Landslide susceptibility mapping using machine learning techniques and comparison of their performance at Ziz upper watershed, Southeastern Morocco

doi:10.21203/rs.3.rs-1534262/v1

Download PDF

Research Article

Landslide susceptibility mapping using machine learning techniques and comparison of their performance at Ziz upper watershed, Southeastern Morocco

https://doi.org/10.21203/rs.3.rs-1534262/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Landslides pose a serious threat to humans, their infrastructure and property, especially in mountainous areas. In Morocco, these risks have attracted more attention due to their harms. The current study aims to use machine learning techniques (MLTs) including Multi-Layer Perceptron (MLP), Adaptive Boosting (AdaBoost), Random Forest (RF), Instance-Based K (IBK), Naïve Bayes (NB) and Decision Tree (J48) to model landslide susceptibility and to compare their performance. Initially, 144 landslide sites were inventoried. Then, thirteen factors related to landslides were considered. Finally, the area under the receiver operating characteristic curve (AUC-ROC) method was used to compare models’ performance. The results showed that AUC values for six MLTs range from 53.0% for RF to 96.1% for J48. Both J48 (AUC = 96.1%) and NB (AUC = 89.5%) showed the best performance compared to the others. The results of this work and the generated landslide susceptibility maps can help decision makers to avoid high susceptible areas.

Landslide susceptibility

Machine learning

Geographic information system

Ziz

SE Morocco

Landslides that occur in mountainous regions are considered the most common geohazards (Panchal and Shrivastava 2022; Xing et al. 2021), resulting in a series of material damage and sometimes in human lives (Panahi et al. 2022). The effect of climate change and human activities on the triggering factors of landslides have increased their frequency and magnitude worldwide (Sajadi et al. 2022). The triggering factors for these disasters are numerous including prolonged rainfall, earthquakes, volcanoes and human activities such as the expansion of roads along steep mountainous (Youssef and Pourghasemi 2021). The landslides’ common forms are rockfalls, rock and soil slides, and debris flows.

In southeast Morocco, the mountainous part of Ziz upper watershed (ZUW) has experienced several landslides, especially during rainy periods. These geohazards caused many traffic accidents along the only national road N°13 connecting the two cities of Midelt and Errachidia (Fig. 1). As a result, traffic was interrupted several times due to landslides along the road section between Errich and Errachidia (Fig. 4). However, the literature review showed that few studies on landslides have been conducted and limited to the Rif, in northern Morocco (El Kharim et al. 2021; Es smairi et al. 2021; Benzougagh et al. 2020) and a single attempt by Manaouch et al. (2021) in southeast Morocco. This work therefore fills this gap on landslide studies and compares its results with the only previous attempt of which they used the classical methods.

Due to the available and free geospatial data and the development of computer science, several approaches to studying landslides have been proposed. Most are based on geographic information systems (GIS) and remote sensing data (Chang et al. 2020). However, these approaches require knowledge, determination and preparation of factors related to landslides (topographic, geomorphological, pedological, environmental and physical) which are not always easy. In general, these approaches are based on subjective expert judgments (qualitative) or on objective mathematical methodologies (quantitative) (Bopche et al. 2022). For Dai et al. (2002), these approaches are classified into three major groups: heuristic, deterministic and statistical.

During the last decade, some authors have introduced machine learning techniques (MLT) in the spatial modeling of geohazards, for example landslide susceptibility (Saha et al. 2022; Hussain et al. 2022; Ghorbanzadeh et al. 2019; Zhang 2021), land subsidence (Zhou et al. 2020), slope failure (Zhang et al. 2021).

For landslides, Tien Bui, Zhang, Zhou, Arabameri, Pradhan, Lee and Pham and others first tried to introduce machine learning algorithms: Lee and Pradhan (2007); Arabameri et al. (2020); Pham et al. (2021); Zhou et al. (2020); Chen et al. (2021). Youssef and Pourghasemi (2021) reported that these techniques are advantageous because of their objective statistical basis, their capacity for quantitative analysis, their reproducibility and their potential for continuous updating. Moreover, Goetz et al. (2015) recommended that for accurate modeling, MLTs should be integrated with the processing capabilities of GIS and field data (e. g., historical landslide inventory and geo-environmental factors).

Recently, the use of MLTs in the study conducted by Youssef and Pourghasemi (2021) has shown that they can adjust their internal structures to current landslide data, they are able to extract relationships between big data automatically, and they can quickly build both classifications and regressions to provide accurate models applied at large-scale and more cost effective than classical methods.

For the above reasons, six MLTs were tested including Random Forest (RF), Multi-Layer Perceptron (MLP), adaptive boosting (AdaBoost), Decision Tree J48 (J48), Instance-Based K (IBK) and Naive Bayes (NB). The results of these landslide susceptibility modeling could provide valuable input to the local scientific community that has not yet tried these techniques. Additionally, landslide susceptibility maps can accurately delineate landslide-prone areas, so decision makers can avoid unfavorable locations for future development projects, such as new highways, roads, and urban areas.

Ziz upper watershed (ZUW) covers an area of ~ 4435 km2 within the Draa-Tafilalet region. It is located at the southeastern portion of Morocco between 32° 5’ 48” N to 32° 64’ 19” N, and 4° 11’ 72” W to 5° 46’ 20” W (Fig. 1). ZUW has an elongated shape and dissected by many wadis that drain their waters towards the main Oued of Ziz which runs in the middle towards east then towards south direction.

Recently, the Draa-Tafilalet region has seen an enormous increase in tourist activities and an unprecedented expansion of date palm farms. However, the only national road N°13 linking the two towns of Midelt and Errachdidia has experienced landslides several times (Fig. 3). As a result, these landslides along this road pose a threat to the community, infrastructure and traffic.

Topographically, ZUW’s altitude ranges from 1023 to 3687 m above mean sea level, some of these areas have steep topographic reliefs and the dominant slope class (80%) is 0–15° (Mohamed et al. 2021). In general, precipitations in the Draa-Tafilat region are rare, uncommon and usually occur in autumn and spring. However, their spatial variation in ZUW is high due to the basin’s topography: the western and northwestern parts are mountainous and more wet while the southern and eastern parts are less wet, almost desert due to the Saharan influence, by the very drying winds of the "Chergui" type blowing from the northeast in summer and those of the "Sahéli" type blowing from the southwest in winter (Navas et al. 2013).

According to several field visits (October 2020 and Mars 2021), rainstorms are the main factor causing landslide in some ZUW sites. The annual precipitation for ZUW ranges from 119 to 377 mm (Mohamed et al. 2020). The high rainfall rates in the western parts are largely related to the high elevations that induce orographic rainfall precipitation.

According to the ZUW's geological map (Fig. 2) developed by combining the two geological maps of: Midelt at 1/2000,000 and Rich-Boudnib at 1/200,000. Triassic terrains are composed of detrital deposits, doleritic basalts and sometimes evaporites. They rest in angular unconformity on a Palaeozoic pedestal deformed and structured by several tectonic phases (Michard 1976; Charire 1990; Hinaje 1996). The Jurassic series form an important part of the Mesozoic terrains of the High Atlas of which ZUW belongs. They lie concordantly on the red formations of the Lower Triassic-Liassic. Their lithological nature is essentially dolomites, limestones, marl-limestone alternations and silico-clastic detrital deposits (Nouayti et al. 2015).

On the structural setting, the NW-SE oriented extension regime, which reigned during the Triassic-Jurassic, controlled the formation of the Atlas basins of which ZUW belongs. The rifting is initiated from the Triassic, while the individualization of these basins and their contemporary subsidence took place during the Jurassic. The paleogeographic and morpho-structural situation was dominated by a system of blocks tilted essentially towards the NW or towards the SE. They are bounded by major normal NE-SW to ENE-WSW faults separated by WNW-ESE transfer faults (Nouayti et al. 2015). In the Tertiary, a phase of bi-episodic compression, first NW-SE then N-S, takes up the old normal NE-SW faults in reverse faults or in sinistral strike-slip thrusts. This compression, which initiates the closure of the Atlas basins from the Upper Eocene, materializes during the Oligo-Miocene and the Plio-Quaternary (Michard 1976).

The proposed approach in this work is based on the use of six machine learning classifiers (MLC) for landslide susceptibility modeling. Figure 3 illustrates the overall workflow of the current study.

3.1 Data sources

Several data sources include field surveys, previous studies and national reports were collected to identify the previous landslide events and their frequencies. Moreover, satellite images including those of Landsat 8 Operational Landsat Imager (Landsat 8 OLI) (acquired in April 2020), spatial resolution of 30 m and professional Google Earth data, and DEM of SRTM downloaded from the USGS website (spatial resolution 30 m) were used. Geological maps of Rich and Boudnib with a scale of 1:200 000 were downloaded, digitized, georeferenced and assembled. Finally, these data sets were projected according to UTM-Zone 30, WGS84 Datum.

3.2 Landslide inventory map

To study the susceptibility to landslides, two types of data are needed. The first is the landslide inventory map and the second is related to landslide triggering factors. Relationships between inventory data and triggering factors should be assessed to determine their significance and effect on the occurrence of landslides, since it is generally assumed that landslides will occur under the same conditions as before. In the current study, an inventory map of existing landslides was prepared by integration of the historical records, field investigations, and data extracted from satellite images interpretation. Some field surveys were carried out to verify the historical and landslide extracted from high resolution images and to collect new landslides in the study area. A total of 144 landslides were used which distributed all over the study area (Fig. 1).

According to previous landslide susceptibility studies, the percentages commonly used to divide the inventory dataset are 70% and 30% for the training and validating datasets, respectively. Therefore, the landslide datasets were randomly divided into two groups, 70% for model training and 30% for model testing (Fig. 3). The validation process was executed by comparing the existing landslide locations (validating group) with the acquired landslide susceptibility maps.

3.3. Landslide triggering factors (LTFs)

To achieve high modeling accuracy of landslide susceptibility, selecting and preparing the LTFs database is a crucial step. According to previous studies, many researchers indicated that precipitations and earthquakes are the external triggering factors of landslides (Goetz et al. 2015). Moreover, others mentioned that they depend on the characteristics of the study area (known as internal factors) (Youssef and Pourghasemi 2021). However, Alqadhi et al. (2022) stated that knowledge of optimal conditioning factors should be conducted.

Related to the study area and field investigations, thirteen LTFs were involved including, slope-angle, slope aspect, slope length (LS), distance from roads, geology, distance from streams, distance from faults, elevation, rainfall, erodibility, plan curvature, profile curvature, and land use/land cover (LULC). Accordingly, a database of 13 geo-environmental factors were generated using GIS software (ArcGIS 10.5). Thirteen thematic layers of the same spatial resolution have been prepared (Fig. 5). All layers have UTM coordinate system zone 30 with a Datum of WGS 84. LTFs have different types: nominal (categorical) where data do not have a natural order or ranking such as (geology, slope aspect, and land use/ land cover) and numeric (continuous data) in which the order matters such as (elevation, slope angle, slope length, plan-curvature, profile- curvature, distance from faults, distance from roads, and distance from streams).

3.3.1 Slope angle

The slope or the degree of slope inclination is a major factor in the reproduction of geological hazards in mining areas (Su et al. 2021). The higher the degree of inclination of the slope, the larger the number of the landslides (Youssef and Pourghasemi 2021). Thus, with the increase in slope angle, the gravity's force on the slope also increases. Therefore, the greater the slope, the stronger the anti-weathering ability of the slope rock. The slope angle map of the study area was derived from the DEM-SRTM with a resolution of 30 m, ranging from 0° to 66.08° (Fig. 5a).

3.3.2 Slope aspect

Slope-aspect can influence directly or indirectly landslides by its impact on various process including the surface distribution of solar radiation (Alghamdi and Abdel-Mottaleb 2021), wind directions, precipitation patterns, discontinuities orientations, hydrological processes, evapotranspiration, concentration of the soil moisture, vegetation, and root development. The study of Capitani et al. (2013) showed that slope aspect significantly influenced the distribution of shallow landslide types but did not influence other landslide types. The slope aspect map was generated using DEM-SRTM and classified into nine categories including flat (-1), North (0°–22.5°; 337.5°–360°), Northeast (22.5°–67.5°), East (67.5°–112.5°), Southeast (112.5°–157.5°), South (157.5°–202.5°), Southeast (202.5°–247.5°), West (247.5°–292.5°), and Northeast (292.5°–337.5°) (Fig. 5b).

3.3.3 Slope length (LS)

LS is a topographical parameter that greatly influences the susceptibility to landslides. this factor is incorporated into the slope angle and affects soil loss and hydrological processes of mountainous areas (Pourghasemi and Rahmati 2018). In the current study, the LS factor map, ranging from 0 to 432 m, as shown in Fig. 5i, was extracted from the DEM using ArcGIS 10.5 software according to the following formula of Wischmeier and Smith, developed by Bizuwerk et al. (2003):

LS = (L/22,13)^m (65,41 Sin²(S) + 4,56 Sin(S) + 0,065) (1)

Where: L is the Slope length (m); L = flow accumulation × DEM spatial resolution and S is the slope gradient (in %); m: Constant which is equal to: 0.5 for slopes greater than 5%; 0.4 for slopes of 3.5 to 5%; 0.3 for slopes of 1 to 3.5% and 0.2 for slopes of less than 1%.

3.3.4 Distance from roads

Road construction works in mountainous areas are accompanied by engineering activities such as cutting or excavating slopes resulting in changes in the original geological situations and the weakening of the natural support of the rocky slopes. These anthropogenic activities are an important agent that influences the topography of natural slopes (Xiao et al. 2019). Thus, these works will increase landslides occurrence (Wang et al. 2016). Consequently, proximity to roads can be a potential indicator of landslides. In the current study, the proximity to road map was developed using the Euclidean distance tool in ArcGIS 10.5 (Fig. 5e).

3.3.5 Distance from streams

The extraction of the drainage network (the main streams) was developed using the DEM of the study area. In fact, runoff along wadis plays a crucial role in undermining phenomena and increasing pore pressure of areas adjacent to wadis and initiating landslides (Youssef and Pourghasemi 2021). Therefore, it becomes an essential triggering factor of landslide susceptibility. The map of proximity to streams was generated using the Euclidean distance tool in ArcGIS 10.5 (Fig. 5f).

3.3.6 Distance from faults

Faults greatly affect the surface stability of a given area and their strictures control the distribution, number and magnitude of geohazards. Using GIS platform, the two maps of the geological structure of Rich and Boudnib at 1/200,000 were vectorized, georeferenced, cut and assembled. The fault map of ZUW was generated and illustrated in Fig. 5g above. The map of proximity to faults (Fig. 5g) was developed using the Euclidean distance tool in ArcGIS 10.5.

3.3.7 Lithology

Each lithology is subject to a type of geohazards. Landslides can occur on different lithologies than where land subsidence can occur (Youssef and Pourghasemi 2021). These lithologies vary in their physical and mechanical characteristics, including type, strength, degree of weathering, durability, density and permeability (Henriques et al. 2015).

In the current study, the lithology was extracted from the geological map (scale 1:200,000) acquired from the Geological Service Office of Morocco. According to stratum age, eight lithological units have been identified, including Trias, Aalenian, Toarcian, Domerian, Pliensbachian, Sinemurian, Batho-Bajocian and Quaternary (Fig. 2).

3.3.8 LULC

According to previous studies, the area of rangeland and degraded forest in ZUW accounts for 60.31% and 20% of the total region, respectively. On the one hand, these two types of land cover do great damage to the surface, affecting the occurrence of landslide hazards; and on the other hand, the absence of land maintenance and a large part of ZUW is almost bare provides certain conditions for the occurrence of geological hazards. LULC map of the study area is illustrated in Fig. 5m.

3.3.9 Rainfall

Precipitation is an essential landslide conditioning factor. According to many field surveys, it is observed that rainfall is the main triggering factor to landslide in ZUW.

According to data recorded over the past 36 years at four meteorological stations all located in the study area, the rainfall map was obtained using the ArcGIS 10.5 platform and is shown in Fig. 5h.

3.3.10 Elevation

Many studies showed that slope instability is influenced by many topographic factors including the elevation (Feizizadeh et al. 2014). This factor has been commonly involved in landslide susceptibility predictions. In this study, elevation map was processed according to classification of the study area's DEM. The altitude values ranged between 1023 m and 3687 m above sea level (Fig. 5c).

3.3.11 Plan and Profile curvature

Slope shape and terrain morphology influence the landslide susceptibility in different ways (Haigh and Rawat 2012). Evans (1979) indicated that the curvature is one of the principal terrain variables which would be used in different geomorphological analyzes. For instance, the convergence and dispersion of surface runoffs is affected by plan curvature (Nasiri Aghdam et al. 2016). While the profile curvature influences the deposition rates of materials by controlling the transport of these materials on slope (Xiao et al. 2019).

In this work, plan and profile curvatures were generated from the ZUW’s DEM using the Spatial analyst tool of ArcGIS 10.5. The plan and profile curvatures values ranged between − 5.1 to 7.8 and − 7.3 to 7.9, respectively (Fig. 5j and 5k).

3.3.12 Erodibility index (K)

Erodibility index estimates the susceptibility of particles from a soil surface to detachment under the effect of raindrops or surface runoff or a combination of the two. Erodibility rate depends on the soil's physical characteristics, in particular the percentage of organic matter, clay, structure and permeability.

The ZUW’s erodibility map (Fig. 5l) used in this work is based on previous study carried out by Fenjiro et al. (2020).

3.4 Normalization of LTFs

Landslide’s triggering factors database aims to select and analyze the relation between landslide parameters. Generally, all the parameters are normalized using different approaches such frequency ratio (Zhu et al. 2021). This approach defines the area ratio wherever landslides have occurred in the whole study area for a period (Aditian et al. 2018). The frequency ratio is defined by the ratio of landslide occurrence percentage to the area occupation percentage for different classes of each triggering factor, as given in the Eq. (2). According to Samanta et al. (2018), the more the frequency ratio increases, the more the chances of landslides occurrences increase.

$$F{R}_{i}=\frac{{N}_{i}∕N}{{s}_{i}∕s}$$

Where FRi is the frequency ratio of the i^th class of a triggering factor; Ni and Si, respectively, represent the number of landslides and the size of the area in the i^th class of this triggering factor; N is the total number of landslides in the study area; and S is the total study area.

The FR values of the 13 triggering factors (each with several classes) are calculated and presented in Table 1 and Table 2. Using the normalize filter method, the normalized value for each class of the LTFs were rescaled from 0.1 to 0.9. A normalized value close to 1 recommends a more distinguished association between landslides and the factor. However, a value close to 0 symbolizes a weaker association.

Table 1

Spatial relationship between LTFs and landslides using frequency ratio (FR) and normalized classes
LTF	Class	Number of pixels	% of pixels	Number of landslides	% of landslides	Frequency Ratio
LTF	Class	Number of pixels	% of pixels	Number of landslides	% of landslides	FRi	FRnormalized
Slope	Gentle (0–15°)	1807422	0,37	11	0,08	0,21	0,10
	Moderate(15–30°)	1249754	0,25	21	0,15	0,57	0,16
	Steep (30–45°)	881213	0,18	28	0,19	1,08	0,23
	Extreme (45–60°)	671457	0,14	36	0,25	1,83	0,35
	Very strong (> 60°)	303773	0,06	48	0,33	5,39	0,90
Aspect	Flat (-1-40)	505437	0,10	20	0,14	1,35	0,73
	N (40–80)	414655	0,08	5	0,03	0,41	0,10
	NE (80–120)	531783	0,11	7	0,05	0,45	0,13
	E (120–160)	754482	0,15	30	0,21	1,36	0,73
	SE (160–200)	772236	0,16	36	0,25	1,59	0,89
	S (200–240)	456295	0,09	9	0,06	0,67	0,27
	SW (240–280)	398328	0,08	9	0,06	0,77	0,16
	W (280–320)	467495	0,10	11	0,08	0,80	0,16
	NW (320–360)	617270	0,13	17	0,12	0,94	0,45
Elevation	1	393526	0,08	18	0,13	1,56	0,47
	2	850896	0,17	21	0,15	0,84	0,29
	3	837542	0,17	9	0,06	0,37	0,17
	4	885384	0,18	6	0,04	0,23	0,14
	5	655483	0,13	14	0,10	0,73	0,26
	6	552065	0,11	29	0,20	1,79	0,53
	7	409303	0,08	22	0,15	1,84	0,54
	8	240897	0,05	23	0,16	3,26	0,90
	9	92885	0,02	2	0,01	0,74	0,27
Geologic unit	Sinemurian	360359	0,07	33	0,23	3,19	0,90
	Quaternary	597887	0,12	7	0,05	0,41	0,17
	Batho-Bajocian	2317137	0,46	103	0,72	1,55	0,47
	Domerian	268712	0,05	1	0,01	0,13	0,10
	Pliensbachian	532940	0,11	0	0,00	0,00	0,10
	Trias	60809	0,01	0	0,00	0,00	0,10
	Aalenian	423867	0,08	0	0,00	0,00	0,10
	Toarcian	454640	0,09	0	0,00	0,00	0,10
Distance to road	1	857267	0,17	47	0,33	1,87	0,90
	2	707668	0,14	31	0,22	1,50	0,72
	3	634743	0,13	24	0,17	1,29	0,62
	4	549073	0,11	30	0,21	1,87	0,90
	5	2170701	0,44	12	0,08	0,19	0,10
Distance to stream	1	147774	0,03	4	0,03	0,92	0,13
	2	134610	0,03	4	0,03	1,02	0,14
	3	143612	0,03	5	0,03	1,19	0,15
	4	100359	0,02	5	0,03	1,70	0,18
	5	4393097	0,89	126	0,88	0,98	0,14
Distance to fault	1	927639	0,19	71	0,49	2,61	0,90
	2	854753	0,17	33	0,23	1,32	0,47
	3	814911	0,17	19	0,13	0,80	0,29
	4	739301	0,15	10	0,07	0,46	0,18
	5	583038	0,12	4	0,03	0,23	0,10
	6	999810	0,20	7	0,05	0,24	0,10
Rainfall	1	1906686	0,39	46	0,32	0,82	0,21
	2	2118988	0,43	9	0,06	0,15	0,10
	3	410111	0,08	47	0,33	3,92	0,70
	4	224347	0,05	34	0,24	5,18	0,90
	5	259320	0,05	8	0,06	1,05	0,24
Length of slope	1	21828	0,00	142	0,99	222	0,90
	2	76772	0,02	0	0,00	0,00	0,10
	3	1015571	0,21	0	0,00	0,00	0,10
	4	1949329	0,40	0	0,00	0,00	0,10
	5	1855923	0,38	2	0,01	0,04	0,10
Plan of curvature	1	249457	0,05	138	0,96	18,9	0,90
	2	974262	0,20	0	0,00	0,00	0,10
	3	2332702	0,47	1	0,01	0,01	0,10
	4	1101346	0,22	2	0,01	0,06	0,10
	5	260214	0,05	3	0,02	0,39	0,12
Profile of curvature	1	208449	0,04	12	0,08	1,97	0,89
	2	774445	0,16	28	0,19	1,23	0,45
	3	2241836	0,46	42	0,29	0,64	0,10
	4	1348631	0,27	42	0,29	1,06	0,35
	5	344620	0,07	20	0,14	1,98	0,90
Erodibility (K)	1	23208	0,46	100	0,69	1,51	0,40
	2	620	0,01	7	0,05	3,95	0,90
	3	17820	0,35	36	0,25	0,71	0,24
	4	8724	0,17	1	0,01	0,04	0,10
Land use / Land cover (LULC)	Agricultural fields	176032	0,04	1	0,01	0,19	0,10
	Degraded forest	450223	0,09	21	0,15	1,59	0,90
	Rangeland	4245821	0,87	122	0,85	0,98	0,55
	Water bodies	28178	0,01	0	0,00	0,00	0,10

Table 2

Normalized classes of non-landslides conditioning factors used in this study.
LTF	Class	Number of pixels	% of pixels	Number of landslides	% of landslides	Frequency Ratio
LTF	Class	Number of pixels	% of pixels	Number of landslides	% of landslides	FRi	FRnormalized
Slope	Gentle (0–15°)	1807422	0,37	17	0,12	0,32	0,10
	Moderate(15–30°)	1249754	0,25	127	0,88	3,47	0,90
	Steep (30–45°)	881213	0,18	0	0,00	0,00	0,10
	Extreme (45–60°)	671457	0,14	0	0,00	0,00	0,10
	Very strong (> 60°)	303773	0,06	0	0,00	0,00	0,10
Aspect	Flat (-1-40)	505437	0,10	12	0,08	0,81	0,16
	N (40–80)	414655	0,08	13	0,09	1,07	0,40
	NE (80–120)	531783	0,11	25	0,17	1,61	0,90
	E (120–160)	754482	0,15	21	0,15	0,95	0,29
	SE (160–200)	772236	0,16	22	0,15	0,97	0,30
	S (200–240)	456295	0,09	10	0,07	0,75	0,10
	SW (240–280)	398328	0,08	10	0,07	0,86	0,20
	W (280–320)	467495	0,10	13	0,09	0,95	0,29
	NW (320–360)	617270	0,13	18	0,13	1,00	0,33
Elevation	1	393526	0,08	28	0,19	2,43	0,90
	2	850896	0,17	34	0,24	1,36	0,54
	3	837542	0,17	33	0,23	1,35	0,53
	4	885384	0,18	20	0,14	0,77	0,33
	5	655483	0,13	26	0,18	1,35	0,53
	6	552065	0,11	2	0,01	0,12	0,11
	7	409303	0,08	1	0,01	0,08	0,10
	8	240897	0,05	0	0,00	0,00	0,10
	9	92885	0,02	0	0,00	0,00	0,10
Geologic unit	Sinemurian	360359	0,07	2	0,01	0,19	0,10
	Quaternary	597887	0,12	20	0,14	1,17	0,70
	Batho-Bajocian	2317137	0,46	99	0,69	1,49	0,90
	Domerian	268712	0,05	2	0,01	0,26	0,14
	Pliensbachian	532940	0,11	0	0,00	0,00	0,10
	Trias	60809	0,01	0	0,00	0,00	0,10
	Aalenian	423867	0,08	8	0,06	0,66	0,39
	Toarcian	454640	0,09	13	0,09	1,00	0,60
Distance to road	1	857267	0,17	127	0,88	5,06	0,90
	2	707668	0,14	12	0,08	0,58	0,18
	3	634743	0,13	4	0,03	0,22	0,13
	4	549073	0,11	0	0,00	0,00	0,10
	5	2170701	0,44	1	0,01	0,02	0,10
Distance to stream	1	147774	0,03	60	0,42	13,9	0,90
	2	134610	0,03	16	0,11	4,06	0,32
	3	143612	0,03	13	0,09	3,09	0,26
	4	100359	0,02	8	0,06	2,72	0,24
	5	4393097	0,89	47	0,33	0,37	0,10
Distance to fault	1	927639	0,19	10	0,07	0,37	0,10
	2	854753	0,17	11	0,08	0,44	0,13
	3	814911	0,17	27	0,19	1,13	0,41
	4	739301	0,15	51	0,35	2,36	0,90
	5	583038	0,12	22	0,15	1,29	0,47
	6	999810	0,20	44	0,31	1,50	0,55
Rainfall	1	1906686	0,39	48	0,33	0,86	0,25
	2	2118988	0,43	51	0,35	0,82	0,23
	3	410111	0,08	23	0,16	1,92	0,59
	4	224347	0,05	19	0,13	2,89	0,90
	5	259320	0,05	3	0,02	0,40	0,10
Length of slope	1	21828	0,00	137	0,95	214	0,90
	2	76772	0,02	0	0,00	0,00	0,10
	3	1015571	0,21	1	0,01	0,03	0,10
	4	1949329	0,40	2	0,01	0,04	0,10
	5	1855923	0,38	4	0,03	0,07	0,10
Plan of curvature	1	249457	0,05	0	0,00	0,00	0,10
	2	974262	0,20	4	0,03	0,14	0,10
	3	2332702	0,47	103	0,72	1,51	0,90
	4	1101346	0,22	33	0,23	1,02	0,61
	5	260214	0,05	4	0,03	0,52	0,32
Profile of curvature	1	208449	0,04	0	0,00	0,00	0,10
	2	774445	0,16	4	0,03	0,18	0,10
	3	2241836	0,46	103	0,72	1,57	0,90
	4	1348631	0,27	33	0,23	0,84	0,48
	5	344620	0,07	4	0,03	0,40	0,23
Erodibility (K)	1	23208	0,46	100	0,69	1,51	0,90
	2	620	0,01	0	0,00	0,00	0,10
	3	17820	0,35	38	0,26	0,75	0,42
	4	8724	0,17	6	0,04	0,24	0,10
Land use / Land cover (LULC)	Agricultural field	176032	0,04	106	0,74	20,5	0,90
	Degraded forest	450223	0,09	24	0,17	1,81	0,17
	Rangeland	4245821	0,87	11	0,08	0,09	0,10
	Water bodies	28178	0,01	3	0,02	3,62	0,24

3.5 Modelling procedure using MLTs

The following paragraphs describe the machine learning classifiers used in this study.

3.5.1 AdaBoost

According to Freund and Schapire (1997), the AdaBoost algorithm contains a series of individual classifiers which are generated iteratively, and each classifier tries to precisely classify the training data thus this algorithm is one of the most widely used boosting algorithms. The classifier uses an adaptive resampling technique in order to choose training samples, i.e., a misclassified dataset produced by a previous classifier is selected more often than correctly classified ones so that a new one classifier can work well in a new dataset. Each iteration assigns a weight to the dataset so that the next integration focuses on reweighted datasets that were misclassified previously.

The final classifier is a weighted sum of the ensemble predictions. The advantage of the AdaBoost algorithm is that it works well for solving two-class problems, multi-class one-label problems, multi-class multi-label problems, single-category problems. label and regression issues (Hong et al. 2018).

3.5.2 Artificial Neural Network (MLP)

Inspired by the biological nervous system of the human brain, Multi-Layer Perceptron are a form of artificial neural networks (ANN), which is a type of models. They can do simulations of complex functions such as decision making and model generation (Liakos et al. 2018). MLPs, like the human brain, consist of a set of processing units called "neurons", which are interconnected. Each unit is a multi-input, single-output nonlinear element (Ding et al. 2013).

Neurons mostly act in parallel and are arranged in multiple layers which feature a data input layer; hidden layer of learning; and an output layer (Liakos et al. 2018).

MLPs could detect complex nonlinear relationships through a learning process that involves adjusting the weighted connections that exist between neurons. After choosing the number of hidden layers and the number of processing units in an individual layer, the ANN starts learning from the training samples (Aditian et al. 2018).

3.5.3 Random Forest

RF is a group of DTs that form an ensemble learning model used for classification and regression problems (Al-Najjar and Pradhan 2021). These models are effective for prediction because they utilize the strength of each tree and their correlations and less sensitive to over-fitting problems. The difference between RF and DT is that a decision tree is built on a whole dataset, utilizing all the variables of interest, while a random forest randomly adopts observations and specific variables to construct multiple decision trees from, and then averages the results. In the present study, samples for landslide and non-landslide events were selected to construct the classification tree (30% of the samples were kept aside from the training and 500 nodes were set as a favorite value).

3.5.4 Naïve Bayes (NB)

The NB classifier is a machine learning algorithm based on Bayes' theorem, which assumes that all learning variables are independent, and that each variable contributes equally to the classification problem (Tsangaratos and Ilia 2016). This algorithm relies on simple iterative parameter estimation schemes; therefore, it is easily developed. According to Tien Bui et al. (2012), NB classifier has good robustness to noise and irrelevant attributes.

According to Khosravi et al. (2018), the NB classifier application passes in four steps: (1) the choice of training variables, (2) the calculation of the posterior probability of each class (Landslides sites/non-landslides sites), (3) estimating the class level along with a covariance matrix and (4) creating a discriminant function for each class level.

3.5.5 Instance-Based K (IBK)

IBK classifier is considered as one of the K-Nearest Neighbors (KNN) classifiers which can select appropriate value of K based on cross-validation and can also do distance weighting.

According to Pedregosa et al. (2011), the classification of a data point by the KNN classifier is based on the properties of neighboring data points. The probability that a data point belongs to such a class is defined by the classification of its nearest neighbors (Bröcker and Smith 2007).

For a set of unclassified points, KNN calculates the distance from each point to find K nearest neighbors. The classification of these neighbors is then used for voting, and the classification with the maximum votes is assigned to the unclassified data point (Abraham et al. 2021).

3.5.6 Decision Tree (J48)

Decision Tree model (J48) is a supervised and nonparametric machine learning algorithm that is operable without prior knowledge about data distribution, with easy interpretation and capability to model as well as it handles the reduction of data complexity and the relationships between variables. Compared to other models, it is a flexible, fast, and robust algorithm that can be used to control the nonlinearity between the input features and discrete classes so that nonlinear relationships between parameters do not affect tree performance. Moreover, DT models are simple to construct and clarify for decision-makers (Kadavi et al. 2019; Al-Najjar and Pradhan 2021).

3.6 Evaluation of LTFs

After the preparation of the landslide inventory map and thirteen causative factor maps, the five algorithms AdaBoost, NB, RF, IBK, MLP and J48 were used to produce the landslide susceptibility maps.

In addition to the 144 landslide sites inventoried in the study area, the same number of points were chosen as non-prone landslide sites (144) to provide essential knowledge on stable or unfavorable conditions of landslide occurrence (Su et al. 2021).

Then, the total of 288 points were randomly separated into two parts: 1) The first one contains 70% data, used for the models training step; 2) The second part including the remaining 30% data adopted to verify these models and confirm their accuracy.

The three measures of accuracy, specificity and sensitivity have been adopted to evaluate performance after applying a dataset. Also, the Receiver Operating Characteristics (ROC) of each model were plotted, and the Areas Under the ROC Curves (AUCs) were obtained. Landslide susceptibility maps were classified into five classes: very low, low, moderate, high, and very high.

3.7 Modelling prediction and performance

Using the trained landslide datasets and saving models, the landslide susceptibility index (LSI) for each pixel in the entire study area was calculated. Then, the produced landslide susceptibility maps (LSMs) were reclassified into five susceptibility classes, including areas of very low, low, moderate, high, and very high susceptibility.

To evaluate the prediction accuracy and performance, the results can be produced quantitatively and graphically by different means. Accordingly, several precision indices (RMSE, Recall, Confusion matrix, …) have been used in landslide susceptibility modelling but many authors reported that AUC is exceptional as the most useful statistical precision index of the overall performance (Lee et al. 2004).

In the current study, the ROC curve is obtained by plotting the true positive rate (sensitivity) on the X-axis and the false positive rate (1-specificity) on the Y-axis. According to Youssef and Pourghasemi (2021), the AUC values can help to extract valuable practical information such as: (i) the quantitative evaluation of the prediction capacity of the models (ii) the identification of the prediction approach or the mathematical functions which give better predictions, (iii) comparing the predictive capability of multiple landslide susceptibility models produced by different methodologies and (iv) the evaluation of the capacity of each model to discriminate the most prone zones and the least susceptible areas to landslide (the safest areas).

4.1 LTFs contribution

To assess the importance of LTFs, an attribute evaluator called “CorrelationAttributeEval (correlation attribute evaluation)” was used to obtain the predictive capability of the thirteen landslide triggering factors. The importance of each factor is shown as a percentage that appeared in x-axis (Fig. 6). The results from the CorrelationAttributeEval indicated that erodibility (gives 74.7%) is the most important contribution factor, followed by profile curvature (62.5%), geology (51.3%), distance from streams (48.4%), followed by slope aspect (44.2%). However, Distance to faults (5.4%) and LULC (1.3%) seemed to have relatively less importance for landslide susceptibility modeling.

4.2 Landslide Susceptibility Models

In order to achieve the comparative analysis, the six landslide susceptibility maps provided by the RF, NB, J48, MLP, AdaBoost and IBK classifiers are presented in Fig. 7A, 7B, 7C, 7D, 7E and 7F. Five susceptibility classes, namely very low, low, moderate, high and very highly susceptible areas were assigned for each landslide susceptible classifier. The relative area percentages of these classes for each model were then computed.

4.3 Accuracy assessment and comparison

To assess the model performance, the AUC method was calculated. Interestingly, according to AUC values (Fig. 8), the variation in classifiers’ performance among the MLCs was very different. The J48 (AUC = 96.1%) and NB (AUC = 89.5%) had the highest performances. They are followed by the IBK (AUC = 60.5%), AdaBoost (AUC = 60.0%), MLP (AUC = 59.7%) and RF (AUC = 53.0%). Both J48 and NB exposed a sufficient performance (AUC > 70%).

To model the spatial distribution of certain natural phenomena such landslides, there are two main components which must be distinguished: the study area's characteristics and the approach used. For the study area's variables (scale, geo-environmental conditions, human activities and climate change), they affect the modeling results. For the approach, the choice between heuristic, deterministic or statistical must be succussed.

Regarding the characteristics of the study area, Pourghasemi and Rahmati (2018) concluded that topography and geomorphology play a major role. Moreover, other authors stated that if all variables are considered simultaneously, the modelling results will be affected. Recently, Dornik et al. (2022) reported that the sampling design on landslide susceptibility modeling in lithologically heterogeneous areas affects the model outputs.

As for the approaches, the MLTs identify automatically the relationships between dependent (LTFs) and independent variable (landslide occurrence) while statistical models are slow (Yilmaz 2010). For Youssef and Pourghasemi (2021), they reported that the accuracy of landslide susceptibility mapping depends on the right choice among the wide variety of MLTs. For Zhou et al. (2018), landslide susceptibility modeling should be done by support vector machine (SVM), and ANN given their accurate predictions.

Whereas Tien Bui (2012) stated that one of the strengths of MLTs is that they are relatively easy to apply, and their prediction accuracy often exceeds qualitative techniques such as AHP.

In this work, the evaluation of the susceptibility to landslides by six different models gave different results. According to their performance, the order is as follows: J48, NB, IBK, AdaBoost, MLP then RF. Obtaining NB before RF is totally different from what several authors have obtained including Imtiaz et al. (2022), who applied RF (89–90.5%), SVM and NB (86–87%); Liu et al. (2022), who applied convolutional neural network (CNN), RF (89.92%), LR and SVM; Sajadi et al. (2022), who applied Deep Neural Network (DNN), RF(AUC = 97.9%), NB (AUC = 92.9%) and SVM.

According to the analysis of the maps of susceptibility to landslides obtained by the six models, the conclusions drawn are like those already concluded by certain authors. These maps could not be used for specific planning and evaluation purposes (Lee et al. 2017). This point requires a great deal of attention when using these models for the characterization of specific sites, and the scale of the analysis must be considered very carefully. Landslide susceptibility models depend on the number and type of variables used to implement them. It was noticed that RF performed well in one specific domain (Catani et al. 2013) but gave poor performance in other domains like this study.

In general, MLTs are very easy to apply for landslide hazard modeling. However, it is essential to understand the differences between MLTs to choose an optimal model for a specific objective and/or field of study (Brenning 2008). In this work, the results indicated that the six MLTs could be of vital importance as preliminary techniques to help decision makers identify vulnerable areas for existing projects and for future new planning areas. Afterwards, areas potentially prone to landslide risk can be identified for detailed investigation. Moreover, the results showed that the six MLTs used did not all generate landslide susceptibility maps with good performance (AUC ranging from 53% for RF to 96.1% for J48). The comparison of the results of this study with those of which Manaouch et al. (2021) applied the AHP method in ZUW shows that the AUC of NB (89.5%) is close to their AUC (88.5%).

In this work, the six advanced MLCs, namely, NB, J48, IBK, AdaBoost, RF and MLP, were applied for landslide susceptibility modelling in ZUW (SE, Morocco) for the first time in Morocco. Moreover, the performances of these classifiers were systematically evaluated and analyzed to select the optimized one for the study area. In general, the following conclusions can be reached:

According to the LTFs analysis based on FR values and corresponding predictive ability, it is noticed that erodibility and slope length have relatively higher predictive capability compared to the remaining factors.

Areas with high susceptibility to landslides are mainly due to an erodibility index (K) of 0.16, slope length < 100m, north and northwest-facing slopes, a geological stratum mainly consisting of limestone, distance 1000–2000 m from faults, distance < 800 m from roads, distance < 200 m from rivers and rangeland and degraded forest (LULC).

Geo-environment marked by poor lithological formation and scarce vegetation cover are the significant cause of landslides in ZUW.

The NB and J48 classifiers have the highest predictions (0.96) and (0.89) compared to other classifiers. Therefore, the J48 classifier is considered as the optimized model in the current study. The NB classifier with a satisfactory prediction rate (0.89) also proved to be a promising model that could be used for landslide susceptibility modeling.

However, further case studies should be conducted to explore the potential and assess the overall performance of six MLCs. Also, it is known that multi-classifiers ensemble methods provide an efficient tool to improve the results of single classifier by combining their results.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgment

This research is carried out using a data provided by the agency of Hydraulic Basin of Guir-Ziz-Rhris (ABH-GZR). The authors would like to thank Mr. N. Khan for his valuable help in validating process.

Abraham MT, Satyam N, Lokesh R, Pradhan B, & Alamri A (2021) Factors Affecting Landslide Susceptibility Mapping: Assessing the Influence of Different Machine Learning Approaches, Sampling Strategies and Data Splitting. Land 10(9):989
Aditian A, Kubota T, & Shinohara Y (2018) Comparison of GIS-based landslide susceptibility models using frequency ratio, logistic regression, and artificial neural network in a tertiary region of Ambon, Indonesia. Geomorphology 318:101-111. https://doi.org/10.1016/j.geomorph.2018.06.006
Alghamdi M, and Abdel-Mottaleb M (2021) A Comparative Study of Deep Learning Models for Diagnosing Glaucoma from Fundus Images. IEEE Access 9, 23894–23906 doi:10.1109/access.2021.3056641
Al-Najjar HA, & Pradhan B (2021) Spatial landslide susceptibility assessment using machine learning techniques assisted by additional data created with generative adversarial networks. Geoscience Frontiers 12(2):625-637
Alqadhi S, Mallick J, Talukdar S, Bindajam AA, Van Hong N, & Saha TK (2022) Selecting optimal conditioning parameters for landslide susceptibility: experimental research on Aqabat Al-Sulbat, Saudi Arabia. Environ. Sci. Pollut. R. 29(3):3743-3762
Arabameri A, Saha S, Roy J, Tiefenbacher JP, Cerda A, Biggs T, & Collins A.L (2020) A novel ensemble computational intelligence approach for the spatial prediction of land subsidence susceptibility. Sci. Total Environ. 726 :138595.
Benzougagh B, Meshram SG, Baamar B, Dridri A, Boudad L, Sadkaoui D, & Mimich K (2020) Relationship between landslide and morpho-structural analysis: a case study in Northeast of Morocco. Appl. Water Sci. 10(7):1-10
Bizuwerk A, Taddese G, Getahun Y (2003) Application of Gis for Modeling Soil Loss Rate in Awash River Basin. International Livestock Research Institue (ILRI), Addis Ababa, Ethiopia, pp 1–11.
Bopche L, Rege PP, & Joshi RD (2022) Landslide susceptibility mapping: an integrated approach using knowledge-based numerical rating scheme, remote sensing, and multiple overlay analysis. J. Appl. Remote Sens. 16(1):014503
Brenning A (2008) Statistical geocomputing combining R and SAGA: the example of landslide susceptibility analysis with generalized additive models. SAGA — Seconds Out (=Hamburger Beitrȁge zur Physischen Geographie und Landschaftsȍokologie) 19:23–32
Bröcker J, Smith LA (2007) Increasing the reliability of reliability diagrams. Weather Forecast 22:651–661
Capitani M, Ribolini A, Bini M (2013) The slope aspect: a predisposing factor for landsliding? C. R. Geosci. 345:427–438 https://doi.org/10.1016/j.crte.2013.11.002
Catani F, Lagomarsino D, Segoni S, Tofani V (2013) Landslide susceptibility estimation by random forests technique: sensitivity and scaling issues. Nat. Hazard. Earth. Sys. 13 (11):2815–2831
Chang Z, Du Z, Zhang F, Huang F, Chen J, Li W, & Guo Z (2020) Landslide susceptibility prediction based on remote sensing images and gis: Comparisons of supervised and unsupervised machine learning models. Remote Sensing, 12(3):502
Charrire A (1990) Héritage hercynien et évolution géodynamique alpine d’une chaine intracontinentale : le Moyen Atlas au S de Fès (Maroc). Thèse Doct. Etat, Toulouse, p589.
Chen W, Chen X, Peng J, Panahi M, & Lee S (2021) Landslide susceptibility modeling based on ANFIS with teaching-learning-based optimization and Satin bowerbird optimizer. Geosciences Frontiers 12(1):93-107
Chen Y, Ming D, Ling X, Ly X, & Zhou C (2021) Landslide susceptibility mapping using feature fusion-based CPCNN-ML in Lantau Island, Hong Kong. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 14:3625-3639
Dai FC, Lee CF, Ngai YY (2002) Landslide risk assessment and management: an overview. Eng. Geol. 64:65-87
Ding S, Li H, Su C, Yu J, Jin F (2013) Evolutionary artificial neural networks: a review. Artif. Intell. Rev. 39:251–260
Dornik A, Drăguţ L, Oguchi T, et al. (2022) Influence of sampling design on landslide susceptibility modeling in lithologically heterogeneous areas. Sci Rep 12:2106 https://doi.org/10.1038/s41598-022-06257-w
El Kharim Y, Bounab A, Ilias O, Hilali F, & Ahniche M (2021) Landslides in the urban and suburban perimeter of Chefchaouen (Rif, Northern Morocco): inventory and case study. Nat. Hazards 107(1):355-373
Es-smairi A, El Moutchou B, & Touhami AEO (2021) Landslide susceptibility assessment using analytic hierarchy process and weight of evidence methods in parts of the Rif chain (northernmost Morocco). Arab. J. Geosci. 14(14):1-18
Evans IS (1979) An Integrated System of Terrain Analysis and Slope Mapping. Final Report on Grant DA-ERO-591-73-G0040. University of Durham, England.
Feizizadeh B, Blaschke T, Nazmfar H (2014) GIS-based ordered weighted averaging and Dempster–Shafer methods for landslide susceptibility mapping in the Urmia Lake Basin, Iran. Int. J. Digital Earth 7 (8):688–708
Felicísimo A, Cuartero A, Remondo J, Quiros E (2013) Mapping landslide susceptibility with logistic regression, multiple adaptive regression splines, classification and regression trees, and maximum entropy methods: a comparative study. Landslides 10:175–189
Fenjiro I, Zouagui A, & Manaouch M (2020) Assessment of Soil Erosion by RUSLE Model using Remote Sensing and GIS - A case study of Ziz Upper Basin Southeast Morocco. Forum Geografic, 19(2):131-142
Freund Y, Schapire RE (1997) A desicion-theoretic generalization of on-line learning and an application to boosting. J. Comput. Syst. Sci. 55:119–139
Ghorbanzadeh O, Blaschke T, Gholamnia K, Meena SR, Tiede D, Aryal J (2019) Evaluation of different machine learning methods and deep-learning convolutional neural networks for landslide detection. Rem. Sens. 11:196
Goetz JN, Brenning A, Petschko H, Leopold P (2015) Evaluating machine learning and statistical prediction techniques for landslide susceptibility modeling. Comput Geosci 81:1–11 https://doi.org/10.1016/j.cageo.2015.04.007
Haigh M, Rawat JS (2012) Landslide disasters: seeking causes – a case study from Uttarakhand, India. Management of Mountain Watersheds, pp. 218–253 https://doi.org/10.1007/978-94-007-2476-1-18
Henriques C, Zezere JL, Marques F (2015) The role of the lithological setting on the landslide pattern and distribution. Eng. Geol. 189:17–31
Hinaje S (1995) Apport de l’analyse de la tectonique cassante tardi et post-panafricane à la modélisation de la mise en place des miniralisations dans la boutoniére de Bou-Azzer (Anti-Atlas, Maroc). Thèse de 3ème cycle, Fac. Sci. Rabat, p238, + carte h.t au 1/2000
Hong H, Liu J, Bui DT, Pradhan B, Acharya TD, Pham BT... & Ahmad BB (2018) Landslide susceptibility mapping using J48 Decision Tree with AdaBoost, Bagging and Rotation Forest ensembles in the Guangchang area (China). Catena 163:399-413
Hussain MA, Chen Z, Kalsoom I, Asghar A, & Shoaib M (2022) Landslide Susceptibility Mapping Using Machine Learning Algorithm: A Case Study Along Karakoram Highway (KKH), Pakistan. J. Indian Soc. of Remot. 1-18
Imtiaz I, Umar M, Latif M, Ahmed R, & Azam M (2022) Landslide susceptibility mapping: improvements in variable weights estimation through machine learning algorithms—a case study of upper Indus River Basin, Pakistan. Environ. Earth Sci. 81(4):1-14
Kadavi PR, Lee CW, & Lee S (2019) Landslide-susceptibility mapping in Gangwon-do, South Korea, using logistic regression and decision tree models. Environ. Earth Sci. 78(4):1-17
Khosravi K, Pham BT, Chapi K, Shirzadi A, Shahabi H, Revhaug I, ... & Bui DT (2018) A comparative assessment of decision trees algorithms for flash flood susceptibility modeling at Haraz watershed, northern Iran. Sci. Total Environ. 627:744-755
Lee S, Hong S, Jung H (2017) A support vector machine for landslide susceptibility mapping in Gangwon Province, Korea. Sustainability 9 (1):48 https://doi.org/10.3390/su9010048.
Lee S, Pradhan B (2007) Landslide hazard mapping at Selangor, malaysia using frequency ratio and logistic regression models. Landslides 4:33-41
Lee S, Ryu JH, Won JS, Park HJ (2004) Determination and application of the weights for landslide susceptibility mapping using an artificial neural network. Eng. Geol. 71:289–302
Liakos KG, Busato P, Moshou D, Pearson S, Bochtis D (2018) Machine learning in agriculture: A review. Sensors 18:2674
Liu R, Yang X, Xu C, Wei L, & Zeng X (2022) Comparative Study of Convolutional Neural Network and Conventional Machine Learning Methods for Landslide Susceptibility Mapping. Remote Sensing 14(2):321
Liu Z, Gilbert G, Cepeda JM, Lysdahl AOK, Piciullo L, Hefre H, & Lacasse S (2021) Modelling of shallow landslides with machine learning algorithms. Geoscience Frontiers 12(1):385-393
Manaouch M, Mohamed S, & Imad F (2021) Coupling Fuzzy logic and Analytical Hierarchy Process (FAHP) with GIS for Landslide Susceptibility Mapping (LSM) in Ziz upper watershed, SE Morocco.
Manaouch M, Sadiki M, & Fenjiro I (2021) Integrating GIS-based FAHP and WaTEM/SEDEM for identifying potential RWH areas in semi-arid areas. Geocarto Int. 1-24.
Michard A (1976) Elément de géologie marocaine. Notes et Mém. Ser. Géol. Maroc, 252, 408
Mohamed M, Anis Z, Imed F (2020) Regional-scale modeling of water erosion and sediment yield in a semi-arid context: A case study of Ziz upper watershed in south-eastern Morocco. GEOIT4W-2020: Proceedings of the 4th Edition of International Conference on GeoIT and Water Resources 2020, Al-Hoceima, Morocco, March 11 -12, 2020. http://doi.org/10.1145/3399205.3399209
Mohamed M, Anis Z, Imad F (2020) Assessment of the Effects of Land Use/Land Cover Changes on Soil Loss and Sediment Yield Using WaTEM/SEDEM Model: Case Study of Ziz Upper Watershed in SE-Morocco. Curr. Appl. Sci. Technol. 21(2):337-350
Nasiri Aghdam I, Varzandeh MHM, Pradhan B (2016) Landslide susceptibility mapping using an ensemble statistical index (Wi) and adaptive neuro-fuzzy inference system (ANFIS) model at Alborz Mountains (Iran). Environ. Earth Sci. 75 (7):1–20
Navas A, Machin J, Gaspar L, Sadiki A, Kabiri L, Faleh A (2013) Les sols dans le pays du Ziz (Sud-est marocain) caractéristiques et aspects de dégradation, Ouvrage financé par le projet n° AP/037737/11, du Programme de Coordination interuniversitaire d’Investigations Scientifiques, du Secrétariat d’Etat de Coopération Internationale AECID.
Nouayti N, Khattach D, & Hilali M (2015) Evaluation de la qualité physico-chimique des eaux souterraines des nappes du Jurassique du haut bassin de Ziz (Haut Atlas central, Maroc)[Assessment of physico-chemical quality of groundwater of the Jurassic aquifers in high basin of Ziz (Central High Atlas, Morocco)]. Journal of Materials and Environmental Science 6(4):1068-1081
Panahi M, Rahmati O, Rezaie F, Lee S, Mohammadi F, & Conoscenti C (2022) Application of the group method of data handling (GMDH) approach for landslide susceptibility zonation using readily available spatial covariates. Catena 208:105779
Panchal S, & Shrivastava AK (2022) Landslide hazard assessment using analytic hierarchy process (AHP): A case study of National Highway 5 in India. Ain Shams Engineering Journal, 13(3), 101626.
Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, et al. (2011) Scikit-learn: Machine learning in Python. J. Mach. Learn. Res. 12:2825–2830
Pham BT, Vu VD, Romulus C, Phong TV, Ngo TQ, & Prakash I (2021) Landslide susceptibility mapping using state-of-the-art machine learning ensembles. Geocarto Int. 1-25
Pourghasemi HR, Rahmati O (2018) Prediction of the landslide susceptibility: which algorithm, which precision? Catena 162:177–192 https://doi.org/10.1016/j.catena.2017.11.022
Saha S, Sarkar R, Roy J, Saha TK, Bhardwaj D, & Acharya S (2022). Predicting the Landslide Susceptibility Using Ensembles of Bagging with RF and REPTree in Logchina, Bhutan. In Impact of Climate Change, Land Use and Land Cover, and Socio-economic Dynamics on Landslides (pp. 275-298). Springer, Singapore.
Sajadi P, Sang YF, Gholamnia M, Bonafoni S, & Mukherjee S (2022) Evaluation of the landslide susceptibility and its spatial difference in the whole Qinghai-Tibetan Plateau region by five learning algorithms. Geosci. Lett. 9(1):1-25
Samanta RK, Bhunia GS, Shit PK, & Pourghasemi HR (2018) Flood susceptibility mapping using geospatial frequency ratio technique: a case study of Subarnarekha River Basin, India. Model. Earth Syst. Environ. 4(1):395-408
Su Q, Tao W, Mei S, Zhang X, Guo J, & Yang Y (2021) Landslide Susceptibility Zoning Using C5. 0 Decision Tree, Random Forest, Support Vector Machine and Comparison of Their Performance in a Coal Mine Area. Front. Earth Sci. 1181
Tien Bui D, Pradhan B, Lofman O, Revhaug I (2012a) Landslide susceptibility assessment in Vietnam using support vector machines, decision tree, and Naive Bayes Models. Math. Probl Eng. 974638
Tsangaratos P, Ilia I (2016) Comparison of a logistic regression and Naïve Bayes classifier in landslide susceptibility assessments: the influence of model’s complexity and training dataset size. Catena 145:164–179
Wang Y, Song C, Lin Q, Li J (2016) Occurrence probability assessment of earthquake triggered landslides with Newmark displacement values and logistic regression: the Wenchuan earthquake, China. Geomorphology 258:108–119
Xiao T, Yin K, Yao T, Liu S (2019) Spatial prediction of landslide susceptibility using GIS-based statistical and machine learning models in Wanzhou County, Three Gorges Reservoir, China. Acta Geochim. 38:654–669 https://doi.org/10.1007/s11631-019- 00341-1
Xing Y, Yue J, Guo Z, Chen Y, Hu J, & Travé A (2021) Large-scale landslide susceptibility mapping using an integrated machine learning model: A case study in the Lvliang mountains of China. Front. Earth Sci. 622
Yilmaz I (2010) The effect of the sampling strategies on the landslide susceptibility mapping by conditional probability (CP) and artificial neural network (ANN). Environ. Earth Sci. 60:505–519
Youssef AM, & Pourghasemi HR (2021) Landslide susceptibility mapping using machine learning algorithms and comparison of their performance at Abha Basin, Asir Region, Saud Arabia. Geoscience Frontiers, 12(2):639-655
Zhang J, Wang Z, Hu J, Xiao S, & Shang W (2021) Bayesian machine learning-based method for prediction of slope failure time. Journal of Rock Mechanics and Geotechnical Engineering.
Zhang T, Quevedo RP, Wang H, Fu Q, Luo D, Wang T, ... & Renno CD (2022) Improved tree-based machine learning algorithms combining with bagging strategy for landslide susceptibility modeling. Arab. J. Geosci. 15(2):1-19
Zhang X, Song J, Peng J, & Wu J (2019) Landslides-oriented urban disaster resilience assessment—A case study in ShenZhen, China. Sci. Total Environ. 661:95-106
Zhou C, Yin K, Cao Y, Ahmed B, Li Y, Catani F, Pourghasemi HR (2018) Landslide susceptibility modeling applying machine learning methods: a case study from Longju in the Three Gorges Reservoir area, China. Comput. Geosci. 112:23–37 https://doi.org/10.1016/j.cageo.2017.11.019.
Zhou D, Simic-Milas A, Yu J, Zhu L, Chen B, & Muhetaer N (2020) Integrating RELAX with PS-InSAR Technique to Improve Identification of Persistent Scatterers for Land Subsidence Monitoring. Remote Sens. 12(17):2730
Zhou S, Zhou S, & Tan X (2020) Nationwide Susceptibility Mapping of Landslides in Kenya Using the Fuzzy Analytic Hierarchy Process Model. Land 9(12):535
Zhu L, Wang G, Huang F, Li Y, Chen W, & Hong H (2021) Landslide susceptibility prediction using sparse feature extraction and machine learning models based on GIS and remote sensing. IEEE Geosci. Remote S. 19:1-5

Download PDF

Version 1

posted

You are reading this latest preprint version

Landslide susceptibility mapping using machine learning techniques and comparison of their performance at Ziz upper watershed, Southeastern Morocco

Status:

Version 1

Abstract

Figures

1 Introduction

2 Study Area And Geological Setting

3 Methodology

3.1 Data sources

3.2 Landslide inventory map

3.3. Landslide triggering factors (LTFs)

3.3.1 Slope angle

3.3.2 Slope aspect

3.3.3 Slope length (LS)

3.3.4 Distance from roads

3.3.5 Distance from streams

3.3.6 Distance from faults

3.3.7 Lithology

3.3.8 LULC

3.3.9 Rainfall

3.3.10 Elevation

3.4 Normalization of LTFs

3.5 Modelling procedure using MLTs

3.5.1 AdaBoost

3.5.2 Artificial Neural Network (MLP)

3.5.3 Random Forest

3.5.4 Naïve Bayes (NB)

3.5.5 Instance-Based K (IBK)

3.5.6 Decision Tree (J48)

3.6 Evaluation of LTFs

3.7 Modelling prediction and performance

4 Results

4.1 LTFs contribution

4.2 Landslide Susceptibility Models

4.3 Accuracy assessment and comparison

5 Discussions

6 Conclusion

Declarations

References

Status:

Version 1