2.1 Site of the study area

Al-Marqab river basin is located in the mountainous region in western Syria. The study area extends between 35°7′32′' N and 34°58′39′' N latitudes and between longitudes 36°16′30′' E and 35°58′00′' E with an approximate area of 190 km² (Fig. 1). Geographically, the investigated basin is bounded from the north by the river basins of Basiya and Jobar, from the south by the Hussain River basin, from the east by the Orontes River basin and the west by the Mediterranean.

Close

Fig. 1

Location of the study area.

Export to PPT

2.2 General geology

The geological formations differ from the Jurassic to the Quaternary. Jurassic and Cretaceous structures are composed of marl, limestone, dolomite and sandy limestone. Neogene formations are composed of basalt and sedimentary rocks.

2.3 Climate

Climatically, the basin belongs to the Mediterranean climate according to the Köppen classification (Mohammed et al., 2020), which is characterized by mild rainy winters and hot, dry summers. The average annual rainfall ranges between 900 and 1450 mm. The average annual temperature in August is 20.5 °C, and in January is 10.1 °C. The population of the study area depends economically on agriculture, with other human activities, including tourism and services (Younes et al., 2023; Alsafadi et al., 2022).

3.1 Data

Present investigation assesses landslide susceptibility in Marqab river basin using SVM, RF and XGBoost algorithms depending on the multi-sources dataset (Table 1). A digital elevation model (DEM) obtained from NASA (https://vertex.daac.asf.alaska.edu/, accessed on 18 May 2022) was used to map the topographical layers, including slope angle, slope aspect, curvature, plan curvature, profile curvature, elevation, streams, stream power index (SPI) and topographic Wetness Index (TWI) layers. These layers were extracted from DEM relying on Surface Analysis and Hydrology tools in ArcMap v. 10.4.1 software. To produce the Normalized Difference Vegetation Index (NDVI) layer, the Sentinel 2A imagery obtained from The European Space Agency (ESA) was used (https://www.esa.int/, accessed on Jul 7, 2022). Precipitation data was gathered from the meteorological corporation in Damascus Governorate. In this respect, the interpolation tools process was utilized to construct rainfall layer. Road information was acquired from the Syrian Public Transport Authority. Faults and geological information were gained from General Geological Authority in Syria. The distance from the faults and road network was analysed by utilising the Euclidean distance tool in ArcMap v. 10.4.1. Soil depth information was obtained from the Ministry of Agriculture in Damascus Governorate (MoA, 2023). The Landuse/Landcover (LULC) map was created based on Global 10 m LULC Datasets developed by ESRI.

3.2 Landslide inventory

In the current assessment, historical records, field observations, aerial images and satellite images were utilized to produce a detailed inventory map of landslide events. A total of 200 landslide sites were specified and overlayed. The same number of non-landslide points as landslide sites were also specified randomly for training the models. Landslide events were randomly classified into a training dataset of 140 points (70 %), and a test dataset of 60 points (30 %).

3.3 Predisposing factors

At the present evaluation, sixteen causative factors were determined according to the related literature and geographical specificity, namely: slope degree, slope aspect, curvature, profile curvature, plan curvature, elevation, SPI, TWI, proximity to road, proximity to stream, proximity to faults, LULC (May 2022), rainfall (Perioed: 1992 – 2023), soil depth, lithology and NDVI.

3.4 Multicollinearity assessment

Using an extensive number of affecting factors in susceptibility modelling does not imply the precision of outputs and may eventually lead to significant deviation in predicted outputs (Zeng et al., 2023). Thus, evaluating the importance of influencing factors before the susceptibility modeling process is of great importance to reduce high-dimensional data. In this assessment, the importance of triggering factors was analyzed using a multicollinearity test. The variance inflation (VIF) and tolerance (TOL) are calculated to determine the level of multicollinearity (Eq. (1).

3.5 ML algorithms

Utilizing ML algorithms is considered one of the most useful manners to produce prediction maps for natural hazards, especially landslides. These algorithms suggest susceptibility modelling with greater accuracy compared to traditional statistical methods. In this evaluation, inventory data with driving factors were used to carefully train ML models. The following subsections present the details of the selected algorithm.

3.6 Performance assessment of algorithms

Receiver operating characteristic (ROC) curve is utilized for assessing the modelling precision. The area under the curve (AUC) is employed to illustrate the capacity of the proposed model to predict the accuracy of conducted landslide susceptibility modelling (Yao et al., 2023). An AUC value close to 1 indicates a high reliability of the model (Basharat et al., 2023). Additionally, a set of statistical indexes were utilised to consider the accuracy of performance, including precision (Eq. (2), recall (Eq. (3), F1 score (Eq. (4), accuracy (ACC) (Eq. (5), and Matthew’s correlation coefficient (MCC) (Eq. (6).

4.1 Multicollinearity assessment

Current results illustrated that all landslide influencing factors have VIF values less than 10, and TOL values greater than 0.1. The lowest TOL and greatest VIF values were 0.591 and 1.727, respectively (Table. 2). Consequently, there is no multicollinearity among factors. Thus, all of the factors were assessed in the modelling phase.

4.2 Landslide influencing factors

In this assessment, sixteen raster layers were produced representing the criteria used in landslide susceptibility modeling. The current results indicate a considerable spatial variation in the slope degree values in Al-Marqab river basin. However, the steep slopes are mainly concentrated along the river beds. The slope angle layer was classified into four categories: <10°, 10°–20°, 20°–30°, and <30° as reported in Fig. 2a. Fig. 2b shows the slope aspect layer which was classified into nine directions.

Close

Fig. 2

Maps of causative factors (a) slope angels, (b) slope aspect, (c) curvature, (d) profile curvature, (e) plan curvature, (f) elevation, (g) NDVI, (h) LULC, (i) rainfall, (j) TWI, (k) SPI, (l) lithology, (m) soil depth, (n) proximity to faults, (o) proximity to streams, (p) proximity to roads.

Export to PPT

Referring to the spatial distribution of negative, flat and positive curvatures, Fig. 2c, d and e illustrate the maps of curvature, profile curvature and plan curvature, respectively. The eastern regions are characterized by great elevation exceeding 700 m, as indicated by Fig. 2f, which presents five spatial domains of elevation. The vegetation (wild and agricultural vegetation) and urban covers occupy most area. Fig. 2g and h depict the LULC and NDVI maps, respectively.

It can be observed that most parts of the study basin feature the greatest values of precipitation, mainly due to the influence of mountainous and plateau topography. Fig. 2i, however, describes the five rainfall classes. Stream networks are characterized by the highest values of topographic moisture, driven by the high runoff rate, thus a higher capacity of sedimentary transport.

Fig. 2j and k show the spatial distribution of TWI and SPI values, respectively. Referring to geological units, the Jurassic and Cretaceous lithological formations cover the largest area of the study area as Fig. 2l illustrates. Moderate shallow soil is spread in most parts, while the eastern parts feature very shallow soil as Fig. 2m shows. The extracted distance from faults, roads and streams was organised into four as depicted in Fig. 2n, o and p, respectively.

4.3 Significance of conditioning factors

Fig. 3 shows the importance of landslide contributing factors in each algorithm used. All models used demonstrated that LULC is the most significant contributing factor for landslide triggering, pursued by curvature factor for the SVM model and the NDVI factor for the RF and XGBoost models. Further, the rainfall and soil depth factors have the least effect on the landslide occurrence according to the XGBoost model, soil depth and curvature factors in the RF model and TWI factor in the SVM model.

Close

Fig. 3

Factors importance of landslide.

Export to PPT

4.4 Landslide susceptibility maps

In this investigation, three ML models were utilised to create three landslide susceptibility maps. These maps were categorised into five zones of susceptibility utilising Natural Breaks method in a GIS environment as Fig. 4a, b and c illustrate. Further, the percentage of the landslide susceptibility categories was computed for statistical investigation of the models used, as Table. 3 shows. According to the results, 24 %, 23 % and 9 % of the investigated basin were classified as very high landslide susceptibility in the SVM, RF and XGBoost models, respectively.

Close

Fig. 4

Landslide susceptibility maps (a) SVM, (b) RF, (c) XGBoost.

Export to PPT

Close

Fig. 4

Landslide susceptibility maps (a) SVM, (b) RF, (c) XGBoost.

Export to PPT

4.5 Validation

Fig. 5 shows the AUC values using the test dataset for the three models. The outcomes indicate that the AUC value for the SVM, XGBoost and RF models was 0.90, 0.94 and 0.96, respectively. Thus, these acceptable AUC results indicate the reliability of the models performance in Al-Marqab river basin. Moreover, the RF algorithm showed the most elevated performance of prediction and classification (AUC = 0.96), indicating provides more objective spatial results than the rest of the models. Moreover, the models accuracy was evaluated using five metrics as Table. 4 illustrated.

Close

Fig. 5

ROC plots of SVM, XGBoost and RF algorithms.

Export to PPT