Sensitivity analysis

Sensitivity analysis was performed in order to assess the sensitivity of the susceptibility model to changes in the input data and investigate the importance of each factor in the model’s predictive performance. In general, the output of a robust (least sensitive) statistical model should not change significantly if the input data are changed within a reasonable range (Guzzetti et al. 2006).

The model’s sensitivity was assessed by applying variations of equation 3.17 using different factor combinations and evaluating the predictive performance of each model. Ten models were initially tested by excluding one factor in each trial and calculating the area under the prediction curve. If by excluding a factor the model’s performance was lower than an all factor model the factor was assumed to be relatively important. Conversely, if the removal of the factor resulted in the same or higher predictive performance the factor was assumed redundant. Four nine-parameter models obtained by removing the curvature, rainfall intensity, land cover or soil drainage factors indicated the same or slightly improved predictive performance compared to the ten-parameter model. The remaining six nine-parameter models obtained by excluding slope, geology, proximity to faults,

106 proximity to streams, slope aspect and soil induration showed slightly lower performance. An additional 22 other factor combinations were also examined (Appendix 4; Table A4.3).

Of the 31 different factor combinations in total, a six parameter model using the slope, lithology, proximity to faults, proximity to streams, slope aspect and soil induration factors demonstrated the highest performance (0.734) whereas a three parameter model using only the slope, rainfall and curvature demonstrated the worst performance (0.667). The difference in performance of the order of ±0.02, observed by the removal of only one factor in each trial does not necessarily mean that the curvature, rainfall intensity, land cover or soil drainage do not affect landslide susceptibility. It is likely to reflect differences in scale and/or accuracy of the input thematic maps as well. For example, the initial resolution of the rainfall intensity map was 2 km and it was changed to a finer resolution of 25m in order to match the cell size of the other data layers. However this process didn’t improve the accuracy of the information (i.e. the spatial distribution of rainfall intensity) as within an area of 4 km² the rainfall intensity remained uniform. On the other hand, the 24h maximum rainfall intensity of a 10 year design rainstorm may not be the optimum predictor factor of shallow landslide and debris-flow occurrence, and other characteristics such as the mean and maximum monthly rainfall (Schicker 2010) might be more appropriate. However, it should be noted that the effect of rainfall on slope instability is not completely excluded from the model, even after the rainfall information is removed. That is because its effect is reflected through the slope aspect parameter, as the west and northwest facing slopes that are exposed to the prevailing direction of humid air masses demonstrate higher landslide densities.

3.5 Runout

Estimating the “runout” of mass movements including their velocity and travel distance is an essential component of any landslide hazard or risk assessment (Hungr et al. 2005). It provides information not only on the potential affected area and associated risk, but also allows investigation of how the deposited material interacts with other geomorphic processes and may initiate hazards such as landslide dams and consequent dam-beak floods, river aggradation or tsunami waves.

Despite the progress in landslide hazard and susceptibility assessment techniques, modelling the post-failure motion of mass landslides is still a very challenging task, especially in regional-scale studies (Carrara et al. 2008; Hungr et al. 2005). Existing methods for identifying the source area, runout path and deposition zone as well as the kinematic parameters of mass movements are generally classified as empirical approaches, physical-based and dynamic modelling (Chen & Lee 2004).

107 The empirical approaches are based on relationships between landslide characteristics (e.g. volume), topographic parameters and the distance travelled by the landslide debris. These include geomorphological (Costa 1984; Jackson et al. 1987), geometrical (Hsu 1978; Corominas 1996; Dai &

Lee 2002) and volume change methods (Cannon 1993; Fannin & Wise 2001). Although they don’t address material rheology or provide any information on kinematic parameters during runout (Chen

& Lee 2004) their main advantage is that they are simple and easy to implement in GIS for preliminary runout assessment. In physical-based modelling the parameters are derived from field measurements or laboratory experiments (Davies & McSaveney 1999; Major & Iverson 1999).

Dynamic models provide more detailed quantitative estimations of the runout process based on numerical and rheological models including lumped mass (Hutchinson 1986), distinct element (Hart 1993) and continuum mechanics models (Crosta et al. 2003). Hungr et al. (2005) and Chen & Lee (2004) provide comprehensive descriptions of the various methods.

Several parameters such as topography, soil properties, land use, debris volume as well as the water content can affect the runout behaviour of a mass movement (Guinau et al. 2007). Given the difficulty of obtaining and/or predicting these parameters for the entire study area, an empirical approach in GIS environment was developed in this study. The approach assumes that all locations downslope from a source zone are potentially affected until the energy from the mass movement is depleted. Several authors (Michael-Leiba et al. 2003; Jaboyedoff & Labiouse 2003; Hungr et al. 2005;

Toyos et al. 2007; Horton et al. 2008; Blahut et al. 2010; Dahl et al. 2010; Jaboyedoff & Labiouse 2011; Kappes et al. 2011) have used the concept of the angle of reach in GIS environment as a simple rule to determine this depletion point and identify where the movement stops. The angle of reach is the angle of a hypothetical line connecting the head of the landslide source to the distal margin of the displaced mass used as an index to express the runout behaviour of landslide mass (Hsu 1978; Corominas 1996; Dai & Lee 2002). Corominas (1996) studied the runout behaviour of different landslide types and observed a continuous reduction of the angle of reach with increasing volume that starts at the smallest volumes. However, for landslide hazard assessments over large areas, the relationship between the angle of reach and the volume of landslide mass may not be a practical method as it is yet very difficult to predict the volume of future landslides (Dai & Lee 2002).

Furthermore, Davies & McSaveney (1999) studying the behaviour of small scale granular avalanches under laboratory conditions and comparing their results with well-documented field events, argued that the ratio of fall height to travel distance is not adequate to estimate the runout distance, especially for high mobility material. They also found that granular avalanches in the range from 0.1 L to about 10⁵ m³ show consistent runout behaviour which significantly changes for large volume avalanches greater than about 10⁶-10⁷ m³.

108 Considering the above findings and limitations, and as the landslide susceptibility model developed herein does not provide any information on the volume of potential future slope failures, the application of the angle of reach cannot provide meaningful results in the study area. As an alternative, a different approach is proposed in order to estimate the runout distance of landslides, based on the morphology of debris-flow fan formations, which essentially delineate the runout zones of previous events. Note that this approach specifically excludes consideration of large landslides, whether rainfall- or earthquake-generated.