International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
rf
kon MR
E
^
* s Lh
SAT, i
A mir,
Maximum Likelihood Method
RCC: 50.93 Iteration: 10000
BCC: 69.59 RCC: 76.38
RMSE: 0.3445 BCC: 88.08
Overall Acc.: 92.22 RMSE: 0.2734
Hidden Neurodes: 6
Overall Acc - 93 57
Hidden Neurodes: 10
[teration: 3000
RCC: 75.53
BCC: 93.16
RMSE: 0.2214
Overall Acc : 95 24
(a) (b)
(c)
Figure (9): a) Maximum-Likelihood Results. b) (3|6]1) neural network's results c) (36|10|1) network's results
Figure 9.a is the result of Maximum Likelihood method which
uses the same training set and has been shown for comparison
between neural networks and statistical methods. Figure 9.b is
the result of optimum simple neural network that just uses
spectral information of a single pixel in its input layer. Figure
9.c is the answer of optimum improved network that its input
vector is consisted of spectral values and normalized distances
of all pixels in a 3*3 window.
5. Summery and discussion
[n this article the impact of input parameters on neural
network's ability for road detection from high resolution
satellite images was tested on multi-spectral Ikonos and Quick-
Bird images. A back-propagation neural network was
implemented with different hidden layer sizes and it was trained
with different iteration times to prevent over-training problem.
As roads are homogeneous areas in high resolution images,
employing neighbour pixels in input parameters can improve
road detection ability of the network, while using the distance
of each pixel to the road mean vector can develop network's
ability in background recognition.
The combination of both mentioned input parameters made
the network powerful in both road and background detection
and also reduced the requested hidden layer size and iteration
time.
It was discovered that there is no need to design more than 10
neurodes in hidden layer as it does not improve results
noticeably and just makes the training and recalling stages more
time consuming.
RMSE proved to be the most reliable parameter to be used as
termination condition since it begins to deteriorate when the
network is about to get over-trained.
References
Benediktsson, J.A., Swain, P.H., and Erosy, O.K. (1990).
"Neural Network Approaches Versus Statistical Methods in
Classification of Multisource remote Sensing data", IEEE
Transaction on Geosciences and Remote Sensing, Vol. 28: pp.
540-551.
Civco, D.L, and Waug, Y. (1994).“Classification of
Multispectral, Multitemporal, Multisource Spatial Data Using
Artificial Neural Networks”. In proceeding of the ASPRS 1994
Annual Convection, Reno, NV, USA. pp. 123-133.
Foody, G.M., McCulloch, M.B., and Yates W.B. (1995).
*Classification of Remotely Sensed Data an Artificial Neural
Network: Issues Related to Training Data Characteristics",
Photogrammetric Engineering and Remote Sensing, Vol. 61,
No. 4, pp. 391-401.
Heerman, P.D., and Khazenie, N. (1992). “Classification of
Multi-spectral Remote Sensing data Using a Back-Propagation
Neural Network”, IEEE Transactions on Geoscience and remote
sensing, Vol. 30, No. 1, pp. 81-88.
Paola, J.D., and Schowengerdt, R.A. (1997). “The Effect of
Neural Network Structure on a Multi-Spectral Land-Use/Land-
Cover Classification”, Photogrammetric Engineering and
Remote Sensing, Vol. 63, No. 5, pp. 535-544.
Richard, J.A. (1993). “Remote Sensing Digital Image Analysis:
Introduction”. Second Edition, Springer, ISBN 0-387-5480-8,
New York.
Yang, G.Y. (1995). “Geological Mapping from Multi-Source
Data Using Neural Networks”, MSc Thesis, University of
Calgary, Canada.
EX
“ZKU
KEY W
ABSTR.
Traditior
used for
is implet
based cl:
spectral
pan-shar
informat
Land us:
existing
Classific
present.
with the
Ikonos i
and they
of inte
interven!
classical
eCogniti
or more
by the a:
the segn
Object-c
informat
crucial
meaning
classific
be set ac
objects.
homogei
segment
resolutic
object c
the com
resolutic
some u:
2001a, t
In this s
in the |
eCogniti
* Corr