bul 2004
also lead
mproved
greatly
ser must
1 a small
range of
» process
nd helps
learning.
3100.9.
i. Higher
ymmonly
training
en alpha
, the user
the alpha
ntrolling
> transfer
id weight
functions
yperbolic
5) will
e will be
function
closed at
ate to be
yperbolic
y.
stopping
monitors
ing when
the point
| test the
ained the
> network
d output.
sd Cross
hich is a
ock of the
aiwan on
Quickbird
ind roads
e LIDAR
; by using
plying an
the false
Roads are
e function
re 72.
abstracted
1 program
ite image.
of inverse
gradients.
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
The criteria of location and direction are fulfilled by
synergizing these results.
3.2 Results
The area size of the test area is in 1229x1209 pixels with pixel
size of 2.44 m. A correlation analysis is carried out for the four
factors, namely NDVI, slope, direction, and locations (Table 3).
As indicated in Table 3, correlation coefficients are very low in
general except V3 and V4 with a coefficient of 0.19. The
highest value of correlation with the target V5 is colour tone V1
with a value of 0.25; and then the slope V2. Principle
component analysis is also applied to extract 4 components, and
thus to reduce the correlation between factors. The relation
between factors and the targets are also reduced accordingly.
Therefore, the components are not adopted for the input of
neural network. Subsequently, information obtained by visual
interpretation as shown in Figure 4. is used to extract inputs of
neural network by extracting 5%, 10%, 15%, 20%, and 25% of
data. Under 4-6-1 neural network structure, various subsets of
random samples apply on 1000 training cycles. Learning errors
for ANN training are shown in Table 4. The MSE (Mean Square
Error) is higher than the threshold of 0.1 required by ANN. The
correlation coefficient is 0.64, indicating that input datasets are
not highly correlated with the targets. So many as 1000 training
cycles are applied to observing learning error curve to see
whether it is possible to reduce the MSE to as low as 0.1. As
shown in Figure 5, after 100 training cycles the network
becomes stable. Classification is further conducted using the
trained network as shown in Table 5. A successful rate of
classification is 85% for landslide and 73% for non-landslide.
The omission and commission error are 0.27 and 0.15,
respectively. The accuracy could be affected by following
factors:
a. The criteria for visual interpretation are not suitable for ANN
in terms of the correlation between the factors and the target.
Part of the reason may be attributed to the mismatch of the
date of various information sources such as Quickbird images
were taken on 15 Jan 2003; the LIDAR point clouds, in May
2002; digital vectors, in August 2002; the landslides, in 1999.
Evidence is given by that the NDVI of manually-interpretated
landslide area was as high as 0.25, indicating the area is
vegetated other than bare.
b.Selected signatures are not good enough to represent the
features as required. Criteria for manual interpretation such as
the cut-off slopes and others are not implemented in this study.
Correlation between Vectors of Values
VI V2 V3 V4 V5
V] 1.000 013 012 ‚009 230
V2 013 1.000 -.044 011 087
V3 012 -.044 1.000 161 -.010
V4 .009 011 .161 1.000 .002
V5 230 087 -.010 .002 1.000
Table 3. Correlation between four signatures and target.
(a) 5% samples (b) 10% samples
MSE 0.46 MSE 0.44
ERROR(%) 13.8 ERROR(%) 13.0
r 0.62 r 0.64
577
(c) 15% samples (d) 20% samples
MSE 0.39 MSE 0.41
ERROR(%) 11.6 ERROR(%) 12.3
r 0.68 r 0.67
(e) 2596 samples (f) all samples
MSE 0.38 MSE 0.51
ERROR(%) 112 ERROR(%) 14.8
r 0.70 r 0.55
Table 4. Learning errors for ANN training
(a) 5% landslide | Non- | (b) 10% landslide | Non-
landshde 84% 16% | landslide | 81% 19%
Non- 28% 72% | Non- 23% 71%
(c) 15% landslide | Non- (d) 20% | landslide | Non-
landslide 85% 15% | landshde | 86% 14%
Non- 21% 79% Non- 23% 77%
(e) 25% landslide | Non- (f) all landslide | Non-
landslide 86% 14% | landslide | 88% 12%
Non- 22% 78% | Non- 45% 55%
Table 5. Accuracy for ANN Training
When the pixels with NDVI larger than 0.25 are filtered out for
the area according to manually-interpretated landslides. Result
shows that the correlation between colour tone and the target is
raised to 0.47. With this condition, the MSE becomes accepted
with a value smaller than 0.1 in a new ANN training cycle. And
the correlation between the factor and the target becomes 0.75.
However, the accuracy of non-landslide is not improved.
4. CONCLUSIONS
Some of the criteria for manual interpretation such as shape
criterion and shadow criterion have not been implemented in
this study due to inadequacy of information. This can be the
reason that the final successful rate of identification of landslide
is only 85%. Further research is required to improve both the
spatial analysis algorithm and the data sources. Nevertheless,
some findings are concluded in this study.
l. It is feasible to gain a synergy of information on high
resolution images, digital terrain models, existing roads and
drainage systems and automate the information for landslide
identification.
2. The correlation analysis of the four criteria for manual
interpretation shows that only direction and location criteria
are correlated. And, only colour tone criterion is better
correlated with the target.
3. Under 4-6-1 ANN network structure, the MSE is 0.43 after
training cycles, not acceptable to the threshold of 0.1.
Furthermore, a correlation coefficient of 0.64 indicates that
the neurons and the targets are not highly correlated. These
could be due to the mismatch of the date of various data
sources.
4. Result of the classification shows a successful rate of 85%
for landslide and 75% for non-landslide. The omission and
commission error is 0.27 and 0.15, respectively ©
5. As shown in this study, GIS functions such as buffering,
spatial intersection, overlay, and terrain analysis are
employed. A system for landslide interpretation would
require capabilities both from a GIS and an image analysis
system.
