Landslide Susceptibility Assessment in Constantine Region (NE Algeria) By Means of Statistical Models

Preparing landslide maps is a preliminary step in landslide susceptibility assessment [29]. This allows to control the distribution, extent, types and patterns of landslides in relation to geomorphological, geological and environmental parameters [30]. The inventory map of the study area enumerates 81 slope failures (78 rotational slides, a planar slide, a solifluction and a debris flow) (Fig. 2a). It was carried out depending on the surveys conducted by the Algerian of Highways (ADA), as well as the visual interpretation of air photos and satellite images. During field studies, all landslides were checked and mapped with the aid of a global positioning system (GPS). The slope failures seem to be more frequent in the NE part of the study area, dominated by weak formations. In this research, the landslide inventory was randomly divided into 70% (57 events) for training the model and 30% (24 events) for its validation (Fig. 3) [21].

Figure 2

Figure 3

The geoenvironmental characteristics of an area control the initiation of slope failures in many ways. They could be considered as conditioning factors in the phenomenon prediction. In our research, both predisposition and triggering factors were selected among those most commonly used in landslide susceptibility assessment [31]. The data source used in this study comes from Landsat 7 Enhanced Thematic Mapper Plus (ETM+) images, aerial photographs, geologic and topographic maps, etc. A contour map at 10 m intervals was digitized from (1/25000) topographic map and subsequently employed for generating a digital elevation model (DEM) of the study area. The related physiographic features were directly extracted from the DEM to construct the landslide susceptibility models.

2.2.1

Lithology

Geologic factors play an important role in slope failures and consequently in landslide susceptibility assessment [32]. The lithological map of the study area (Fig. 4a) was defined by the digitization of outcrop formation from “El Aria” geological map (No 74) published by the Algerian Geological Survey Agency (ASGA). The variability of the lithologies in the study area is classified into seven main lithological groups.

Figure 4

In this research, landslide susceptibility mappings (LSMs) were designed by means of landslide inventory and FR, WoE and IV techniques. In order to evaluate the consistency of the three models, area under the ROC curve (AUC) was plotted using the Microsoft Excel software [41]. This method is commonly used to measure how much the model was successful in modelling and predicting the results. All LSMs were classified into five landslide susceptibility classes using Natural Breaks (NB) classifier [14].

The resulted IV, WoE and FR susceptibility maps show close similarities between overall the three methods. The conditioning factors were classified into classes, and weights are presented in columns 7, 10 and 13 of Table 1.

According to Table 1, the relation between landslides and slope gradient shows that most of the landslides were observed for slopes >30°. This means that there exists a good correlation between slope angle and the occurrence of landslides. Counter to the slope, there is no specific correlation between the elevation and the occurrence of landslides. For the slope aspect, the most susceptible classes are N and SW. Regarding the stream density factor, the most suitable class is 0.8–1. In the case of NDVI factor, the susceptibility increases gradually with the decrease in NDVI values. The relation between the lithology and landslide inventory of the study area shows that landslides occurred mainly in (i) Mio-Pliocene and Numidian clay, marl of the Tellian unit and clayey marl of Priabonian, current superficial deposits and colluvium and (ii) the flysh of Numidian unit. Concerning rainfall factor, the classes greater than 550 mm are the most susceptibles. The relation between landslide occurrences and distance to lineaments and streams shows that a large number of the landslides were observed in the area with a distance less than 400 m and 200 m, respectively.

The produced LSMs of the three methods (Fig. 5a–c) were subdivided into five hierarchic classes (very low, low, moderate, high and very high) using the NB classifier. The very high susceptibility zones are located in the NW and SE of the study area spreading out 22% (IV), 19% (WoE) and 9% (FR) of the total area. The moderate susceptibility has a close distribution with 21.59% (IV), 25.17% (WoE) and 29.49% (FR) of the total area. Finally, the low to very low susceptibility is spread in the N, NE and SW of the study area within the average of 24.1% (IV), 27% (WoE) and 39.35% (FR).

In order to verify quantitatively the consistency of the models, the landslide density (LD) for each class was calculated. The LD values reach 3.09 (IV), 3.35 (FR) and 3.38 (WoE) in very high and high susceptibility zones. This confirms the proneness to landsliding in these areas when compared to those of moderate, low and very low susceptibility. The very low susceptible area has the lowest values of LD with 0.0323 (IV), 0.04 (FR) and 0.02 (WoE). According to Table 2, it can be explained that there is a gradual decrease in LD from the high to the low susceptible area.

Figure 5

In order to test the compatibility of the models and to determine their prediction ability, the AUC (success and prediction rates) was plotted. The prediction rates of the FR and WoE methods are in the range of 0.7332–0.7907, indicating a good performance of the models, with a slight differentiation for the WoE model. In the case of the success rates, the IV method shows a low score of 0.5881. This value is less convincing than the two remaining models (Fig. 5d).

From these calculations, we can say that the slope gradient is the most important parameter for slope stability analysis, because the driving force of mass movement increases with increasing slope. In the study area, the gradients over than 25° in soft rocks and/or soils are the most susceptible to landsliding. Because of their physical behaviour, the clay, marl and colluvium formations are more prone to shear stress and therefore most susceptible to landslide occurrence. It is clear that elevation factor influences less the landslide occurrence, against to other factors. This is probably due to the accumulation of weathered materials on low altitudes (corresponding to gentle slopes) and hardness of outcropping formations in high altitudes. The streams network is a further controlling parameters of landslide occurrence, as streams’ erosion may undercut the foot of the slopes and saturate their lower part. Moreover, for the fault proximity, we cannot extract a clear correlation between the nearness to faults and the landslide occurrence. The relation between landslide occurrence and NDVI shows that most of the landslides occur in areas with low values. The relation between aspect and landslide occurrences shows that most of the landslides were observed in areas with N and SW-facing slopes. This is probably due to the more exposure to rainfall coming from these sides.