JATA
ifferent
s. Two
|. From
| of an
ind the
Terrain
height
th low
terized
sure of
le. The
ise, the
mporal
relation
can be
of each
attering
solution
overall
Dear as
position
system
ositions
ting the
d-cover
tion will
the type
| exists.
e borne
emporal
of two
S-1 and
assed in
resulting
a set of
ell. The
dividual
ositional
distribution of them are regular or irregular.
If the scatterers are regularly distributed and if they have the
same scatter characteristic, the cell is called homogeneous.
Small changes in structure or in chemical composition of the
ground-cover in such a cell will not change the phase if the
imaging geometry did not change.
In a non homogeneous cell the dominant scatterers will
have the highest influence on the phase therefore variation
in the position of dominance will vary the phase
significantly.
Apart from temporal changes, decorrelation can also be a
consequence of an improper (too large) distance (base)
between the two sensors during the data acquisition. To
reduce the baseline decorrelation a careful selection of the
orbits of the sensor system(s) is required. For the extraction
of topographic and thematic information different constraints
are involved. If a data pair is used to extract spatial and/or
height information by means of an interferogram, the base-
line between the two orbits should range from 200 to1000m
to realize an acceptable height resolution. In case of the
extraction of thematic information, the decorrelation related
to the base line should be minimized. If the decorrelation
related to the baseline is zero, the remaining decorrelation
is caused by temporal changes in composition and structure
of the ground cover within the cells .
In practice a zero baseline will not exist, but with a short
baseline almost no baseline decorrelation exists. So
differences that appear can be considered to be caused by
temporal decorrelation. The data can be used as an
additional dimension in the feature space for the combined
image analysis.
1.1 Neural Network classification
Neural networks are based on a model of the human brain,
using certain concepts of its basic structure. The network
consists of many simple processing elements (neurons)
ordered in layers. These layers are separated into an input,
one or more hidden layers and an output layer. The
elements in the hidden layer(s) are connected with all or
with some elements in the next/previous hidden , input or
output layer. In an operational neural network, these
connections are weighted in a training stage. The training of
the network is based on a set of vectors of which the class
membership is known. The neural network classifiers are
able to learn from sample patterns. These classifiers do not
need a particular frequency distribution as required by some
conventional statistical classifiers.
1.2 SAR Coherence image
For the creation of a coherence map, two SAR complex
data sets have to be registered and the coherence
computed. Coherence is a measure for the relation of the
phase information of corresponding signals. To reduce large
fluctuations in the map the coherence is computed in a
window.
According to Schwábish and Winter there are several
factors which decrease the coherence:
- thermal noise
- temporal changes in atmospheric conditions
- phase errors due to processing
- temporal changes in the object phase
* different viewing positions
171
2. DATA PREPARATION
The "ground truth" data was collected in the field and their
positions indicated on a topographic map. The SPOT image
was georeferenced and geocoded to the geometry of that
topographic map. Before the classifications were performed,
the data sets (SPOT and Coherence Image) were
registered. Further a neural net was initiated and trained.
2.1 Data description
For the experiment a data set is selected consisting of an
optical image (SPOT-XS) and a tandem of SAR images
(ERS-1, ERS2) in single look complex format (SLC). In
Figure 1: SPOT XS band 3
figure 1 , band 3 of the Spot image is shown. The SPOT
image is acquired on 02 August 1995, the ERS-1 image
(figure 2) on 19 August 1995 and the ERS-2 (figure 3) on 20
August 1995. Because of the small time interval between
the acquisition dates it is expected that the types of ground
cover of the sensed area have not changed dramatically.
The weather conditions during the data acquisition of the
two ERS images were perfect for the experiment; without
rain but with different wind force and direction.
The baseline of the two images has a horizontal component
of 38 meters and a vertical component of 82 meters
Figure 2: ERS-1 Intensity image
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996