(in this
tion of
ftware
/. The
hich is
Is it is
ges in
tensity
essing
en as
scene.
nd the
istered
ifficient
to be
he fine
portant
| is the
)metric
jle the
vanted
iuth * 2
of the
jram is
rrected
art from
it orbits
try. So
ige will
g steps
| PCIS
ution of
ito one.
covers
image.
J range
and resample to 2020 meter resolution.
flip-over the image in line direction to create an almost
west-east image. This step is required in case of images
acquired in descending orbit.
apply an affine transformation to register the ERS-1
image to the SPOT image.
- extract form the coherence map the same window as is
used in ERS-1.
- execute the same steps, with the same parameters for
the transformations, on the coherence map as are used
for the registration of the ERS-1 image.
3. CLASSIFICATION
3.1 MLH classification
Optical data
For the classification of the optical SPOT-XS data all the
bands (3) were used. The classification was performed with
the ILWIS software. The positions of the training samples
were located in the image by visual inspection, i.e. relating
the position of SPOT samples in the topographic map and
the corresponding positions on the false color composite
map. Since the Maximum Likelihood classifier was applied,
the numbers of training samples representing a class were
taken into consideration because of the validity of the
statistical estimates (mean and variance-covariance matrix
). Per class, samples with a size of more than 100 pixels
were taken. In the classification process a threshold, in the
Mahanalobis distance, of 14 was used. Those pixels that
were not classified applying this threshold were considered
to belong to the NULL-class. In table 1 the confusion matrix
and the corresponding accuracies and reliabilities are
shown.
Optical and radar data
The same parameters and sample set are used for the
classification of the SPOT image plus the coherence map.
Only the feature space dimension is increased by one.
In table 2 the result of the classification is shown.
3.2 NN classification
Mask
Neural network classifications in general do not recognize
automatically a NULL class. That means that all pixels are
appointed to one of the classes the network was trained for.
So also pixels that belong to a not sampled land use will be
classified but to a wrong class. To avoid this, a mask was
created using the NULL class of the maximum likelihood
classification.
Optical data
For the investigation the EASI/PACE software of PCI
(version 5.3) is used. A three layer network of fully
interconnected processing units is composed. In the input
layer one node represents one input channel. For the
optical data alone three nodes are used. The output layer
comprises 8 units, one for each class. The size of the single
hidden layer is varied. Benediktsson et al. (1990) Civco
(1991) used both 3 and 4 layer networks. They found that
using a four layer network did not improve the classification
accuracy. The momentum and learning rate were set at 0.9
and 0.1 respectively. The PCI programs use a back-
173
propagation network which is trained by means of the
Generalized Delta Rule. This involves two phases. In the
first phase the weights for the inter-unit connections are
initialized to random values in the range of -0.5 to 0.5. The
input data is presented and propagated forward through the
network. Within each processing unit the combined input
contained is modified by the sigmoid function before it is
passed to other connecting processing units. The second
phase is a backward pass through the network, adjusting
the weights to reduce the error between the actual and the
desired output until it is acceptable or is stabilized. During
the classification of the image, each pixel in the output layer
is allocated to the class associated with the unit with the
highest activation level. After the classification, the
previously created mask is used to create a NULL class that
is identical to the one of the MLH classification. In table 3
the result of the classification is shown.
Optical and radar data
The same parameters and sample set are used for the
classification of the SPOT image plus the coherence map.
Only the feature space dimension is increased by one; the
first layer contains four nodes.
In table 4 the result of the classification is shown.
4. CONCLUSIONS
The overall accuracy of the four classifications does not
vary significantly (about 296). Also the variance of the
average accuracy and reliability is low (about 496). The
main reason for the low variances is the almost perfect
classification of the classes 1 (bare soil), 2 (sugar beets), 3
(stubbles), 4 (forest) and 7 (water) in all classifications. Only
the accuracy and the reliability of the classes 5 (maize), 6
(grass) and 8 (reed) are improved. In these classes a
significant difference exists between the result of the neural
network (NN) classifier on the SPOT data plus the
coherence map and the result of the maximum likelihood
(MLH) classification of the SPOT data alone.
For some classes, which have a high accuracy in the
classification of the SPOT data alone, the accuracy is
decreased in the classification of the combined data. This
decrease is caused by outlyers in the feature space of the
combined data set. The outlyers increase the number of
elements appointed to the NULL-class.
From this experiment we can conclude that the input of a
coherence map improves the result of the classifications.
However it does not give a clear answer which of the two
classifiers gives the best result.
REFERENCES
Benediktsson, J.A., Swain, P.H., and Ersoy, O.K., 1990,
Neural network approaches versus statistical methods in
classification of multi source remote sensing data. IEEE
Transactions on Geosciences and Remote Sensing, 28,
540-551.
Civco, D.L., 1991, Landsat TM image classification with an
artificial neural network., Proceedings, ASPRS-ACSM
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996