Introduction

Landslides are one of the most widespread natural hazards on Earth causing casualties and property loses. They are sudden hydrogeomorphic events due to the combination of i) predisposing factors (e.g. lithology and morphology), ii) triggering factors (e.g. excessive and intense precipitations) and iii) accelerating factors (e.g. human activities altering natural slope stability) [1]. Among landslide typologies, shallow landsliding is common on steep topographies covered by colluvial or residual soils whose texture can vary from sand to clayey silt [2], playing a significant role in hill slope denudation and catchment sediment dynamics. The possible time and locations where landslides are likely to occur should thus be identified in advance in order to avoid or reduce the harm. In this regard, establishing a landslide warning system which provides information for evacuation and hazard mitigating is a top priority.

Prediction of rainfall-triggered hill slope disasters has relied mostly on the valley slope [3 and 4], rainfall intensity and duration that can cause hill slope failure [5, 6, 7 and 8]. Recently, theoretical models have been developed to predict landslide susceptibility based on watershed topographic, geologic and hydrological variables as well as changes in landuse [9, 10, 11, 12, 13 and 14]. The rapidly growing availability of relatively detailed digital elevation data, coupled with simple slope-instability mechanism and hill slope hydrological models, has led to advances in physically-based modeling of shallow landslide hazard [9, 15]. Casadei et al., [16] proposed a landslide warming system using a slope-instability analysis and a hydrological model to predict the time and location of landslides that were verified using historical data of landslide events for Montara Mountains of California from 1953–1998 [17].

Given the complex spatial dynamics of the process, most of the hazard evaluation tools make use of Geographical Information Systems (GIS), capable of geospatial and multitemporal data integration. The availability of digital elevation models (DEMs) and precipitation data at higher spatial and temporal resolution [1] encourage the development of more sophisticated techniques for shallow landslide hazard modelling [9, 15, 18, 19, 20, 21 and 13]. Several of these methods are based on the infinite slope stability equation [22] coupled with hydrological models (e.g. [9, 15 and 13]), extensively applied for the estimation of spatially variable soil wetness (e.g. [23, 24 and 25]). SHALSTAB [9 and 26] is one of these models. Various models have been developed in order to assess landslide incidence potential [27, 28, 29 and 30]. Among these techniques, deterministic, heuristic and statistical methods are the most common ones. Deterministic methods deal with the estimation of quantitative values of stability variables, over a defined area where the landslide types are simple with homogeneous intrinsic properties [31]. The required data include soil strength, depth below the terrain surface, soil layer thickness, slope angle and water pressure. These methods have been employed in translational landslides studies [32] and are applicable at large scale over small areas. One of the main drawbacks of these methods is their high degree of oversimplification when the data are incomplete. Another problem is that the data requirements for deterministic models can be prohibitive, and frequently it is impossible to acquire the input data necessary to use the model effectively.

In this context, the present study explores the suitability of pattern recognition techniques to predict the probable distribution of landslide occurrence points based on several environmental layers along with the known points of occurrence of landslides. The aims of the study are:

1. to apply and evaluate Genetic Algorithm for Rule-set Prediction and Support Vector Machine based models for predicting landslides;
2. to compare and assess the results of these analyses;
3. to suggest management strategies for preventing potential losses.

The model utilises precipitation and six site factors including aspect, DEM, flow accumulation, flow direction, slope, land cover, compound topographic index and historical landslide occurrence points. Both precipitation in the wettest month and precipitation in the wettest quarter of the year were considered separately to analyse the effect of rainfall on hill slope failure for generating scenarios to predict landslides.