ul 2004
——
avzoglu
tion
k were
ing rate
ely, and
ing the
(1)
esented
training
ol r is
r isin
stuttgart
Institute
s at the
neural
‘raining
y taking
, which
rion for
stopped
se. The
e tested.
inations
[able 5.
celihood
training
nated in
atistical
rporates
als (ie.
liagonal
ne table
ponding
des, 16
IL
851 |
o |
209 _
410
)516
———
os |
rations
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004
It is clear from Table 5 that classification accuracies produced
for Landsat ETM+ and Terra ASTER images (C1 and C2) are
higher than those produced by these images together with their
principal components (C3 and C4). This shows the
ineffectiveness of PCA bands for land cover type delineation.
While the PCA bands do not introduce different distinctive
characteristics, they increase the dimensions and the
complexity of the data. Therefore, they do not make any
significant contribution to the results produced by the two
classifiers. On the other hand, the use of the combination of the
two multi-temporal images (C5) in classification resulted in
slightly better performances. However, it should be borne in
mind that it is difficult to increase the level of classification
accuracy after a certain point, mainly depending on the
complexity of the data set, the number of output classes to be
recognised and the classification technique used.
After the accuracy assessment stage, subset images were fed
into the trained networks to produce the thematic maps of the
study area including eight land cover classes. As an example,
the ML and ANN classification results for C6 combination are
given in Figure 3. The robustness of ANN classification over
ML classification can be easily observed from the figure,
especially for road class. In the ML classification many pixels
were identified as road since the samples collected for this
particular class are known to be mostly mixed pixels that
encompass a substantial region in the feature space compared
to the others. This forced the ML, which relies on statistical
estimates, to wrongly identify many pixels as road in the
resulting thematic image. The weakness of the ML
classification can be also observed for urban pixels near the
upper right corner of the image. Except for these problems, the
ML classification method performed well for other classes.
Column
(a)
When the result of ANN classifier is analysed from Figure 3,
two points draw attention. The first is related to larger number
of bare soil pixels in the image compared to the ML result. The
analysis of ground truth data showed that ANN classifier
produced more realistic results for this class, as well. The
second problem is about the incorrect classification of sea
pixels as inland water for pixels along the seashore. This is
most likely resulted from the ground truth image due to the
possibility of selected pixels of inland water and water
reservoirs being sand or rock. Therefore, the pixels assigned to
inland water class along the seashore could be representing
sand or rock.
5. CONCLUSIONS
The degree of contributions of multi-temporal image data and
their principal components in the production of a classified
thematic map is investigated in this study. The classification
methods considered are the Maximum Likelihood (ML) and
Artificial Neural Network (ANN) classifiers, sophisticated
methods that have been recently used for many investigations.
The image data available for the study were from Landsat
ETM+ and Terra ASTER images acquired in 2001 and 2002,
respectively. The main objective was to use these images to
determine the effect of principal components in the results in
terms of the classification accuracy. ANNs were trained to
learn the characteristics of the training sets prepared for
combinations of images together with their principal
components. Trained networks were later used in accuracy
assessment and in the classification of subset image
combinations. Several important results can be derived from
the results. Firstly, the ANN classifier yielded more accurate
results than the ML classifier. The robustness of the ANN can
be easily seen from Table 5 and from the classified image
Road
assland
Urban
Deciduaus
Coniferous
Column
(b)
Figure 3. (a) Maximum Likelihood classification result, (b) Artificial neural network result
UA
Inland Water