shows
raining
'arning
utation
oposed
by the
1cation
tion by
indows
tionally
nethods
um
od
at were
aluating
ven: the
'ariance
bes the
ich. are
ger the
chis a
another
ification
:ase the
images,
eference
f testing
e neural
Igorithm
>. better
ation is
> feature
mples is
With the
ion task
ual tool,
s useful
tion and
ible the
spectral
| be used
phasized
when the image presents a highly complex variability of
spectral classes, which may difficult the task of selecting the
classes directly from the original image
Figure 9 shows the classified KCM itself, where one can verify
the separation of classes, building decision regions for the
classification. The non-classified regions, in black, may
indicate the presence of distinct classes, which may be
interactively incorporated in finer and more comprehensive
selections of classes. This may lead to increasingly better
and/or more specific classifications.
In this work we also proposed alternatives for improving the
performance of Artificial Neural Networks in terms of training
time, with the parallel implementation of the SOM and an
advanced learning algorithm for the MLP network, a version of
the Scale Conjugate Gradient algorithm instead of the standard
backpropagation.
The positive results achieved show the feasibility to search for
new techniques either in software or in hardware and their
combination for reaching increasingly better performances as
shown in tables 4 and 5. As for the classification performance
itself, the proposed neural architecture showed superior
performance over a well known statistical classification
method (table 6). The results obtained for neural classification
associated with its performance improvement in terms of time
motivate therefore to continue and to expand research efforts in
this area, since still only a limited part of the power of the
neural nets is actually being utilized.
As an example, another research direction is to further
investigate the applicability and flexibility of dynamic
structural determination techniques for the SOM and MLP
networks, as proposed by (Fritzke 1995) and (Zuben 1996)
respectively. These techniques consist basically in enabling the
network structure to grow up until specific performance criteria
to the application be achieved.
6. ACKNOWLEDGMENTS
The first author thanks to Jesus Christ, the Source of Life and
Hope. The first author also thanks to Nelson Hi Man Pak for
his comments and suggestions. This research was developed in
part with help of CENAPAD (National Center for High
Performance Processing - Sao Paulo - Brazil) project
UNICAMP/FINEP - MCT.
7. REFERENCES
Benediktsson, J. A., Swain, P. H., Ersoy, O. K., 1990. Neural
Network Approaches Versus Statistical Methods in
Classification of Multisource Remote Sensing Data. IEEE
Transactions on Geoscience and Remote Sensing, 28(4), pp.
540-552.
Fritzke, B., 1995. Growing Grid - a self-organizing network
with constant neighborhood range and adaptation strength.
Neural Processing Letters, 2(5), pp. 1-5.
Fletcher, R., 1975. Pratical Methods of Optimization. New
York: John Wiley & Sons.
121
Hepner, G. F., Logan, T., Ritter, N., Bryant, N., 1990.
Artificial Neural Networks Classification using a Minimal
Training Set Comparison to Conventional Supervised
Classification. Photogrammetric Engineering and Remote
Sensing, 56(4), pp. 469-473.
Jacobs, R. A., Jordan, M. L, Barto, A. G., 1991. Task
Decomposition through Competition in a Modular
Connectionist Architecture: The what and where vision tasks.
Cognitive Science, 15, pp. 219-250.
Johnson, R.A., Wichern, D.W., 1982. Applied Multivariate
Statistical Analysis, Prentice-Hall, New Jersey.
Key, J., Maslanik, J. A., Schweiger, A. J., 1989. Classification
of Merged AVHRR and SMMR Artic Data with Neural
Networks. Photogrammetric Engineering and Remote Sensing,
55(9), pp. 1331-1338.
Kanellopoulos, I., Varfis, A., Wilkinson, G. G. and Mégier, J.,
1992. Land-cover Discrimination in SPOT HRV Imagery using
an Artificial Neural Network - a 20-class experiment.
International Journal of Remote Sensing, 13(5), pp. 917-924.
Kohonen, T., 1988. Self-Organization
Memory. Springer-Verlag, New York.
and Associative
Liu, Z. K., Xiao, J. Y., 1991. Classification of Remotely-
sensed Image Data using Artificial Neural Networks.
International Journal of Remote Sensing, 12(11), pp. 2433-
2438.
Moller, M. F., 1993. A Scaled Conjugate Gradient Algorithm
for Fast Supervised Learning. Neural Networks, 6(4), pp. 525-
533.
Oja, E., Lampinen, J., 1994. Unsupervised Learning for
Feature Extraction. Computational Intelligence: Imitating Life,
IEEE Press.
Pearlmutter, B. A., 1994. Fast Exact Multiplication by the
Hessian. Neural Computation, 6(1), pp. 147-160.
Rocha, J.V., 1992. The Influence of Ground Survey Size on
Accuracy of Area Estimates from Satellite Images. Ph.D.
Thesis, Silsoe College.
Rumelhart, D. E. and McClelland, J. L., 1986. Parallel
Distributed Processing: Explorations in the Microstructure of
Cognition. MIT Press, Cambridge, MA.
Schlünzen, E. T. M., 1993. Perspectives on the Application of
Neural Networks in the Classification of Remote Sensing
Multispectral Data. M.Sc. Thesis, Departament of Computer
Engineering and Automation (DCA), School of Electrical
Engineering (FEE), State University of Campinas
(UNICAMP), Campinas, Brazil. (in portuguese)
Zuben, F. J. V., 1996. Parametric and Nonparametric Models
of Artificial Neural Networks and Applications. Ph.D. Thesis,
Departament of Computer Engineering and Automation
(DCA), School of Electrical Engineering (FEE), State
University of Campinas (UNICAMP), Campinas, Brazil. (in
portuguese)
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B2. Vienna 1996
