In: Wagner W„ Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Voi. XXXVIII, Part 7B
*
560
layer and the 7 output layer nodes were defined based on the
number of training site categories. For balance between training
time and overall error reduction, learning rates between 0.01 -
0.2 were used, and to reduce the oscillatory problems,
momentum factor between 0.5 and 0.6 were applied. To
terminate the training process, the accuracy rate was set to 90%
and 10000 iteration were chosen.
The more appropriate network parameters considered in this
study were shown in table 2. For comparing the MLP results of
IARR, PCA and MNF inputs this structure was used separately
for each input. Three Hard classification images based on
different input were produced by MLP approach to showing the
lithological units and distribution of salt plug-affected areas
(Fig 3).
Parameter
Value
Hidden layers
8
Learning rate
0.10
Momentum factor
0.5
Sigmoid constant a
1.0
Accurate rate %
90%
Table 2. The more appropriate network parameters that were
used in this study.
Figure 3. MLP mapping results, (A) result of IARR input to
MLP, (B) result of PCA input to MLP, (C) result of MNF input
to MLP.
4. ACCURACY ASSESSMENT OF INPUT DATASETS
To evaluate the results of MLP classification maps obtain from
three input datasets (IARR, PCA and MNF), the accuracy of
salt-affected areas and lithological units were assessed using
ground reference information by determining the overall
accuracy and Kappa coefficient.
The accuracy results are summarized in table 3. Comparison of
the results of different input datasets shows higher capability of
the MNF input to detect salt-affected areas. The accuracy
results indicated that operating MLP with MNF input has higher
accuracy (85%) than the IARR input with 79% accuracy and
PCA inputs with 82% accuracy, so the hard classification image
produced by MNF input was used to assessment of the salt
plug-affected areas.
Input to MLP
Overall
accuracy
Kappa
coefficient
IARR
79%
0.64
PCA
82%
0.77
MNF
85%
0.79
Table 3. Accuracy assessment of hard classification maps
5. RESULT AND DISCUSSION
This study investigated the utility of the MLP network with
different input (IARR, PCA and MNF) for detecting salt plug-
affected areas southeastern Shiraz, Iran.
The ability to map salt plugs, and extent of the salt plug’s
materials are essential to understand and minimize salt plug’s
environmental impact and provides practical solutions to more
advantageous water resources management.
At first, ASTER datasets of the area were analyzed by using the
PCA method. By this method, the ASTER data were limited to
3 bands. Table 1 shows PCA eigen analysis of ASTER image.
These components were used to find suitable training areas for
the classification, as well as gathering sufficient number of
training samples for each lithological unit with the aim of
existing geological map. The training areas were used to
training the MLP neural network and detecting salt plug-
affected areas. Hard classification image of MNF input to MLP
provide the opportunity to map salt plugs and salt plug-affected
areas, as well as to estimate extent of salt plug materials. This
may be important to identify impacts of salt plug on the
adjacent areas, especially on the Firouzabad River (Fig 4).
The Jahani (central part of the scene), Konarsiah (upper part of
the scene) and Kohe Gach (western part of the scene) were
identified from this neural network method. The Hard
classification image of the southern Firouzabad show Konarsiah
salt plug in elliptical shape. This salt plug is located at the top
of the image, surrounded by salt plug-affected areas along the
slopes and margins of the salt plug (Fig 4). Yellow boxes in
figure 4 indicate polluted areas surrounding the Konarsiah. The
spatial distribution of salt plug-affected areas observed in the
hard classification image revealed three main spatial trends.
Relatively, high distribution of salt plug materials is seen in the
northern, southern and western parts of the Konarsiah. It seems
that morphology of the salt plug plays a major role on the
shaping of the salt plug-affected areas, because it controls the
flow of surface runoff and hence the distribution of salt plugs
materials. The main tributaries that convey water as well as
Konarsiah salt plug materials are shown in figure 4, drainages
1, 2 and 3, including branches that convey the Konarsiah
materials into the Firouzabad River. The branch 1, located at
the northern Konarsiah salt plug, drains its materials toward the
east, but branches 2 and 3 are situated in the eastern and
western sides of this salt plug respectively, draining their
materials toward the south.
The Jahani salt plug is located at the center of the hard
classification image (Fig 4). High distributions of salt plug
materials occur in the eastern, western and south western parts
of the Jahani. This image shows that the salt plug materials are
extending down to the Firouzabad River. The amounts of
materials decrease from the salt plug to the Firouzabad River.
The results show a good differentiation between salt plugs
materials and other lithological units however, some
misclassifications occur in south east Jahani salt plug due to the
spectral similarities. The main tributaries that convey the water
as well as the Jahani salt plug materials are shown in figure 4.