The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B7. Beijing2008
Input Hidden Output
Fuzzv Slone
Fuzzy Distance to Road
Fuzzy Distance to
Service Centre
Fuzzy Distance to
Administrative Area
Fuzzy Distance to
Residential Area
Figure 4. Application of neural nets in determination
relationship between the driving variable grids, output and
network files.
In order to prepare input files for SNNS software in the
required format, all data layers need to be exported to ASCII
files. Each cell in ASCII files was presented to the network will
have a number assigned to it based on the cells relationship
with the variables and the urbanization process. The higher
numbers represent the neural network’s prediction of a more
likely transition of that cell to urban area.
The pattern file contained information from the 5 final input
grids and one output file so that each line in the pattern file
corresponds to one location. The output layer, on its turn, has
only one neuron that corresponds to the map of transition
probabilities for the considered type of land use change. In the
training data set, the desired (target) value in the output layer is
recorded as 1 for a cell that underwent a change in its land use,
and 0 for a cell that had no change. In the extraction of the
training data set for each network, large rectangles were
delimited within the study area containing representative
samples of the ranges of distances (for the maps of variables)
and of land use change. In each cycle one complete
presentation of all training cells to the network was performed,
mean squared error generated by SNNS and each cycle was
stored in a file for analysis. Based on the analysis, it was
concluded that about 6400 cycles were adequate to stabilize the
error level to a minimum value. It can be concluded that not
only the number of iterations with neuro-fuzzy approach
decreases, but also accuracy of land use change model
improves. For testing, SNNS used the pattern file and the
network file to generate an output file of activation values. The
output file contains values ranging from 0.0 (no likelihood of
changing to urban area) to 1.0 (highest likelihood of changing
to urban area).
3.3 Accuracy assessment of the GIS based neuro-fuzzy
approach
In order to evaluate the results, the predicted land use changes
were overlaid to the observed changes in land use from 1980 to
2000. According to this overlay, four categories were achieved.
In two categories one and four, there is no error; because
observed changes and predicted changes have the same results.
However, the other two categories, two and three, had same
error, because the observed changes and predicted changes
have contradicted results. For this reason, we need Kapa
coefficient to assess accuracy of the mode which uses all of the
four categories for assessment. Kapa coefficient was calculated
for this model equal to 78% which seems reasonable. A layer
was created with the following codes at Table 1:
1
No observed change and no predicted change
2
Observed change but not predicted by the model
3
No observed change but change predicted by the model
4
Observed change and predicted change
Table 1: Coding of predicted layer
GIS was used to compare cells that were predicted to be the
transition to urban areas (according to the model output) with
the cells that actually did transition during the study period. For
assessing the performance of the model, percentage of cells
falling into this category was divided by the actual number of
cells which have been changed to obtain a percent correct
match (PCM) metric (Pijanowski et al., 2000).
GIS was used to determine that 1145 cells changed into urban
class in Tehran Metropolitan Area during the 20-year period
1980-2000. Thus, 2073 cells were selected from the output file
that had the greatest change likelihood values; these cells were
then classified as new urban areas. The test was completed by
comparing those cells that were observed as the changed areas,
with those cells with the highest likelihood of the change, based
on the model. The Percent Correct Metric (PCM) is the number
of 4’s divided by the number of cells that were in transition. We
typically picked the cycle which gives us the best PCM for our
region. Neuro-fuzzy approach undertaken reaches 72%
thematic accuracy.
4. RESULTS
ANN proved to have the capability to model non-linear features
and handles well the uncertainties of spatial data. The
methodology described in this paper showed the potential of
implementing NN algorithms as a tool to predict the land use
change based on historical satellite images. The numbers of
cells indicating land use change have been equal to 1145. This
number matches the number of cells between 1980 and 2000
that actually transitioned to urban and were not part of the
exclusionary layer. The actual recorded urban growth to the
neural network’s prediction has been compared. In the case of
the Tehran Metropolitan Area, 800 km 2 of non-urban lands
were converted to urban areas over the 20-year interval
between 1980 and 2000.
Having urban land use change between 1980 and 2000 and
assuming the existence of the same rate of urban change, urban
land use of Tehran has been predicted for 2020. Based on a
concurrent study in land use change detection in the Tehran
Metropolitan Area between 1980 and 2000, it is clear that the
land use change in this region for this time period is
concentrated along west of the city. The land use change
pattern to the south of the metropolitan area is also substantial
but more dispersed (Figure 5).