Predicting the Relative Likelihood of Landsliding in the Central Appennines, Italy

P.M. Atkinson and R. Massari
Department of Geography, University of Southampton, Highfield, Southampton, SO17 1BJ

Introduction
Generally, one may distinguish between landslide mapping and landslide hazard mapping. Landslide mapping amounts to describing the spatial distribution of landslides that have already occurred within a given region. Landslide hazard mapping, on the other hand, amounts to predicting spatially the likelihood of landsliding (that is, landslides that might occur in the future) (Hansell, 1984). Several reseachers have attempted to map the landslide hazard (for example, Carlara et al., 1991; Chacon et al., 1993).

Although slope movements are often the result of a single triggering cause, such as rainfall, earthquakes, excavations, erosion and vegetation clearance, they also depend on the existence of manly properties that make the slope inherently susceptible to failure, for example, slope angle, lithology and ground water. The former "triggering" properties are termed extrinsic and may change over a very short time scale. The latter "susceptibility" properties are termed intrinsic and may be expected to change over a geomorphological time scale only (Siddle et al., 1991).

Where knowledge of extrinsic properties is limited, the spatial distribution of intrinsic properties within a given region of interest determines the spatial distribution of the relative likelihood of landsliding in that region.

Previously researchers have taken two different approaches to mapping the potential for landsliding. In the first approach, physical models are used to predict landslide potential from independent variables only. Recently, geographical information systems (GIS) have become an important tool in this field (Wang & Unwin, 1992; Chacon et al., 1993).

In the second approach. a statistical model of the relation between the potential for landsliding(dependent variable) and a set of intrinsic properties (independent variables) is constructed using statistical techniques (and sample data) This model is then applied to map the landslide hazard (Hansen, 1984; Wang & Unwin, 1992). In this paper, the statistical approach is take within a GIS framework.

In the statistical approach to landslide hazard mapping, one never has data on future landslides. The statistical model must be constructed between a set of selected independent variables and data on past landslides. As a consequence, the predicted map is really a landslide map and not a landslide hazard map. However, for those slopes that have not already failed, the relative differences in the predicted values may point to relative differences in the propensity for landsliding.

Most importantly, the tendency to map the landslide hazard may be increased by selecting intrinsic properties (independent variables) that reflect conditions before slope movement, rather than after slope movement.

Field site and data
The study area is located in the Umbro-Marchean Appennines in central Italy. The site lies between the latitudes 43" 30' and 43" 35' and longitudes 12" 27' and 12" 32' and covers an area of 62 km2. The geological formations outcropping cover a time interval between the Jurassic-Cretaceous (limestone and marls) to the north-east and the Miocene (flysch formations) to the south-west.

For the present analysis six intrinsic properties only were selected: slope angle, slope aspect, geology, density of lineaments, slope-strata interaction, and strata orientation. Slope angle was determined separately for each landslide unit to provide slope conditions prior to failure. The averge slope for a buffer of pixels around the rupture area of each land slide was taken as a good indicator.

Generalised linear model
Generalised linear modelling allows one to form a multivariate regression relation between a dependent variable (in the present case, the presence or absence of landslides) and several independent variables. The variables may be either continuous or categorical or any combination of both types. Where the dependent variable is categorical (and, in particular binary, as in the present example) the logistic model (Equation 1) is an appropriate link function:

where p is the probability of a landslide occurring and the n xi are the independent variables.

In the present analysis, a single pixel was extracted from the centre of each of 380 landslides to give a set of 380 locations where landslides are present. To balance the number of pixels for which landslides are present against the number for which landslides are absent it was necessary to obtain 2000 pixel locations from the remaining 'stable' area.

For each of the 2380 pixel locations values were extracted on the presence or absence of landsliding, and on each of the six independent variables. These data were then input to the generalised lineal modelling algorithm to obtain coefficients for the logistic regression relation. The model was then applied to the set of six data layers for the whole region and a map of the relative likelihood of landsliding produced.

For the areas that have not yet landslided, one may interpret the prediction in relative terms. If one pixel has a larger value than another then it is predicted to have a higher probability of landsliding. In simple terms, large value for a given pixel mean that the independent variables for that pixel take similar values to the independent variables for the areas which contain landslides.

Assessing the accuracy of the predictions is problematic since no test data set is available. To assess accuracy one would require data on future conditions. The only way to assess accuracy is by comparing the predictions to the landslides that have already occurred. We compared the presence or absence of landsliding with the predictions (discretised into ten separate classes). Areas that were predicted to have relative values of between 0.0 and 0.1 (lowest probability of landsliding) corresponded to areas covered less than 10% by landslides. Areas that were predicted to have relative values of greater than 0.7 (highest probability of landsliding) corresponded to areas covered around 42% by landslides. The results are encouraging, but should be treated with care since it is implicit in this method of assessment that areas that have not yet landslided are stable. If this really were the case there would be no need to predict areas susceptible to landsliding.

Conclusions
Generalised linear modelling is proposed as a technique for predicting partially the relative likelihood of landsliding within an area of the central Appennines, Italy. Particular attention has been drawn to identifying independent variables that relate to conditions prior to the occurrence of landsliding.

References
Carrara, A., Cardinali, M., Detti, R., Guzzetti, F., Pasqui, V., and Reichenbach, P. 1991. "GIS techniques and statistical models in evaluating landslide hazard", Earth Surface Processes and Landforms, 16, 427-445.

Chacon, J., Irigaray, C., and Fernandez, T., 1993. "Methodology for large scale landslide hazard mapping in a GIS". In Landslides, Proceedings of the 7th International Conference and Field Workshop, Czech and Slovak Republic, 77-82.

Hansen, A. 1984. "Landslide hazard analysis". In Brunsden, D and Prior, D.B (eds.) Slope instability. Chichester: John Wiley and Son. 523-602.

Siddle, H.J., Jones, D.B., and Payne, H.R. 1991. "Development of a methodology for landslip potential mapping in the Rhondda Valley" In Chandler, R.J. (ed.) Slope Stability Engineering. London: Thomas Telford. 137-142.

Wang, S.-Q., and Unwin, D.J. 1992. "Modelling landslide distribution on loess soils in China: an investigation", International Journal of Geographical Information Systems, 6, 391-405.