m
n
rd
as
ie
es
1e
se
ne
on
ce
es
ee ET EE
Figure 2. Image 09-08-1992 before (left) and after (right) filtering by the Global Classification Filter
X- Y mn(m ^ n — 2)
T 3 3 (D
4m, + nS, m+n
where À , Ÿ are the mean value of set 1 and set 2, respectively,
2 2
S, . S5 are the variance of set 1 and set 2, and m, n are the
number of pixels in set 1 and set 2. Therefore a curve going
through the T-values is achieved. The extreme values of the
curve are selected as the boundaries. Thus the groups without
significant difference are combined. The same procedure is
performed for the mean value of the original image, the second
variable of the characteristic space, based on the result of the
variance grouping.
The reason for choosing T-test method is that the T-values is a
good indicator of the selected variables. The bigger the
difference in mean value between two sets, the larger the T-
value is. And the smaller the variance in both sets, the larger
the T-value is. The reason for selecting the extreme T-values
as the boundaries is that the intermediate T-values represent
gradual changes, but the extreme T-values indicate the abrupt
changes of the characteristics of the pixels from one group to
another. The filter preserves the edge because if the
characteristic variables describe well the difference of the
areas on two sides of an edge, the pixels in different areas will
be grouped into different classes and get different filtered
values.
Different T-values describe different ground objects: in the
tidal flat area, the water surface has high mean and high
variance because of the alternative bright and dark speckles
from the moving water surface; in the dry flats the mean and
variance are all low because of the specular reflection from the
smooth surface. Grouping the pixels from the same object
reduces the speckle and have the edges preserved. After
filtering by the Global Classification Filter, the speckle was
greatly reduced and the contrast between land and water was
enhanced (figure 2).
The real classification of land and water was done by
thresholding with the help of , from histogram analysis or
765
density slicing observation. It was proved that the threshold
value 1s easier to be determined from the Global Classification
filtered image than that of the Lee filtered image. The
continuity of the water or land areas is also improved with our
filter. The land-water classification map is a binary map from
which the outline of the areas was extracted for the final water
lines.
For the images having difficulty in automated classification,
the knowledge based screen digitizing was carried out.
At this moment, the mean and the variance are selected as the
variables. They may not be the best ones. Later other or more
characteristic variables may be considered. For programming
such a filter, the speed and memory of the computer system
should be well handled because the global strategy requires a
large amount of memory and computing time. The present
program run very well on PC both in time and memory.
2.2 Water surface modeling. Use has been made of the
"Wadden Model” provided by the National Institute for Coastal
and Marine Management (RIKZ) (Robaczewska et. al., 1991).
This model provides the simulation of the tidal water levels
and the tidal current velocities in the Wadden Sea. It is based
on calculations involving 26 harmonic components such as
tidal and rest currents at the North Sea, the climatic conditions,
water stowage or drawdown as the result of strong wind and
wave actions, sea floor bathymetry and roughness, the expected
current drag and the interaction in time between the different
channel systems, etc. The model yields the data in a grid
system of 500 m with a temporal resolution of 2.5 minutes and
simulates the tidal conditions quite well.
To evaluate the model, the time series data of the Wadden
Model were compared with the tidal gauging records over the
same period. The differences, which are not constant, are
principally a result of local change of wind velocity and wind
direction. The model takes the wind influence into account
only when the wind velocity exceeds 8 m/s.
To adjust these differences between the Wadden Model and
the real measurements, the model was corrected by the gauging
data at the time of SAR image acquisition. The differences
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996