Internation tl Arca es of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV. Part B2. Istanbul 2004
output neurons. If the observed error exceeds the desired error,
the output signals are fed forward to the hidden layer. In the
same way, the signals in the hidden layer are fed forward to the
input layer. in which it is also calculated to the hidden-layer
error. Finally, the weights between the hidden layer and the
output layer, and the weights between the input layer and
hidden layer are adjusted. The iterative procedure involving the
previous two stages continues until the output layer error is
within the specified threshold.
Upon completion of the iterative training procedure the entire
224 by 224 SAR scene was classified. The result of the
classification of the SAR image is shown in figure 2. In figure 2
the blue, turquoise and yellow colours are associated
respectively to poplar, bushes and background. For the two-
element vegetation classification into poplar and bushes in well
differentiated regions the probability of yielding poplar and
bushes classification was found to be 100%. When the scene
includes transition regions, i.e. boundaries between trees and
bushes, the resulting probability of correct classification is
reduced marginally. The same is observed in the boundaries
between the trees and the background. It is worthwhile stating
at this point that cultivated areas like those exhibited in figure 1
often exhibit significantly different intensity (digital number)
values; this 1s evident from a cursory inspection of figure |.
This will lead to confusion in decision boundaries and so, result
in erroneous classification when conventional classification
schemes are applied. The addition of wavelet decomposition
fields reduces the confusion that might arise from consideration
of the reflectance fields alone as areas of poplar and bushes
exhibit a similar high energy content —1.e. 1, 2 and 3.
Figure 2. The result of classification of the SAR image.
5. CONCLUSIONS
This paper demonstrates the integration of multispectral SAR
data coupled to representations of the image by wavelets. The
resulting nonlinear nature of the data forms a complex set for
classification. A modified neural network is therefore employed
as a means to classification. The proposed approach has shown
good results on a three-element classification into poplar,
bushes and background in a region where the features are
186
visibly well differentiated. This illustrates that SAR imagery
can be classified effectively where there are distinct regions of
each class and when additional information is provided on
texture and structural features with rather strong orientation. An
important consideration in the whole process of classification of
SAR images, however, is that it involves a multiple-exploitation
of techniques to interpret land surface characteristics. This
particular combination of methods was consistent in this
context but extension into further environments, both different
and more complex, requires further research.
REFERENCES
1. . BRYAN, M.L., 1979, The effect of radar azimuth angle
on cultural data. Photogrammetric Engineering and
Remote Sensing, 45, 1097-1107.
2. BARBAROSSA, S., and PARODI, L., 1995, SAR image
classification by wavelets, [EEE International Radar
Conference, 462-467.
3. CHUI, C.K., 1992, An introduction to wavelets, (Boston
Academic Press).
4; | .DAUBECHIES, LD 1998, \ Orthonormal bases of
compactly supported wavelets. Comm. Pure and Applied.
Math, 41, 909-996.
5. GURNEY, K., 1997, An introduction to neural networks,
(UCL Press Limited)
6. LIPPMANN, R.P., 1987, An introduction to computing
with neural nets. IEEE ASSP Magazine, 4, 22.
7. MALLAT, S., 1989, A theory for multiresolution signal
decomposition: the wavelet representation. [EEE
Transactions on Pattern Recognition and Machine
Intelligence, 11, 674-693.
8. PAOLA, J.D. and SCHOWENGERDT, R.A., 1995, A
detailed comparison of backpropagation neural network
and maximum-likelihood classifiers for urban and land
use classification. /EEE Transactions on Geoscience and
Remote Sensing, 33(4), 981-996.
9. RUMELHART, D.E., HINTON, G.E., and Williams R.J.,
1986, Learning internal representations by error
propagation. Parallel Distributed Processing: Explorations
in the Microstruction of Cognition. D.E. Rumncelhart and
J.L McClelland, (Eds.). Cambridge, MA, The MIT Press.
I: 318-362.
10. SIMARD, M.. SAATCHI, S. and DeGRANDI, G.. 2000,
Comparison of a decision tree and maximum likelihood
classifiers: application to SAR image of tropical forest,
IEEE 2000 International Geoscience and Remote Sensing
Symposium, 5, 2129-30.
If. Kirkpatrick, 5S, €. D. Gelatt 4r; M. P. Vecchi,
"Optimization by Simulated annealing", Science, 220,
4598, 671-680, 1983.
Laborato
316-317,
KEY WOI
ABSTRAC
It is estima!
to be usefu
general use
become the
fits the ind
access, anc
automate a
geocomput:
services, gx
service stai
knowledge
knowledge-
managemer
Data is a 1
meaning m
be associat
dominant f
that more tl
far is the ;
many aspe
environmer
importance
money hav
and distribu
data is appt
For exampl
two petaby
data arrivin
Information
means of tl
an organize
In order for
be extractec
Because of
disciplinary
useful infor
The traditi
involves in,
are the exp
volume of t
of the gcos
people afte
lacking of
