TWORK
dy of the
e primary
lassify by
s. Then
stering is
nethod is
ssful and
level. it
sification.
spectral
lasses is
algorithm
lasses at
| idea of
jssed in
t the test
scussion.
>.
rget old
on. The
vork can
1g things
s able to
1g codes
t can be
available
tudy, the
because
Figure
s of two
modules.
vo fields:
category
/, X, V, U.
p, and q, on the F1 field within which bottom-up input
patterns and top-down predicted prototypes are matched.
The activities of these layers can be found in the figure.
When the input pattern reaches a stable state in field F1,
a bottom-up signal is transmitted to the F2 field. The
neurons on the field then compete with one another for
the signal. If one of the neuron wins, its corresponding
prototype encoded in the top-down connections is
initiated. The prototype and the bottom-up pattern are
then matched to each other within the F1 field. If the
goodness of the match is satisfied with certain criterion,
the input pattern is regarded as being in the category
represented by the F2 neuron. If the match is poor, the
competition process proceeds until all the neurons in F2
field are inspected. f no neuron wins in the competition,
a new neuron is created to represent the new prototype
(or new category).
2.2 Hierarchical Clustering
The spectral classes obtained at ART2 stage are
regrouped using hierarchical clustering method at this
stage. The spectral distance between two classes is
calculated for all classes, then regroups two classes that
has the minimum distance into one class. The
regrouping process continues between the classes until
the class numbers reach the desired numbers. As a
result, the final product will provide the information
classes.
3. DATA
The data used in this study includes a simulated image
and a Spot High Resolution Visible image. The
simulated image is used to test the performance of the
proposed classification algorithm, while Spot image is
employed to demonstrate the practical applications of the
method. The simulated image is designed to know both
gray values and the corresponding class of each pixel in
order to evaluate the classification accuracy more easily.
It is generated first by applying an unsupervised
ISODATA clustering algorithm to a subimage (512 x 512
pixels) selected from an arbitrary Spot image. Seven
classes are produced in this case. Next, the statistical
parameters such as the mean and standard deviation are
calculated for each class in all three bands. Then the
simulated image is formed by specifying the gray value of
each pixel in each band according to its class and the
statistical parameters. In this study, each gray value is
randomly chosen using its corresponding mean and up to
three times of the standard deviation. The resulting
simulated image is shown in Figure 2. It appears that
the image is rather complex and could be a challenge for
testing any classification performance. The real test
image of size 512x512 pixels is selected from a Spot
image. The image was acquired on September 19,
1994 and its corresponding test site is located at Chung-li
area of northern Taiwan. The site mainly is a land-use
mixture of agriculture and urban. The major land-
cover/land-use types can be identified as rice paddy,
grass, barren land, lake, and built-up land. Figure 3
shows the test image.
113
4. RESULT and DISCUSSION
The performance of the proposed method is tested firstly
by using the simulated image. At first, 40 spectral
classes are generated by ART2. The relatively small
standard deviation of gray values (below 3.5) of each
class indicates that the classes generated by ART2 are
rather pure and homogeneous, an indication of fine
classification of ART2. These 40 classes are then re-
grouped to 7 classes. Figure 4 shows the classified
simulated image. In addition, the accuracy analysis is
performed by comparing with original class image. The
overall accuracy of 99% demonstrates the excellent
performance of the proposed method.
The land-use mixture of agriculture and urban features in
the test site basically forms a rather complicated image.
The spectral responses in this area can be separated into
a lot of different spectral classes which may represent
significant or insignificant information classes. It is
obvious that the conventional classification approach will
have difficulty to perform a fine separation of this sort of
surface features. The proposed ART2 approach, in fact,
accomplishes a delicate and fine classification by
separating the test Spot image into 51 spectral classes. .
By inspecting the topographic maps and the ground truth,
it is obvious that the dominant surface appearance in the
test site can be categorized into six major land-
cover/land-use classes: lake, grass, rice paddy, barren
land, highly-reflective roof, and built-up land. Therefore,
51 spectral classes obtained from ART2 are re-grouped
into 6 classes using the hierarchical clustering method.
Figure 5 illustrates the classified Spot image. The
accuracy analysis, based on a random process carried
out on the topographic maps and the classified image,
indicates that an overall accuracy of 9596 can be reached
for the test Spot image.
5. CONCLUSION
An unsupervised neural network classification for
remotely sensed imagery is proposed in this study.
Firstly, the adaptive resonance theory 2 (ART2) neural
network is employed to perform a fine classification.
The main objective of ART2 is to produce the spectral
classes as fine as possible from multispectral remote
sensing data. Then the hierarchical clustering method is
used to re-group the spectral classes to form the
significant information classes. The proposed method is
tested by using a simulated image and a Spot image.
The analysis demonstrates that the proposed approach
needs only few user-specified parameters to perform
unsupervised classification while still keeps an overall
accuracy above 95% for both simulated and Spot images.
The test results indicates that the proposed method is
very promising for the practical application of remotely
sensed image.
REFERENCES
Bischof, H., W, Schneider, and A. Pinz, 1992.
Multispectral classification of Landsat images using
neural networks, IEEE Trans. on Geoscience and
Remote Sensing, Vol. 30, No. 3, pp. 482-490.
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B7. Vienna 1996
