a
poete:
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B3. Istanbul 2004
Table2: Normalized distance as forth input parameter
ele poe E
5 1000 71.45 92.63 | 0.2153 94.32
10 1000 71.52 92.33 | 0,2155 94.44
15 2000 72.49 92.97 | 0.2160 94 47
20 2000 72.55 93.28 | 0.2146 94.34
While background is consisted of different objects with
different spectral behaviours, distance parameter, accentuating
on spectral differences, has improved network ability in
background detection and has brought about decrease in RMSE
value. The decrease in RCC can be interpreted as numerical
problems since d; values for road pixels are very small in
amount near zero.
4.3 Participating neighbours pixels in input parameters
In this section, normalized spectral information in a 3*3
window around each pixel is extracted as 9 red, 9 green and 9
blue values to form input parameters in that order. Accordingly,
input layer involves 27 neurodes. Figure 5 shows the network
structure and obtained results are presented in Table 3.
The comparison between Tables 3 and | shows that RCC has
increased highly, while BCC has decreased in value which
means network's ability in road detection has improved while
background detection ability has been deteriorated. Also spikle
noises were observed that caused increase in RMSE and
decrease in overall accuracy.
1|213
4316 — [0 . 1]
7.1819
5.10,15.20
Neurodes
By increasing the hidden layer size and iteration time, results
got improved but network's training stage took more time in
return.
While designed input parameters in 4.2 and 4.3 have opposite
results as improving background and road detection ability in
comparison with simple network in 4.l, it seems the
combination of both input parameters can make the network
more powerful in both sides. This combination is examined in
4.4.
4.4 Spatial information and normalized distance as input
parameters
In this section the normalized distances of all 9 pixels in the
mentioned window to the mean vector of road pixels are added
to previous stage's input parameters. Thus the input layer is
consisted of 9 red, 9 green, 9 blue and finally 9 normalized
distances to the road mean vector which means 36 neurodes are
designed in this layer. Figure 6 shows network's structure and
Table 4 presents obtained results.
| 192 [3
4 [5 |6 —» [o . 1]
71s T9
5.10.15.20
Neurodes
Figure (6): Network structure when neighbor pixels with their
normalized distances form input parameters
Table (4): Spatial information and normalized distance as input
Internat
MIO UE
Figure (5): Network structure when neighbour pixels form input
parameters
Table 3: Participating neighbour pixels in input parameters
parameters
Neurods | imation | RCC | BOC | RMsE P
5 1000 75.62 93.88 | 0.2008 95.13
10 1000 75.22 94.62 | 0.1999 95.18
15 1500 73.53 95.59 | 0.2012 95.18
20 1500 76.63 95.26 | 0.2006 95.19
DL TT EE
5 3000 80.58 86.80 | 0.2811 91.73
10 3000 81.01 86.54 | 0.2739 92.22
13 3000 80.45 87.59 | 0.2666 92.96
20 4000 80.83 88.33 | 0.2507 93.61
As roads are presented like homogeneous areas in high
resolution satellite images, participating neighbour pixels in
input parameters enables the network to extinguish road pixels
more efficiently and it caused RCC parameter to be improved.
The comparison between Table 4 and Tables 1, 2, 3 shows that
designed input parameters in this section could improve
network's ability in both road and background detection.
Although input layer enlargement can make training stage more
time consuming, but this problem is compensated to some
extent by decrease in requested hidden layer size and iteration
time.
Finally, as a comparison with statistical methods, obtained
results from Maximum-Likelihood classification method and
best network from 4.1 and 4.4 are shown together in Figure 7
with their accuracy assessment parameters