Geoscience Reference
In-Depth Information
by using the random selection technique. Subsequently, the validation of predic-
tion performances was compared by using prediction rate curve and the unused
landslide points.
5 Results and Discussion
A total of 219 landslide locations were randomly divided into 30 % (66) validation
data, and 70 % (153) training data (Althuwaynee et al.
2012b
) to build LSM1. In
this chapter, the same amounts of training data (153) points were tested by NN
index ratio to identify the spatial nature pattern (Fig.
2
).
The NNI test showed a ratio of 0.53, which proved the cluster nature of
landslides pattern. Also, the expected mean distance was about 1457 m. Since the
expected mean distance represents the limit distance which separates between the
non-random and random distribution in the current study area. For that reason, all
the points' distances with less than the expected mean distance range will be used
for model training, and others will be ignored.
A total 132 or 86 % out of 153 points, have an expected mean distance less than
or equal to 1457 m, and will be used to train the EBF model. Additionally, the
remaining 21 or 14 % out of 153 points were tested with NNI, and the results
confirmed its dispersed pattern as shown in Fig.
3
.
The EBF model was applied with 132 clustered training landslides to produce
LMS2 (Fig.
4
). The prediction rate curve with unknown spatial pattern data of 67
landslide points were used to evaluate the prediction performance of LMS1 and
LMS2. The prediction rate curve showed higher percentage in LMS2 (0.8) than
LMS1 (0.75) as shown in Fig.
5
.
Here two major points can be highlighted: first, the current technique used only
132 cluster points, showed better prediction than 153 random selected points. Also
it confirms the clustering pattern of landslides recorded in the past 25 years in
Kuala Lumpur and vicinity areas. Second, LSM2 showed high contrast and sharp
identified zones edges, with less diffused and uncertain susceptible areas than
LMS1.
