Under the semi-diurnal tides, the water exchange between the
North Sea and the Wadden Sea continuously remodels the
coastal configuration and the bottom topography of the channel
systems and related ebb and flood deltas. The ecological
balance should be carefully maintained.
Disturbance of this fragile ecosystem looms from many sites.
Excessive shell fishing threatens to destroy the mussel banks.
Toxic discharge carried by Rhine and Meuse into the North
Sea enters the Wadden Sea through the inlet channels and gets
concentrated here. Drilling for oil and gas exploration carries
the danger of environmental pollution. Excessive tourists and
recreational activities disturb the tranquillity of the region. For
monitoring such a sensitive area, constant updating of
topographic and bathymetric maps is required.
Conventional mapping methods consist of terrestrial levelling
over the tidal flats and echo sounding over the water covered
areas. The survey and mapping of the Wadden Sea has been
conducted periodically by the North Netherlands Division,
Ministry of transportation and Public Works of the Netherlands
(Rijkswaterstaat). These methods are labour intensive, costly
and time consuming. The ship-borne echo sounding is carried
out during the period of high tide, under the assumption that
the water surface in that period is horizontal. The water level
at ship's location is on reference of water level measurements
from nearby gauging stations. The original echo-sounding data
were recorded along the survey line in 2-5 metres interval,
while the spacing between the survey lines is about 200
metres. Height approximation between the lines are obtained
by interpolation.
2. METHOD
The water line procedure consists of three parts: the image
processing for water line delineation, the water surface
modeling for height extraction and the evaluation of the
accuracy by comparing with the existing DEM.
2.4 Water line delineation and the Global Classification
filter. Radar imaging is near weather independent and sun
light independent, therefore it can acquire sufficient data in a
relatively short period. Variations in the radar backscatter from
moving water surfaces due to changes of surface roughness
modified by the wind and tide may cause difficulties in the
land-water classification.
The following generalizations can be made from the analysis
on basis of 20 ERS-1 SAR images over the study area
(Koopmans et al, 1995):
1). For delineation of the water line, more than 60% of the
images acquired can be used.
2). Wind velocity has obvious influence on the differentiation
of the water surface with the emerged flats (land). When the
wind velocity is higher than 5 m/s, the water surface shows a
continuous light grey tone which has strong contrast with the
dark toned smooth surface of the flats. However, when the
wind velocity is lower than 5 m/s, the water surface varies
very much in tone. Patches of dark toned water surface are
confused with the flats.
3). The recognition of the channel and gullies is not clearly
wind related. Wind velocities between 3.4 and 10 m/s seem
764
to be more favourable. Low and outgoing tides seem to form
the best tidal conditions for channel pattern delineation
when stream velocities are in the order of 1-2 m/s.
4). For visibility of the ebb tidal delta, no straight forward
relationship with the tidal situation or the wind force has
been found.
Nine SAR images acquired during 1993 in this area were used
for water line delineation. All the images were geocoded to the
Dutch national rectangular coordination system (RD system) by
ground control points selected from the coastal constructions.
Some images with good contrast between water and dry flats
allow automatic classification. The main problem is speckle
effect in the SAR image. Several adaptive filters such as Lee,
Kuan, Frost and GMAP filters were tested. They are all aiming
on the reduction of speckle and preservation of the edges and
based on the noise distribution model, which is believed as the
multiplicative noise model. The removal of the noise from the
observed intensity are derived from calculation referencing to
the local window. It was found that for classification it is not
sufficient and efficient by using the local window. The local
window strategy has problems in selecting the window size for
speckle reduction. Using large window size has more
capability of reducing speckle but less effective in the edge
preservation. The smaller window size has the opposite effect.
There is always a trade-off.
The real intensity of the radar return signal is ground object
dependent, therefore the estimation of the real intensity should
be made from all the pixels of the same ground object. The
issue is, how to find those pixels belonging to the same kind of
object, that is, how to classify the image?
In our proposed filter, the Global Classification Filter, the
global strategy was used. The speckle will be reduced by
referencing to a class of pixels of the same object distributed in
the entire image. The mean value of the class will be the new
pixel value, thus the speckle is reduced and different objects
are separated.
The Global Classification filter 1s performed in two steps. The
first step is to define a characteristic space. The space includes
one or more characteristic variables. The selection of the
variable(s) is determined by the application purpose
(segmentation, line detection, etc.). Therefore each pixel of the
image will have characteristic vector(s) in a distribution
pattern in the characteristic space.
The second step is to group these variables by a classifier.
There are many well developed classifiers. The application
adaptive classifier is more relevant. For testing the idea, a
classifier was designed. Two characteristic variables, the local
statistical mean and variance, are chosen for the classifier.
Firstly, all pixels of the original image with the same variance
value (rounded in integer) are put into one group. Then all
groups are arranged according to their variance from small to
large. Some neighbouring groups are not significantly different
in statistics, therefore they should be combined. For this
purpose, the boundaries should be found.
They can be determined by T-test method. Each variance
separates all groups into two sets, say set 1 and set 2. T-values
of these two sets is calculated by
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
thi
cu
Sig
va
va
Th
dii
va
the
as
gr:
ch
an
ch
are
be
va
tid
va
frc
va
sir
re
fil
gri
en
Th
thi