**3.4 Landslide susceptibility analysis using artificial neural network**

The landslide susceptibility analysis is carried out by using artificial neural network (ANN) based on the above statistical analysis results.

ANN program is a "computational mechanism able to acquire, represent, and compute a mapping from one multivariate space of information to another, given a set of data representing that mapping, which is independent of statistical distribution of the data, can resolve the nonlinear problem and get high prediction accuracy for classification problem especially for large amount samples (Garrett, 1994). The applications of ANN to landslide susceptibility evaluation have been made by many researchers (e.g. Ermini,L., et al., 2005; S. Lee et al.,2006; Pradhan.B et al.,2010). Nefesilioglu et al. (2008) showed that ANN could give a more optimistic evaluation of landslide susceptibility than logistic regression analysis. Ermini et al. (2005) compared two neural architectures: probabilistic neural network and multi-layered perceptor, and obtained a better prediction result.

In this study, the neural network tool SPSS clementine is used since very few parameters are required. One group of the total slopes are randomly selected for training. 611 collapsed slopes of 885 landslides (70%) and 3300 of 55014 un-collapsed slopes (6%) are randomly selected for this group.

Earthquake Induced a Chain Disasters 397

**Causative factors Case 1 Case 2**  Slope gradient 0.367 0.446 Elevation 0.253 0.327 Slope range 0.193 0.163 Aspect 0.038 ×

Specific catchment area 0.037 × Distance to the fault 0.054 0.016 Distance to the stream 0.035 0.048 Lithology 0.023 × Accuracy 95.05% 96.39% ANN Structure 8\*16\*1 5\*3\*5\*1

By comparing with real landslides, it can be found that 877 of 885 landslides and 3268 stable slopes are identified as high susceptible level, which means 99% of landslides can be

On the other hand, 1 of 885 landslides and 48373 stable slopes are identified as low susceptible level, which means 99.4% of predications are correct and less than 0.2%

0.0-0.01 Low 1 48372 48373 96.53 0.01-0.1 Medium 7 3375 3382 6.05 0.1-0.9966 High 877 3267 4144 7.42

Fig. 10. Earthquake-induced landslide susceptibility map by using ANN analysis.

**Practical Analysis result** 

**Collapsed Stable Slope number (%)** 

Total 885 55014 55899 100.00

predicated by the model but 78.8% of predictions would be false alarm.

Table 3. Weight of each factor in 2 cases

landslides would not be alarmed.

**LSI Susceptible** 

**level** 

Table 4. Characteristics of the three susceptibility zones

Two cases have been analyzed. Case 1 used the 8 factors mentioned as the statistical analysis and Case 2 used the 5 factors, 3 factors with the smaller weight values ware removed from the 8 factors. Also, different layers are used for the two cases.

The weights for each factor, experiments structures and the accuracy from the analysis results are shown in Table 3. It can be seen that the accuracy of Case 2 is a little bit higher than Case 1. Therefore, the ANN model from Case 2 is used for the landslide susceptibility classification.

Fig. 9. Landslide frequency and slope unit numbers (%) for each category of causative factor

The output landslide susceptibility indices (LSI) were converted to GIS grid data in three susceptible levels as shown in Table 4. There are 4145 slopes identified as high susceptible level (dangerous slopes), 48,373 slopes identified as low susceptible level (stable slopes), 3,382 slopes identified as medium susceptible level (gray zone).

Two cases have been analyzed. Case 1 used the 8 factors mentioned as the statistical analysis and Case 2 used the 5 factors, 3 factors with the smaller weight values ware removed from

The weights for each factor, experiments structures and the accuracy from the analysis results are shown in Table 3. It can be seen that the accuracy of Case 2 is a little bit higher than Case 1. Therefore, the ANN model from Case 2 is used for the landslide susceptibility

Fig. 9. Landslide frequency and slope unit numbers (%) for each category of causative factor The output landslide susceptibility indices (LSI) were converted to GIS grid data in three susceptible levels as shown in Table 4. There are 4145 slopes identified as high susceptible level (dangerous slopes), 48,373 slopes identified as low susceptible level (stable slopes),

3,382 slopes identified as medium susceptible level (gray zone).

the 8 factors. Also, different layers are used for the two cases.

classification.


Table 3. Weight of each factor in 2 cases

By comparing with real landslides, it can be found that 877 of 885 landslides and 3268 stable slopes are identified as high susceptible level, which means 99% of landslides can be predicated by the model but 78.8% of predictions would be false alarm.

On the other hand, 1 of 885 landslides and 48373 stable slopes are identified as low susceptible level, which means 99.4% of predications are correct and less than 0.2% landslides would not be alarmed.


Table 4. Characteristics of the three susceptibility zones

Fig. 10. Earthquake-induced landslide susceptibility map by using ANN analysis.

Earthquake Induced a Chain Disasters 399

from the colliding with the vibrating slope during earthquake; (2) the force of friction between a falling stone and the slope can decrease since the normal force varies with the contact condition during earthquake; (3) The flying and rotation movement of a falling stone

In order to consider these effects, we divide a period of wave is divided into two phases: *P*phase and *N*-phase as shown in Fig. 11. The *P*-phase is defined as the period when the slope is moving in the outer normal direction of the slope surface. The slope is pushing the falling stones on its surface and lets them obtain kinetic energy in the *P*-phase. The *N*-phase is defined as the period when the slope is moving in the inner normal direction of the slope surface. Since the normal force will decrease when the slope surface moves apart from the

By the repeated exchange of two phases during an earthquake, the falling stone get multiplex accelerated. The MAM model can be seen more clearly by apparent friction angle

Supposing that a stone with mass *m* moves from position A to position B during a landslide without earthquake (see Case 1 in Fig. 12), the potential energy decreases by *mgh*. Based on the energy conservation law, it is easy to obtain the following equation for a falling stone

tan cos 0

 

*mgh l mgk* (1)

*i i si i*

The first term here is for potential energy and the second term is for the work of friction force between the slope and the falling stone, where the sliding movement is considered and the whole curve path is divided into finite linear segments. And *m* = mass, *g* = gravity acceleration, *h* = the falling height, *l* = the segment length, *θ* is the segment slope angle, *φ* is the friction angle, *k* is the coefficient of conveying from static to dynamic friction and *i* is the

1

 

*i*

*n*

may occur much easily in earthquake induced landslides.

Fig. 11. *P*-phase and *N*-phase definition in MAM

movement in the case without earthquake.

analysis.

index of segment.

falling stones, the force of friction will get decreased in the *P*-phase.

The landslide susceptibility map is made from the ANN results (Fig. 10). The high susceptible zone occupies 7.42%, the low susceptible zone 86.53% of the whole area. In addition, 6.05% of the area is gray zone.

### **3.5 Conclusions**

Landslide susceptible analysis has been carried out in Qingchuan County. 55,899 slope units have been extracted and used for the analysis. The relationship between landslide distribution and the individual causative factor has been investigated by statistical analysis. The clear relationship can be identified for slope gradient, elevation, slope rang, the distances to the fault and the distances to a stream. The ANN analysis also showed the same results, that is, slope gradient, elevation, slope range, distance to the fault and distance to a stream have relatively larger weight. By removing the other three factors with smaller weights, the ANN analysis accuracy got improved. By comparing landslide occurrence locations with susceptibility zones, it has been shown that 99% of landslides can be predicated by the obtained ANN model, but 78.8% of predictions would be false. On the other hand, 99.4% of stable predications are correct and less than 0.2% landslides would not be alarmed. In addition, the gray zone occupies 6% of the whole area. Therefore, the landslide susceptibility classification presented in this study is acceptable.
