enhances noise as well as
signal, figure 2.
It has been demonstrated in many studies (Rosenfeld, 1976;
Baxes, 1984) that our eye/brain system applies a Laplacian-
like enhancement to everything we view. Consequently, if an
“edge” image is added to its original, the resultant sum image
will have a “natural sharpening”.
0.1.0 |
j Ü | 1:0 I m 6 |
|o 1 0, |
VE Lorene :
{ Line
ETE | FT EU or
1-11] ] [e ;o]o |
} * i ë + 1 1
-18 -1| +0 1/0|= 8| |
-11 -1| 0 0/0 | |^!
: ~ Laplacian | Paint
1 aem r^ EEE
i gemmas dis mets 5 I fe ~~
oiii] =
0.1 1 |a |
M —
Figure 2. Effect of Laplacian filter on a line, a point and an
edge
5. COLOUR SATURATION ENHANCEMENT
Colour coding and image enhancement aim to present the
data in such a way that the eye/brain can extract the
maximum information for a given purpose and under given
