In the recent decades, the growth of population, man-made facilities, infrastructures, and lifelines at the expense of landslides-prone areas has been responsible for an exponential increase in human and economic losses in many parts of the world. In the Moulay Yacoub region, where marly hills dominate, the interaction of the semi-urban and rural socioeconomic development and landslides significantly increases, which urges identifying and prioritizing areas of risk in order to maximize harm reduction and to avoid the disastrous outcomes as is the case of Moulay Yacoub town. This paper aims to develop a landslide susceptibility map in a highly affected sector of the province, where no previous landslide data have been produced, and to find the most predisposing parameters. This goal is attained using two robust methods for landslide susceptibility mapping, the Frequency Ratio and the GIS Matrix Method. Before that, the correlation of 11 predisposing factors was tested. The results show that the anthropogenic factors, particularly the agricultural practices, were highly involved, and the field investigations proved that cereal farming slopes are the most affected. The success rate was about 0.75 (75%) for both models showing good quality results for the two susceptibility maps. Therefore, the two models could be efficiently used, and the new agricultural projects located in landslide-prone areas of the province must include such reliable methods of landslide hazard analysis to minimize the risk, which would put human lives, ecosystems, food production, and infrastructure in threat.

Au cours des dernières décennies, la croissance de la population, des installations artificielles, des infrastructures et des lignes de vie au détriment des zones sujettes aux glissements de terrain a été responsable d’une augmentation exponentielle des pertes humaines et économiques dans de nombreuses régions du monde. Dans la région de Moulay Yacoub, où dominent les collines marneuses, l’interaction du développement socio-économique semi-urbain et rural et des glissements de terrain augmente considérablement, ce qui oblige à identifier et à prioriser les zones à risque afin de maximiser la réduction des dommages et d’éviter les conséquences désastreuses, comme est le cas du centre urbain de Moulay Yacoub. Cet article vise à développer une carte de susceptibilité aux glissements de terrain dans un secteur fortement affecté de la province, où aucune donnée antérieure sur les glissements de terrain n’a été produite, et à trouver les paramètres les plus impliqués. Cet objectif sera atteint en utilisant deux méthodes robustes, le rapport de fréquence et la méthode de la matrice-SIG. Avant cela, la corrélation de 11 facteurs a été testée. Les résultats montrent que les facteurs anthropiques, particulièrement les pratiques agricoles, ont été fortement impliqués, et l’enquête de terrain a montré que les versants céréaliers étaient les plus touchés. Le taux de réussite était d’environ 0,75 (75%) pour les deux modèles montrant des résultats de bonne qualité pour les deux cartes de susceptibilité. Par conséquent, les deux modèles pourraient être utilisés efficacement, et les nouveaux projets agricoles situés dans les versants de la province sujettes aux glissements de terrain doivent inclure des méthodes d’analyse aussi fiables, afin de minimiser le risque, ce qui mettrait les vies humaines, les écosystèmes, la production alimentaire, et les infrastructures en péril.

Landslides represent the most destructive geological hazard in word wide (Bao et al., 2020; Petley et al., 2005; Schuster 1996; Schuster and Highland 2001), especially in the Mediterranean rim where the environments are known by an erratic distribution of intense rainfall events, and the neo-tectonic activity is coupled with a high-energy relief (Van Beek, 2002). Sometimes, the interaction between these geo-hazards and the local practice of land use causes serious damages to infrastructures and agricultural lands and threaten local socioeconomic activities (Lee and Choi 2004; Conforti et al., 2014; Del Soldato et al., 2017; Harmouzi et al., 2019; El Kharim et al., 2021; Obda et al., 2022). According to the International Federation of Red Cross and Red Crescent Society, in the last decades, more than 3.2 million people have been affected by landslides with more than 1.700 million dollars losses (Hölbling et al., 2015). The large damaging effects of these phenomena have incited adopting several prediction strategies to identify high-risk areas that are the most vulnerable to future landslides based on local physical parameter, such as the relevant geology, geometry and slope forming processes (Fell et al., 2008). With such information, people and decision makers can foster the prioritization of slopes requiring follow-up actions and take appropriate preventive measures to spare expensive socioeconomic losses and to plan safe future development activities (Pantelidis 2011). Therefore, in the last decades, landslide susceptibility mapping has become an essential subject in international geomorphology, engineering geology and regional/urban territory planning (Bălteanu et al., 2020; Monsieurs et al., 2019; Tien Bui et al., 2019), especially since the development of GIS-aided mapping procedures has enhanced the capability of preparation of landslide susceptibility maps, thus allowing the analysis at a broad scale (Rybar et al., 2002). However, the quality landslide susceptibility mapping (LSM) depends mainly on the completeness and precision of the landslide inventory (Pereira et al., 2012; Gaidzik and Ramírez-Herrera 2021; Bounab et al., 2022) and the quality of causal factors used for modelling (Pereira et al., 2012; Mind’je et al., 2020; Cao et al., 2021).

In fact, the inventory of past occurrences remains the most influential input factor on the calculation of the spatial probability of landslide occurrence (Pereira et al., 2012; Bounab et al., 2022), which makes the landslides mapping a decisive step. Landslide inventory maps are prepared for several objectives, such as defining the location and type of landslides in the study area (Antonini et al., 2002; Cardinali et al., 2001), showing the abundance of slope movements (Degraff 1985; Guzzetti et al., 2000), determining the frequency-area statistics of slope failures (Guzzetti et al., 2002; Hovius et al., 2000, 1997; Malamud et al., 2004), and providing relevant information to construct landslide susceptibility and hazard models (Soeters and Van Westen 1996; Chung and Fabbri 1999, 2003, 2005; Guzzetti et al., 2005, 2006).

Actually, recent studies proved that a proper computation of landslide frequency-area distributions (FAD) remains an efficient tool for testing the completeness of the landslide inventory (Van Den Eeckhaut et al., 2007), quantifying the risk and estimating erosion and sediment yields (Guzzetti et al., 2002; Fu et al., 2020) as well as comparing different geomorphological contexts in terms of landslide predisposition (Qiu et al., 2020; Sahrane et al., 2022). However, good probability size distribution models can only be obtained from complete geomorphological inventories (Galli et al., 2008).

As for the conditioning factors, which are considered as the second key parameter in landslide susceptibility modelling, recent studies usually assess about ten variables considered as predisposing factors of slope instability, e.g., Lee and Choi, 2004 (15 variables); Van Den Eeckhaut et al., 2010 (9 variables); Byou et al. (2020), (11 variables). However, most LSM studies, especially those published before the 2020s, tend to mainly focus on the effects of different statistical techniques used to calculate the landslide susceptibility model and to validate its goodness of fit without assessing the correlation rate between landslides and the variables chosen (e.g., Süzen and Doyuran 2004; Schicker and Moon 2012; Kavzoglu et al., 2015; Othman et al., 2015; Aditian et al., 2018). Moreover, the choice of these parameters (data of influencing factors) could depend on the spatial extent of the study area, where a local spatial scale requires more accurate input data (Zêzere et al., 2017) as compared to a regional scale, where the choice of input is more generalized (Thiery et al., 2007; Holec et al., 2013). However, the selection step of these influencing parameters requires a geomorphologic synthesis by experts of the geomorphological context of the study area to achieve data consistency and to avoid objective evaluation of landslides spatial susceptibility (Chung and Fabbri 2005; Bounab et al., 2022). This could provide physical explanations of the observed statistical correlations and contribute to advancing our understanding of the underlying physical processes controlling the evolution of hillslopes.

Two types of approaches are commonly used to perform the landslide susceptibility maps, qualitative and quantitative methods. Each of the latter has both advantages and drawbacks (Li et al., 2017). Qualitative approaches are based on the expert point of view (Akgun 2012; Bourenane et al., 2015; Chabok et al., 2019; Demir et al., 2013; Pourghasemi et al., 2012; Zhou et al., 2016), who should have enough knowledge about the local conditioning factors involved in landslides manifestations. Quantitative methods are based on the analysis of the functional relations between the factors of instability and the distribution of landslides and can be categorized into bivariate and multivariate methods. In the bivariate ones, each factor is combined with the landslide distribution map. As for the multivariate approaches, they are usually more adapted to irregular data distribution (Kavzoglu et al., 2014).

In the Rif Mountain belt, Northern Morocco, landslides are amongst the most frequent natural hazards and are considered as a major risk to the socio-economic development of the country after floods and droughts (WBG 2021). Besides, the high energy relief and the lithological factor, dominated by loose formations, combined with human practices made landslides very abundant (Mastere et al., 2013; Rouai and Jaaidi 2003; Tribak 1997), especially in the region of Moulay Yacoub-Fez-Taza, where the lithology, represented by weathered clays and marls, combined with steep slopes and water saturation are the main factors involved in slopes instability (Hassan et al., 2015). However, the study of the anthropogenic factors (especially the land use) and their impact on landslide occurrence in the Moroccan territory (especially in the Rif mountain belt) is uncommonly investigated while international studies focusing on this important topic have increased exponentially since the 1960s (Beguería 2006; García-Ruiz et al., 2010; Guns and Vanacker 2013; Notti et al., 2015), with most irrigation-induced landslides occurring in dry climates (Chevesich et al., 2021). Following this, recent studies have shown clear evidence of a strong and increasing control of the land use practices, especially the land cover change, agricultural activities and even urban extension, on landslide occurrence (Chevesich et al., 2021; Guns and Vanacker, 2013; Obda et al., 2022). Therefore, the integration of this parameter cannot be omitted in the landslide hazard analysis. Furthermore, landslide susceptibility analysis in the Pre-Rif region is still very poorly investigated. Recently, projects of construction of aptitude for urbanization maps launched by the ministry of the national territory planning aim to provide the whole of the Moroccan territory with reference documents making it possible to consider the dimension of the risk of natural disasters, including the landslide hazard, during the urban planning process and particularly, during the development phase of the various urban planning documents. The later cannot be established without basic expertise and analysis of landslide susceptibility in predisposed areas.

The aim of the present study is first, to create a landslide database and to identify the characteristics specific to the regional study context and which have a direct impact on the predisposition of slopes to landslides. The second goal is to develop a landslide susceptibility map for the first time in the MY region, whose slopes are widely affected. Thus, future prevention measures could be proactive and efficient. To achieve this, we have chosen two among the most robust and accurate methods used in landslide susceptibility studies. The bivariate statistical Frequency Ratio (FR) method is one of the effective and widely used processes in landslide susceptibility mapping. It provides better understanding of the influence of the factors selected (whether geomorphological, climatic or anthropogenic) on landslide occurrence (Razavizadeh et al., 2017; Vakhshoori et al., 2016). Moreover, the FR outperformed many other methods in landslide susceptibility mapping (Weights of evidence, Fuzzy gamma model, Support vector machine, Evidence belief function, etc.) in terms of accuracy (Huang et al., 2018; Vakhshoori et al., 2016) and the ease of data manipulation in a GIS platform (Zhang et al., 2020). On the other hand, the landslide susceptibility results pointed out the quality of the maps drawn by means of the GIS Matrix Method (GMM) in comparison with those made by other bivariate-statistical techniques (Perálvarez et al., 2009) and those made by the multiple-regression method (Irigaray et al., 2007). Its application showed satisfactory results and can be tailored for a broad scale where the same predisposing factors prevail (Irigaray et al., 2007; Boualla et al., 2019) which constitutes a similar context of our study area. In this study, we opted for these two models because in addition to the satisfactory results they provide according to the authors cited above, the FR method is counted among the most frequently used models, vice versa, the matrix method is still very little used in the modelling of landslide susceptibility, which encouraged us to test it and compare its results with the FR method in a specific and highly landslide-prone area as the MY province.

Study area

The study area is located in the west part of Moulay Yacoub province (MY) northern Morocco (Fig. 1), and includes 3 rural municipalities: Mikkes, Sebt Loudaya and Laajajra. The total area covered is 335 km2. MY province occupies the extreme south of the Pre-Rif unit (sub-unit of the Rif belt) and is limited to the south by the South-Rif overlapping front, which separates the Pre-Rif from the plain of Saïs (Fig. 1). The Pre-Rif consists of marly overthrust nappes that belong to Cretaceous and Tertiary ages (Sendide 2002; Winckel 2002; Lakhdar et al., 2006; Charroud et al., 2007). The marls domination affords to the region a landscape of hills with 5 to 20% average slopes. The highest hills rarely reach 900m altitude and are usually armed with sandstone or limestone ledge. The observed processes of water erosion (sheet, rill and gully erosion) reveal the relative soils impermeability, their friability and the strong surficial flow (Heusch 1970; Tribak et al., 2009).

The drainage pattern is non-perennial and quite dense. Moreover, the study area is known by subarid climatic conditions with an average of 314 mm of annual precipitations. Rainfall data of Fez and MY’s hydraulic stations show that most precipitations fall as stormy rain events, sometimes exceeding 50 mm, during very short periods (2 to 10 days), which expresses a severe seasonal contrast (DGM Moulay Yacoub; Obda et al., 2022). For the structural geodynamic of the region, the last studies confirmed recent tectonic activities at this Southern-Rif border (Agharroud et al., 2021; Chalouan et al., 2014; Poujol et al., 2017).

The main activity of the population in MY province, which does not exceed 200 000 residents (RGPH, 2014), is the agriculture which is primarily based on cereal farming and olive tree plantation. The housing regime is mostly rural except for the municipality of MY where the urbanization is closely linked to touristic and commercial activities provided by the existence of the thermal springs.

Landslide inventory

In the province of Moulay Yacoub, no slope movements data have been developed previously. The inventory map was established according to the classification of Hungr, 2014, who proposed a more detailed update to the Varnes classification. The landslide mapping was based on geomorphological field mapping, on visual interpretation of the aerial photos of the 1991 and 2005 missions at a scale of 40000 and 30000 respectively, and on the use of imageries like Google Earth and Bing maps, where most of landslides were mapped on the December 2018 image. Then, seven field investigation surveys were conducted in February 2020, July 2021, and October 2021 to check and validate concealed processes by land cover (especially brushwood and trees), and relatively eroded ones. Earthflows, slides, and rock falls were observed and inventoried in the study area (Fig. 2a). However, since rock falls are fairly localized, particularly at the foot of a few rocky escarpments, they were discarded from this study.

The term earthflows is used because their materiel is composed of marly soils (fine-grained material), has a flow-like surface morphology (Appendix 2), is essentially ductile and does not significantly lose strength during deformation (Hungr et al., 2014; Varnes et al., 1996). As a result, earthflows move slowly and intermittently without requiring steep slopes to be triggered (less than 12°) (Schuster et al., 1996). In fact, they have long periods of relative dormancy alternated with more rapid “surges”, allowing infrastructure and man-made projects to be built on them (Picarelli et al., 2005; Hungr et al., 2014).

Area-Frequency distribution analysis

Nowadays, modelling landslide size distribution is considered as essential in quantitative hazard and risk assessment because of the assessment dependency on the relationship between landslide size and frequency distribution (Corominas et al., 2008; Malamud et al., 2004). Moreover, several studies in the last decades have shown that landslide populations exhibit some power-law scaling across a portion of their size distribution (Guzzetti et al., 2002; Malamud et al., 2004; Brunetti et al., 2014; Li et al., 2016; Qiu et al., 2020; Sahrane et al., 2022). In fact, the tail of landslide size probability distribution for medium and large landslides is a power law where α controls the power law decays for medium and large landslide sizes, while β controls the maximum probability density of landslide distribution in square kilometres, and γ controls the rate of decay for small landslide areas in square kilometres. In addition to its ability to characterize the size distribution of the landslides affecting the slopes of the study area, the probability size distribution also helps to assess the completeness of the landslide inventory carried out (Eeckhaut et al., 2007; Guzzetti et al., 2002; Li et al., 2016). In the present work, an Area-Frequency distribution analysis will be carried out to check the completeness of the landslide inventory but also to investigate further how the land use practices affect the size of landslides in the study area. This will give us a better understanding of the role of vegetation types in landslide occurrence (Guns et al., 2014), which is essential to integrate the land use practices in the protection programs against the landslide risk in the region.

Landslide influencing parameters

Many factors have been used to establish landslide susceptibility maps (Catani et al., 2005; Wang et al., 2012). There are two types of involved factors in the occurrence of landslides, intrinsic and extrinsic factors. The first type is directly connected to slope instability, as the types of terrain, the presence of discontinuities, to the morphology of the slopes and to the hydrological conditions. The second type includes rather triggering factors such as intense rainfall, earthquakes and anthropic activities (railway or road constructions, mining, deforestation, drilling, different uses of lands, etc.) (Crozier 1984; Hansen 1984). For this study, the factors taken into consideration were those where data and maps were available. In the first step, slope, aspect, elevation, curvature, surface roughness, topographic wetness index (TWI), proximity to drainage network, proximity to roads, proximity to faults, lithology and land use were selected as the involved parameters of landslide influence in this study (Appendix 1). The first six factors have been calculated from the 12 meters DEM downloaded from ALOS-PALSAR platform, using GIS software. Lithology and faults were digitised from the Fes-Ouest geological map. As for the land use, roads, and the hydrographic network, they were extracted from the topographic maps and have been detailed on google earth imagery. The aspect has an impact on hydrological phenomenon, evapo-transpiration, it governs soil moisture, vegetation and grass, and root penetration. Therefore, slope aspect has an indirect link on landslide occurrence and its degree of susceptibility (Neuhäuser et al., 2012; Pourghasemi et al., 2018). Moreover, recent susceptibility mapping studies have shown that surface roughness proves to be an important parameter with a strong influence on the landslides occurrence (Komac 2006). As for the TWI, it is usually used to study the on-going hydrological processes of the sub soils. The topographical character of the terrain directly affects the soil moisture, which is considered as a conditioning parameter of shallow gravitational processes, and is recently used for their forecasting (Marino et al., 2020). To calculate the values of the index of the topographic moisture index we can use spatial data, namely the digital terrain model (Krivoguz and Bespalova 2017). This prompted us to evaluate the correlation of these parameter with landslide occurrence. Table 1 summarises the class range chosen for each parameter.

After selecting the influencing parameters in slope instability, a geostatistical connection between these parameters and each landslide type was estimated by analysing the association coefficients of contingency tables (Tab. 1). Then, a cross-tabulating of each factor grid layer and the landslide vector layer was analysed using SPSS statistical software to quantify the relative correlation. This step allows, on the one hand, to select from the factors previously considered as those which most contribute to each movement type giving rise to the landslides inventoried, and to square parameters that faintly influence the instability on the other. Good-man-Krustal’s Gamma (G-K) and Pearson’s R were the two correlation coefficients calculated (Davis 1973; Goodman and Kruskal 1979). The predictor variables are classified as “effective” (EFF) if the condition G-K index > 0.5 and R > 0.4 applies, and as “non effective” if it doesn’t (Fernández et al., 2003; Costanzo et al., 2012).

Despite that the cross-correlation were carried out for each landslide type, the G-K and R coefficients showed close values for earthflows and slides (Tab. 2). Therefore, the susceptibility maps were developed for the two landslide types combined.

GIS Matrix method (GMM)

The GMM is based on the construction of three matrixes in which the following data are recorded (Irigaray 1995; Chacón et al., 2006):

  • The first matrix TSM (Total Surface Matrix): was calculated by computing the selected factors as effective parameters, making all possible combinations among the selected factors, and then calculating the area occupied by each combination.

  • The second matrix LM (Landslide Matrix): was calculated by cross tabulating the reclassified landslide inventory binary map with the first resulting matrix TSM, this step allows us to calculate the area affected by landslides for each unique combination. (Irigaray et al., 2007),

  • The third Matrix SM (Susceptibility Matrix): is the result of dividing the LM by the TSM, the results can be calculated by a percentage that represent an assessment of relative susceptibility levels (0<LSM<100).

Then, the matrix of susceptibility (SM) is established from these data where the percentage of each of the combinations occupied by rupture zones for each type of movement is recorded (Tab. 3). The higher the percentage, the more susceptible is the corresponding combination of factors to landslide phenomena (Fernández et al., 2003). Finally, the values obtained are visualized showing 5 susceptibility levels (very low, low, moderate, high, very high). Moreover, it should be noted that there are no statistical rules for automatically categorizing continuous data. So, most researchers use their own expert opinion to develop class boundaries (Ayalew et al., 2004; Chaudhary et al., 2016). In our case, the landslide susceptibility levels of the GMM map were categorised using the Quantile method (groups that contain an equal number of values).

Frequency ratio method (FR)

In order to investigate landslide susceptibility at medium and small scales, statistical methods are the most frequently adopted approaches (Yalcin et al., 2011). Frequency ratio is one of the statistical methods frequently used (Lee et al., 2006; Pourghasemi et al., 2018; Yalcin et al., 2011). The FR method estimates the area correlation of landslide occurrence with each of the landslide-inducing parameters (Kannan et al., 2013). The reclassified factor maps were combined with the inventory map to calculate the frequency ratio of each factor class according to the equation below (Youssef et al., 2015). The higher it is the more it represents a positive correlation and high landslide susceptibility. In other words, if the FR is greater than 1, it means that a high association exists between slide locations and different classes, and a value lower than 1 means lower correlation.

FR=Np(LXi)i=1mNp(LXi)Np(LXj)j=1mNp(LXj)
graphic

with FR = frequency ratio of class i of factor j, Np (LXi) = number of pixels with landslides within class i of factor variable X, Np (Xj) = number of pixels within factor variable Xj, m = number of classes in the factor variable Xi, n = number of factors in the study area.

The Landslide Susceptibility Index (LSI) represents the relative susceptibility or hazard to landslide occurrence. To obtain the sliding susceptibility index, each factor’s frequency ratio values were summed to the training area as in the equation below (Yalcin et al., 2011; Hong et al., 2017; Sun et al., 2018). The higher LSI value indicates higher susceptibility to landslide. As or the landslide susceptibility levels of the FR map, they were categorised using Natural bricks method

LSIi=j=1mFRi(j)
graphic

Validation of susceptibility maps

The results quality assessment is an essential step in any mapping of susceptibility and/or hazard. In this research, the quality-control of the landslide susceptibility maps were assessed through the cross correlation between these maps and a previous inventory composed of 200 instability movements. In each case, two robust validation methods were used:

(i) In the first method, the degree of fit or concordance (DF) was determined (Goodchild 1986). This statistical method has been used and judged valid by several researchers studying instability movements (Chacón et al., 2006; Zieher et al., 2016; Boualla et al., 2019). The degree of fit was obtained according to the following equation:

DFi=Zi/SiZi/Si
graphic

where Zi is the area occupied by the landslide at each susceptibility level and Si is the area of the i class of susceptibility.

For DF results, the lower the values in the low and very low susceptibility classes and the higher the values in the high or very high susceptibility classes, the higher the quality of the susceptibility map (Fernández et al., 2003; Irigaray et al., 2007; Boualla et al., 2019).

(ii) The Area Under the ROC Curve (AUC) was the second validation approach. it is used to assess the performance of the statistical model, and larger AUC is considered better (Zhou et al., 2018). The AUC estimates the quality of the statistical model by describing its power to predict the occurrence or non-occurrence of landslide events (Feizizadeh and Blaschke 2013). Values close to 1 (100%) show excellent prediction, conversely if it is around 0.5 (random fit).

Landslide inventory

In the study area, 1113 earthflows and 843 slides were identified (Fig. 2a), which expresses a total sliding surface of 17,74 km2. Panels b and c of the same figure shows that for both landslide types, earthflows and slides, small size and therefore shallow movements (<10000m2) are much more common. This is undoubtedly controlled by the predisposing factors specific to the regional context of the study area.

The statistical analysis of the landslide inventory shows that the most affected slope gradient is the range of 10−15° with 1003 movement which represents 51% of the total occurrences (Fig. 3a). This suggests that, in the context of the study area, the moderate slopes of the MY’s hilly landscape are steep enough to be prone to slope movements, which, according to the landslide inventory, are shallow processes. Moreover, the relatively low occurrences in steeper slopes (>20°) can be explained by the fact that the latter promote more surficial water erosion (gullying) as strong inclinations may not allow enough infiltration to saturate the shallow marly layer. As for the material affected, about 1250 movements affects marly slopes and 484 were mapped in the calcareous marls, corresponding to respectively 63% and 25% of total occurrences (Fig. 3c). This means that the MY marls are the most prone to slope movements probably giving their low carbonate content compared to the calcareous marls and which seems to provide higher cohesion to the latter.

Furthermore, earthflows represent 57% of the total sliding area. Field investigation showed that most of them occur in cereal farming slopes (Fig. 4 and Appendix 2). The later alone represents 92% of the area affected by all the landslides, corresponding to 1637 movements (Tab. 1 and Fig. 3b), although the seasonal agricultural activity has erased morphological features of several landslides, especially the shallow ones. However, earthflows could be instantly reactivated after heavy rain events (Bardi et al., 2017). These results suggests that cereal-growing areas not only promote landslides, mostly shallow ones since they are more dominant according to the movement size analyses (Fig. 2b, c and d), but also represent a determining factor in the predisposition to shallow landslides.

As for the slides, they represent 43% of the movements mapped. Field visits show that many slides affect the road network which means that the latter is suspected of being a significant influencing parameter in slides triggering. Moreover, several cases of compound slides were inventoried in the field and by photo interpretation. The multiple scarps along the sliding area attest that the these movements occur along multiple rupture surfaces, which is commonly encountered in soft lithologies (Hungr et al., 2014; El Kharim et al., 2021; Obda et al., 2022). Furthermore, few large and eroded slides are observed in the region, the oldest ones are not included in the present study presuming that the parameters involved in their activity are not the same as today (Guns et al., 2013). The case study of slides investigated in MY town (east of the study area) develop with rotational components (circular rupture surface), showing prominent main scarp and back-tilted landslide head (Hungr et al., 2014). They are usually very to extremely slow (few mm per year) with short acceleration events (El Kharim et al., 2021; Obda et al., 2022).

Quality assessment of LIM

As for the assessment of the quality and the completeness of the landslide inventory, the probability size distribution of the landslides seems to be well fitted with the double pareto and the three-parameter inverse-gamma models (Fig. 2d). The rollover is well presented, and the power law does not exhibit a heavy tailed behaviour, which means that landslide inventory is quite complete for both ranges, small size/shallow (<104 m2) and large/deep landslides (>104 m2) (Qiu et al., 2020; Sahrane et al., 2022) . The quality of these results is quite comparable to the completeness assessing of landslide inventories in Eeckhaut et al. (2007); Fausto Guzzetti et al. (2002), and Tanyaş et al. (2019). Furthermore, the landslide size ranges are between 80 and 316,200 m2. As for the rollover effect (representing the most frequent landslide range), it is observed at 800 m2, which show that the small size landslides are much more frequent. This could be related to differing environmental conditions such as lithology, slope, aspect, and land use practices. This will be statistically evaluated in the section of factor’s correlation analysis.

Factors correlation analysis

After relating each factor class to slope movements distribution in the contingency tables (Tab. 1), the correlation assessment between the factors and the landslides has been performed. Table 2 shows the correlation coefficients (G-K and R) for each factor previously selected. The two factors that have been discarded are the aspect and the curvature, whose correlation coefficients are not high enough to say that these parameters play a determining role in the landslide hazard (G-K<0.5 and R<0.4). The hilly landscape of the region could be responsible for the low correlation value of the aspect parameter, as the low slope gradients mean that there is not a great difference in exposure to rain or sunlight, which generally affects the hydrological properties of the surface layer. The other influencing parameters are listed according to a decreasing order of the R of Pearson absolute value, which does not vary too much showing a moderate to high correlation for all the effective factors. Unlike that, the G-K values indicate a weak to medium correlation for some factors (land use, lithology, dist. from rivers, elevation) and a medium to strong correlation for others.

Slope angle is among the more effective instability factors for all the 3 landslide typologies, having a fairly high G-K and R values (G-K ≥ 0.6). In fact, slope gradient, faults-proximity, elevation, and distance from rivers are strongly connected parameters. In active mountain belts, such as the southern border of the Rif unit (as mentioned before), elevated hillslopes respond to an intense fluvial incision and slope adjustment which generates landslides to compensate for the oversteepening who increases the vertical component of gravity (Donati and Turrini, 2002 and Korup et al., 2010). This domino-effect results in the fact that these 4 parameters usually appear as controlling factors especially in the context of collision belt as the Rif. In the study area, the correlation values of the elevation and the distance to faults are not that high. However, the landslide affected area of the last factor shows that most landslides occur in the first 200 meters from the faults (Tab. 1) attesting that there is an influence even if it is not direct, it can be through a domino effect as mentioned above. As for the slope and the proximity to rivers, the correlation values clearly express their effective role in the landslide occurrence. However, the landslide affected area of the river proximity factor shows that most these gravitational processes occur uphill, at the furthest part from watercourses (>400m, Tab. 1). Such values mean that bank undercutting may not always be a causative factor for the landslides mapped although the strong correlation values. Otherwise, the other ranges of watercourse’s proximity would not have such weak values. On the other hand, the high corelation of the slope parameter confirms the results of the LIM analyses section where moderate slopes promote landslide occurrence while steep ones are prone to gully erosion.

Surface roughness and TWI are also among the main causative factors in our case showing medium G-K values (0.5<G-K<0.6) and slightly high R values (R >0.6). Surface roughness is considered as a significant conditioning factor in LSM (Abdulwahid and Pradhan 2017). In the specific case of landslides, the study of Eeckhaut et al. (2005) have shown that surface roughness can be used to identify landslides shape and estimate their relative age or create an assets landslide map. As for the TWI, it is used to quantify the effect of local topography on hydrological processes and modelling the spatial distribution of soil moisture and surface saturation (Qin et al., 2011). In several landslide susceptibility studies, the TWI appears among the main causative factors (Costanzo et al., 2012; Sun et al., 2021), which highlights the effect of soil moisture and surface saturation on landslide occurrence. This supports the correlation results, since the texture and the hydrological properties of the shallow layers are directly linked with the human practices in our case, and the cereal farming terrain shows the highest affected area (Tab. 1).

The road proximity and the land use are considered as anthropogenic controlling factors that promotes landslides occurrence in this study. The effects of roads and/or trails on slope stability can be summarised as (i) alteration of the natural hydrologic pathways, (ii) undercutting of unstable slopes, and (iii) overloading and oversteepening of the hillslopes (Sidle and Ochiai 2006; Guns and Vanacker 2013). Moreover, field investigations showed that landslides are often associated with bad road drainage and slope oversteepening, which explains the high R values for road proximity despite the limited road network in the MY province. For the land use, the G-K values are slightly weak, but the R ones are the highest. Land cover change induced by human activities significantly controls landslide patterns in the monotonous context of MY, as vegetation influences slope stability. In fact, vegetation modifies the hydrology of the subsurface (unsaturated zone) by controlling the impact of precipitation and evapotranspiration which affect the amount of water stored in the subsoil. Moreover, roots increase water infiltration and soil hydraulic conductivity (Beguería 2006). This explains why shallow movements are mostly found on cereal-growing slopes while wooded areas present relatively deeper processes (Appendix 2 and 3). Vegetation also has a mechanical effect on the subsoil as roots create an apparent cohesion via root fibre reinforcement that promotes slope stability (Roering et al., 2003; Schwarz et al., 2010). Consequently, the conversion from wooded areas to pastures and agriculture lands results in a rapid decline of the sub soil cohesion. According to Rogers and Selby (1980) study, forms of instability begin to manifest after 3 to 5 yr, which corresponds to the time it takes for roots to decay. This is well observed in the study area where field investigation and statistical analysis show that significantly fewer landslides are observed in forests, arboriculture zones and brushland compared with cereal farming areas (Fig. 4 and Tab. 1), knowing that the latter are much more dominant (83% of the total surface). In a recent study (Guns et al., 2013), the analyses of soil physical properties in natural forest soils showed that the saturated hydraulic conductivity is very high (estimated to be 500 mm/h) and the bulk density is very low (0.35 ± 0.13 g/cm3). Also, the observation of high porosity in the forest soils (87 ± 5%) indicates that water can easily infiltrate and percolate through the solum, which reduces the soil wetness and the landslide activity. Thus, in our case, the practice of cereal farming to the detriment of forests and brushland constitutes an effective role in the exposure of slopes to landslides by promoting water infiltration in the sub-surface, which after stormy rain events, triggers shallow movements.

As for the lithology, despite the low G-K value, the R ones are slightly high (Tab. 2). This parameter is considered among the main controlling factors in landslide susceptibility. Moreover, the low G-K values obtained for lithology do not reflect the real weight of this factor and may even lead to it being underestimated. In the present case, most of the slopes are covered by marls and calcareous marls (equivalent to more than 50% of the total area) which reveals the monotony of the lithological factor in the region. Moreover, 64% of the affected area is part of the marls class and 23% of the calcareous marls class. Which means that these two marly formations represent 87% of the landslide affected area. This leads us to believe that the lithological factor has a very effective role in landslide occurrence. Therefore, to estimate the real weight of this parameter, it is necessary to consider the old, inactive, and eroded landslides that affected the slopes, which are difficult to identify, especially when the weathering and the agricultural practices conceal the topographic features of these mass movements. Furthermore, most of these large and eroded landslides are very old and the factors (except the lithology) involved during their period of activity are not the same as today (especially the faults activity and its related parameters), for the reason that the landslide controlling factors are not fixed in time (Guns et al., 2013). However, if some factors are losing weight, the importance of other anthropogenic landslide controlling factors is increasing strongly and cannot be omitted in landslide hazard analysis. Which is the case of the current study where the installation of the road network and the land cover conversion from forests and brushwood to cereal crops are among the most influencing parameters in landslide occurrence.

FAD analysis of the land use parameter

The statistical analysis of the influence of the dominant vegetation on landslide occurrence using the probability size distribution have shown that:

  • 1. Small size/shallow landslides occurring in cereal farming slopes have an area of about 80 m2 while for the olive tree slopes, the smallest landslides have an area of about 250 m2 (Fig. 5)

  • 2. The rollover-effect, representing the area of the highest occurrence density, occur at 630 m2 in cereal farming slopes and at 1000 m2 at olive growing slopes

  • 3. The power law decay of landslide distribution decreases earlier for the cereal farming compared to olive growing, which decreases at a larger landslide size. Moreover, the first decreases more rapidly with an α (angle controlling the power law decay for medium and large landslides) equal to 1.2, while the second decreases with an α of about 1 allowing the power law decay toward larger landslides.

Therefore, based on FAD results, the cereal farming activity promotes more the small size landslides than the olive growing. On the other hand, landslide events occurring in slopes with olive trees, as the main vegetation type, are significantly larger. These findings support the discussions in the previous section. Thus, we can assert that, in the homogenous context of MY province, land use practices control the landslides magnitude and hence their typological difference as:

Susceptibility maps

The landslide susceptibility maps for both GMM and FR methods are presented in Figure 6. From a general comparison, the map resulting from the GMM method shows that very low susceptibility levels cover much more areas compared to the map of the FR method Table 4. On the other hand, areas of low and moderate susceptibility levels cover more areas in the FR map. As for the high and very high levels, they cover comparable surfaces.

A cross-comparison between the causative factor maps and those of the landslide susceptibility at a detailed scale have yielded some interesting conclusions about the direct impact of some factor classes on landslide susceptibility. A first observation shows that the valleys and the beds of the wadis have a very low susceptibility to landslides, especially in the GMM map. On the other hand, areas with high susceptibility are featured in the moderate and steep slopes which explains the strong correlation of the slope parameter. Regarding the lithology, the slopes dominated by marls and calcareous marls show high susceptibility levels while shales, saliferous clays and sandstone-marls outcrops seem to significantly increase the slope stability (Appendix 1 and Fig. 6). Moreover, for the land use, the orchards and the brushwood seem to directly reduce the susceptibility level of the areas they cover, especially in the GMM map (Fig. 7). Unlike that, most terrains of high and very high landslide susceptibility are dominated by cereal farming. This corroborates what was previously mentioned about the consequences of land cover conversion from natural forests, orchards and brushland to pastures and cereal farming on the exposure of slopes to landslides in the study area. Another interesting observation shows that badlands also reduce the landslides susceptibility level of the slopes (Fig. 7). Badlands are geomorphological processes that consist in the mobilization of the parent rock downslope by water erosion. They are characterized by steep slopes, minimal vegetation, and high drainage density. Moreover, they inhibit the agricultural use and are devoid of shrubs and grasses that prevent water erosion and infiltrated run-on water during rainfalls (Howard 2009), which promotes surface water erosion more than other instability phenomena such as landslides. The latter require water infiltration and moisture storage of the sub surface to be triggered (Chevesich et al., 2021).

LSM validation

The DF values for the very low and low susceptibility classes are 2% and 11% respectively for FR and GMM maps while the values for the high and very high susceptibility classes are 87,3% and 68% respectively (Fig. 8). In fact, the quality of these landslide susceptibility maps tested by the degree of fit method shows good values for each landslide susceptibility map, with a slight advantage for the FR method.

As for the AUC curves, they were plotted based on the number of correctly classified pixels (true-positive) and the number of the incorrectly identified pixels (false-positive). The AUC values for FR and GMM methods were estimated as 0,743 and 0,752 respectively (Fig. 8). These comparable results show effective landslide susceptibility modelling, knowing that AUC values between 0.7 and 0.9 indicate good discrimination ability (Swets 1988).

In the present work a detailed analysis of the LSM input database were conducted before computing the susceptibility models in order to investigate how the causative factors impact the landslide size-distribution and hence the hazard. In fact, the anthropogenic factors turn out to be among the most effective influencing parameters in MY region where the cereal farming areas dominating the MY region were highly affected and promoting small scale and shallow movements. The landslides mapped in olive trees slopes and brushland were larger and hence deeper processes. This highlights the role of human practices such as land cover conversion from forest, arboriculture areas and brushland to cereal farming crops in controlling the landslides magnitude and typology. Apart from that and based on the statistical analysis, TWI, slope roughness, lithology, slope gradient, elevation and faults proximity are all effective parameters promoting landslide occurrence in the study area.

Moreover, we tried to map the susceptibility to landslides in a geomorphologically homogeneous context, where influencing factors are monotonously repeated over almost the entire study area. In this case, the most efficient method is the one that is less influenced by this homogeneity and gives liberating results with a little detail, this is recently asserted in Sahrane et al. (2022a,b). Thus, according to the results of the susceptibility maps produced (Fig. 6 and Tab. 4), the FR model seems to amplify the level of susceptibility and is less influenced by the variation of factors especially the land use (Tab. 4 and Fig. 7). On the other hand, the cartographic comparison shows that the classes which, according to the statistical results of the landslide affected area (Tab. 1), are supposed to reduce the susceptibility to landslides, reduce it more in the GMM map than in the FR map (e.g., brushwood, olive trees and orchards, badlands) even in a detailed scale, despite the relative monotony of these classes at a broad scale. This confirms that the GMM model gives more liberative results.

This type of map constitutes an essential database to guide urban planning and food production plans in the region. However, the use of these data must seriously consider the influencing parameters identified in this study. Thus, in areas with high susceptibility levels, the risk mitigation could be well oriented and effective while saving time and costs. Furthermore, the models can trustfully be applied in other regions with similar geomorphological and geological contexts, especially in the Pre-Rif region.

Based on the results of these two accurate and practical methods, the future research could be directed towards improving the accuracy of these models by testing different combination of influencing parameters and comparing the results at a regional scale. This method will allow to find the most effective combination for accurate susceptibility results and then, be reliably applied in similar regional contexts.

     
  • MY

    Moulay Yacoub

  •  
  • LIM

    Landslide inventory map

  •  
  • GMM

    GIS Matrix method

  •  
  • FR

    Frequency Ratio

  •  
  • LS

    Landslide susceptibility

  •  
  • LSM

    Landslide susceptibility map

  •  
  • TWI

    Topographic wetness index

  •  
  • G-K

    Good-man-Krustal

  •  
  • EFF

    Effective

  •  
  • NEF

    Non-effective

  •  
  • LSI

    Landslide susceptibility index

  •  
  • FAD

    Frequency Area Distribution

  •  
  • DP

    Double Paretto

  •  
  • ING

    Three Parameter Inverse Gamma

  •  
  • DF

    Degree of fit

  •  
  • AUC

    Area under the curve

  •  
  • MPI

    Measures of the parameter importance

Cite this article as: I. Obda, O. Obda, R. Sahran, Y. El Kharim, A. Lahrach, M. Ahniche. 2024. Landslide susceptibility mapping using GIS Matrix Method and Frequency Ratio, application in the marly context of Moulay Yacoub Region, Morocco, BSGF - Earth Sciences Bulletin 195: 1.

This is an Open Access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.