The Role of Initial Soil Conditions in Shallow Landslide Triggering: Insights from Physically Based Approaches

In the last years, the shallow landslide phenomenon has increasingly been investigated through physically based models, which try to extend over large-area simplified slope stability analyses using physical and mechanical parameters of the involved material. However, the parameterization of such models is usually challenging even at the slope scale, due to the numerous parameters involved in the failure mechanism. In particular, considering the scale of the phenomenon, the role of transient hydrology is essential. For this reason, in this work we present the outcome of different experimental tests conducted on a soil slope model with a sloping flume. The tested material was sampled on Monte Mario Hill (Rome, Central Italy), an area which has been frequently affected by rainfall-induced landslide events in the past. In this respect, we also performed a physically based numerical analysis at the field conditions, in order to evaluate the response of the terrain to a recent extreme rainfall event. The results of the flume tests show that, for the same material, two different triggering mechanisms (i.e., uprise of a temporary water table and advance of the wetting front) occur by varying the initial water content only. At the same time, the results of the numerical simulations indicate that clayey sand and lean clay are the soil types mostly influenced by the abovementioned rainfall event, since the initial moisture conditions enhance the formation of a wide wetting front within the soil profile.


Introduction
Many of the rainfall-induced landslides occurring all over the world are shallow-type; namely, the sliding surface is located at a depth from a few decimeters to some meters. They generally occur in response to prolonged intense rainfall events and involve either residual weathered soils or transported colluvial deposits. This type of landslides represents a widespread hazard that frequently results in considerable damage to infrastructure and human losses in many mountainous regions of the world, especially in areas characterized by the widespread presence of natural (e.g., [1][2][3]) and/or humanreworked soil cover (e.g., [4][5][6]). For this reason, in recent years great efforts are being made to improve the assessment of the temporal and spatial occurrence of rainfall-induced shallow landslides, especially through physically based models (e.g., [7][8][9][10]). However, the fundamental controls leading to slope failure driven by rainfall are still not well quantified [11], and thus the improvement of current models is still an important research topic [12]. In fact, despite the small size, it is not straightforward to define the complex interaction between hydrological and mechanical processes that develops before and during the triggering process [13]. In principle, the infiltrating water flow may cause both the development of a temporary perched water table, usually at the contact between the soil cover and the less permeable bedrock [14], and a decrease of the resisting effect (apparent cohesion) induced by increasing positive water pressure values in the unsaturated portion [15,16]. In this respect, it has been well recognized that matric suction can play a crucial role in the stability of unsaturated soil slopes [17]. Additionally, rainfall-induced slope failures heavily depend on the relationship between suction and water content which, in turn, are related with unsaturated conductivity functions as well as the rainfall intensity [18]. Considering the complexity of this research topic, a considerable amount of experiments has been conducted on understanding the behavior of waterinduced shallow landslides under controlled laboratory conditions using sloping flumes (e.g., [19][20][21][22][23]). Several authors [24][25][26] also reproduced via numerical modelling the performed experimental tests, assuming a triggering mechanism that commonly occurs when the wetting bands progress into the soil, resulting in loss of suction and in effective stress reduction [27,28]. Although these studies significantly contributed to better understanding the conditions leading to water-induced shallow landslides, there have been relatively few investigations for linking such experimental evidences with real widespread landslide events [29], evaluating in detail the triggering mechanisms [30]. In this sense, physically based models aimed at predicting the occurrence of shallow landslides over large areas (e.g., [31][32][33][34]) can be viewed as reliable tools, unless the uncertainties concerning the input parameter values be reduced as much as possible.
For this reason, in this work we analyze the triggering mechanisms of shallow landslides evaluating the effect of the initial soil conditions through laboratory flume experiments. Specifically, different tests have been performed on a soil sampled on Monte Mario Hill (Rome, Central Italy). This area has been affected by recurring rainfall-induced landslide events in the past, including the one that occurred between January 31 and February 2, 2014. In this sense, the outcome of the experiments has been also analyzed in relation to the results provided by HYDRUS-1D [35], a USDA (United States Department of Agriculture) Salinity Laboratory software package which can simulate the water flow into unsaturated porous media resulting from a rainfall event. Specifically, the insights resulting from the laboratory tests and the numerical simulations have been used to better understand the potential failure conditions for the shallow landslides occurred during the 2014 event.

General Features of the Study Area and the 2014 Event
The Monte Mario Hill, which is located in the northwestern sector of the city of Rome, on the right bank of Tiber River ( Figure 1(a)), represents the highest relief of the city (144 m a.s.l.). From a geological point of view (Figure 1(b)), it is composed at the bottom by a silty-clay succession of lower-upper Pliocene age (Monte Vaticano Formation (MVA)). This formation has a discordant contact with the upper Monte Mario Formation (MTM), a lower Pleistocene succession composed of silty sands [37]. At the base of this formation, it is also possible to identify a portion of reduced thickness (15 m) constituted by clayey silts (Farneto Member MTM 1 ). From a morphological point of view, Monte Mario is characterized by a relatively high slope gradient, especially along the eastern sector, which represents the result of both natural and anthropogenic factors, like fluvial erosion and manmade cuts. Such a slope gradient also enhances the triggering of rainfall-induced landslides, which generally involve a thin (0.5-2 m) layer of eluvial-colluvial superficial deposits overlaying the majority of the slopes. In this respect, between January 31 and February 2, 2014, 68 landslides have been recorded all around the city (Figure 2(a)), of which 12 occurred along slopes of Monte Mario only [38]. According to the data recorded by the Roma Monte Mario station, which is located within the study area, such landslides have been triggered by approximately 250 mm of rainfall cumulated in three days. However, in this time interval, two main subevents may be distinguished. The first (and most severe in terms of rainfall amounts) started in the early hours of January 31 st and ended after about 24 hours, with a total cumulated rainfall of about 190 mm. This subevent was characterized by extremely high rainfall intensity peaks, such as the one that occurred between 4:00 and 5:00 a.m. (46 mm).
Very high values continued to be recorded throughout the morning, with 3-hour and 6-hour rainfall equal to 87.6 and 140 mm, respectively. On the contrary, the second subevent, which took place in the afternoon of February 2 nd from about 1:00 to 6:00 p.m., was of shorter duration and characterized by less intensity peaks, resulting in a total average cumulated rainfall of approximately 40 mm. Most of the triggered landslides involved relatively shallow (less than 1 m) portions of 2 Geofluids weathered soil (Figure 2(b)) and can be classified as translational and rotational slides, earth flows, and debris flows [38]. Despite their limited thickness, such landslides caused the disruption of the road network, with consequent vehicle traffic bans and inconveniences for citizens as well as substantial damages to other infrastructures. ). The material is generally characterized by a medium plasticity and a high content of sand and clay, while the gravel and silt amount seldom exceeds 20-30% (Table 1). Specifically, the sampled material can be identified in four different USCS classes, which are consistent with eluvial and colluvial deposits resulting from 3 Geofluids In correspondence of sampling site S21, further material (approximately 500 kg) has been collected for performing the flume tests described in Section 3.2. In this sense, it is important to specify that the soil unit weight has been determined in this specific site by the sand cone method (Figure 3(d)), and the resulting value (14.9 kN/m 3 , which corresponds to a porosity of 48%, given a unit weight of soil solids of 26.4 kN/m 3 and a field water content of 11.5%) has been properly reproduced within the flume during the experimental tests.

Experimental
Set-Up. The experimental equipment (Figure 4(a)) comprises a rectangular sloping flume 100 cm long, 60 cm wide, and 20 cm high, whose sides were made in plexiglass in order to allow the visual observation of the wetting and triggering processes. To assure the same friction between the soil particles and the base of the flume as of that of particles inside the flume, a rough plastic panel was applied to the surface of the flume base. A stiff permeable barrier was fixed in front of the soil to contain it after the failure, whereas a video camera was used to monitor failure initiation time and location. Two properly placed spray nozzles above the flume guarantee an artificial rainfall having a raindrop size distribution and impact energy consistent with the experiment scale (Figure 4(b)). For assuring the correct functioning of the system, the supplied water pressure has been kept constant at 3.2 bar, resulting in a steady rainfall input of approximately 1/mm/min. During each test, the water content and pore water pressure within the soil were measured using, respectively, 4 EC-5 soil moisture sensors and 4 UMS-T5 mini-tensiometers ( Figure 4(c)). These sensors are connected to a data logger which acquires data each two seconds.
Three different flume inclinations have been used during the experiments (i.e., 27°,3 2°, and 35°) which represent, respectively, the minimum, average, and maximum slope values within the 2014 landslide source areas. The soil initial water content ranges between 11.9% and 19.2%, and it was obtained, before placing the soil into the flume, by wetting a specific quantity of oven-dried soil with the amount of water needed to reach the desired water content value. After the soil was set into the flume, the water content was checked by sampling the soil in different points. The initial porosity was attained overlaying four compacted soil layers parallel to the flume base ( Figure 5(a)). Given the fixed geometry and volume of a 4 cm thick layer, the soil weight required to fill that volume was calculated considering also the fixed initial water content. During the soil placement within the flume, the 8 sensors have been placed at different depths ( Figure 5(b)), with the aim of observing the infiltration process throughout the soil slope model. Finally, in order to achieve the least possible disturbance of slope conditions, a wedge-shaped slope was created in the termination of the material ( Figure 5(c)). The inclination of the wedge-shaped slope was set to 30°(lower than the friction angle) to ensure that failure occurred within the soil volume and not affecting this terminal slope only. This geometry was considered suitable for the experimental purposes, and then, it has been kept constant in each test.

Results
Ten flume tests have been performed by varying the initial water content (w) and the flume slope angle (α) ( Table 2).
According to the obtained results, different observations can be made: (1) as the initial water content or slope increases, the failure time decreases ( Figure 6); (2) by changing the initial water content, the failure time variation is higher in tests with a slope angle (α)of35°than in those with 32°( Figure 6); (3) in tests with α =27°, no failure occurred, but only a gradual erosion of the superficial layers, even with the wettest initial soil conditions (i.e., w =164%); (4) tests n.9 and n.10 have been performed for replicating tests n.6 and n.3, respectively. The aim was to evaluate the 6 Geofluids reproducibility of the assumed boundary conditions and, in turn, the reliability of the obtained results: in both tests, the failure time difference with the corresponding tests is lower than five minutes. As regards the failure mode, it was generally extremely rapid, with evidences of incipient instability only a few seconds before the failure (Figure 7). The detachment generally involved a soil thickness between 7 and 12 cm; thus, considering that the total soil thickness is 16 cm, the failure surface always developed within the soil profile and not at the contact between soil and the flume base. Furthermore, it is worth noting that, if the initial water content is relatively low, failure typically induces the mobilization of greater soil volumes, and vice versa.
Other interesting observations can be made by analyzing the records from the eight sensors placed within the soil during the tests. In this respect, if we compare data from test n.3 (α =32°, w =119%) and test n.5 (α =32°, w =192%), a different soil behavior in response to the rainfall input can be recognized. Specifically, in both tests  7 Geofluids starting from initial negative values of pore water pressure induced by partial soil saturation, soil undergoes progressive reduction of the matric suction, until the development of positive pore water pressures. However, in test n.3 such development occurs first for the deepest mini-tensiometers (MT1 and MT2) and then for the shallowest ones (MT3 and MT4) (Figure 8(a)). On the contrary, in test n.5, the sensors that first record a drastic reduction of matric suction are those located in the most surficial portion of soil (Figure 8(b)). Actually, in this specific test also MT2 recorded positive pore water pressures almost simultaneously to MT3; however, this point can be explained with a not sufficient soil compaction in the area above the sensor, which caused an excessive water percolation in that specific point.
As regards the water content, even if in test n.3 the first sensor that records an increase in water content is the shallowest one (SM4), at the end of the test the highest values are still recorded by the deepest sensors (SM1 and SM2) (Figure 8(c)). Conversely, in test n.5, SM3 and SM4 first recorded an increase in water content, reaching the highest values during the whole test (Figure 8(d)).
The same behavior, in terms of infiltration, can be observed also in tests with α =35° (Figure 9). In such a case, the greater slope angle induces most sudden variations of water content and pore water pressure, also enhancing an earlier response of sensors located closer to the soil wedge (i.e., MT1-MT3, and SM1-SM3). In this respect, it is worth noticing that in test n.6 (w =119%) failure occurs while the sensors farthest from the soil wedge (MT4 and MT2) still recorded negative pore water pressure values (Figure 9(a)).

Discussions
On the basis of the observations deriving from the outcomes of the performed tests, different insights can be inferred. Firstly, a strong sensitivity of the tested material to changing the flume slope angle can be noticed. In fact, if no failure was observed for α =27°regardless of the rainfall duration, for tests with α =32°, failure occurred in about one hour, and the time of failure strongly reduces, increasing by just three degrees the flume slope. Afterwards, a different response of the soil, in terms of infiltration, was observed by varying the initial water content, while keeping constant both slope and porosity, as well as the rainfall input. Specifically, for tests with lower initial water content, the sensors that first recorded variations induced by the infiltrating water flux are those located in the deepest part of the soil. On the contrary, for tests with higher initial water content, the first variations were detected by the shallowest sensors. This point suggests two different triggering mechanisms for the soil slope model. In the first case, it is possible to hypothesize the formation of a temporary water table at the base of the flume, which progressively rises until failure occurs, involving a relatively high amount of material (Figure 10(a)). In the second case, failure is likely induced by the advance of the wetting front, which mobilizes a lower soil thickness with respect to the preceding instance (Figure 10(b)). Therefore, the most interesting insight deriving from the flume tests is that the triggering mechanism may change only considering variation of the initial water content. In preceding works, several authors (e.g., [39][40][41]) highlight how antecedent soil moisture greatly affects the rate and depth of advance of the wetting front during intense rainfall, but not the type of Initial water content (%) Figure 6: Failure time vs. initial water content for different flume slopes. 8 Geofluids triggering mechanism that, instead, is generally associated with the rainfall intensity [28,42] and grain-size soil characteristics. In this sense, [43] assert that shallow landslides affecting fine-grained soils are induced by the reduction in matric suction near the ground surface due to rainfall infiltration, while a significant triggering factor for coarsegrained soils of high permeability like sands is a rise in the water table.
Therefore, the role of initial soil moisture in relation to the triggering mechanism should be examined more in detail for evaluating the potential shallow landslide triggering conditions, especially when physically based numerical models are employed. In this respect, it is considered that one of the main drawbacks of such models relies on the complexity in correctly evaluating the input parameters, especially over large areas. In this respect, different authors (e.g., [44][45][46]) 9 Geofluids have observed that soil cohesion, friction angle, and soil thickness represent the major sources of error for this type of analyses. With regard to the mechanical parameters, the problem is mainly related to the limited sampling [47,48] and the difficulty to properly evaluate the vegetation effect in terms of root cohesion [49,50]. As regards the soil thickness, several authors proposed linear correlations with elevation and slope gradient [51], semi-empirical geomorphologybased approaches [52], and multivariate statistical analyses of terrain parameters [53]. However, the results of the experimental tests demonstrate that a reliable analysis of the (c) (d) Figure 9: Pore water pressure (a and b) and soil moisture data (c and d) recorded during test n.6 (α =35°, w =119%) and test n.8 (α =35°, w = 16%), respectively.  Afterwards, the simulations have been extended to include the main rainfall event causing the triggering of the majority of landslides (specifically, between 1 a.m. of January 31 st and 1 a.m. of February 1 st ). The van Genuchten-Mualem model [54] was chosen as a hydraulic model to simulate water flow, whereas the hydrodynamic parameters θ s , θ r , and K s (Table 3) are predicted from soil grain size distribution using the ROSETTA Lite module [55]. Daily rainfall data (source: Centro Funzionale della Protezione Civile della Regione Lazio-Functional Civil Protection Centre of the Latium Region) have been used as input for the model, whereas evapotranspiration is accounted for by inserting the maximum and minimum temperature values (source: Servizio Idrografico e Mareografico Nazionale-National

Geofluids
Hydrographic and Marine Services) recorded during the investigated period into the Hargreaves equation [56]. With regard to the thickness of the model, we decided to use the average value of the three measurements deriving from the boreholes (1 m), while the soil column inclination has been set in order to obtain a superficial slope equal to 32°.
According to the simulations, it results that SC and CL are the soil types mostly influenced by the rainfall event (Figures 11(a) and 11(b)). In detail, both soils start from relatively high initial (W in ) moisture conditions (between 25% and 35% in the first 90 cm of soil). Afterwards, a clear wetting front forms, in particular after the 5 h rainfall and advances, causing the saturation of the first 20-30 cm. It is worth noting how such advance is greater in the case of SC as a consequence of the higher hydraulic conductivity. In the case of MH soil (Figure 11(c)), although W in is similar to that of the preceding soils, the rainfall effect is substantially irrelevant, at least in the short term, probably due to the lower hydraulic conductivity that enhances the run-off. The rainfall impact is small also for SM ( Figure 11(d)): in this specific case, W in is quite low and homogenous, which coupled with the high hydraulic conductivity of the material, inducing significant water content changes. However, these variations are not still sufficient to form either a wetting front (excluding the very first cm of soil) or a perched water table at the base of the column.

Conclusions
In this study, different flume tests have been performed to analyze the triggering process of rainfall-induced shallow landslides, with a specific focus on the role of the initial hydraulic conditions by changing the slope. In detail, it was observed that the increase in the initial water content anticipates the triggering time, particularly in the case of slope = 35°. This point, together with the lack of failure for tests with α =27°, also suggests the marked sensitivity of the tested material to even small slope variations. With regard to the failure mode, the data deriving from the soil moisture and pore water pressure sensors indicate two potential triggering mechanisms to variations of the initial water content, i.e., failure induced by uprise of a temporary perched water table and by the advance of the wetting front in the case of relatively low and relatively high initial soil moisture, respectively. At this point, for analyzing such process at the field conditions, we performed a numerical analysis with HYDRUS-1D, a physically based model which simulates the water flow into unsaturated porous media. On the basis of the data collected on site (i.e., Monte Mario Hill, Rome), we simulated numerically the rainfall event occurring between January 31 and February 1, 2014, which triggered 68 landslides along slopes of Monte Mario and the surrounding areas. According to the simulations, it results that SC and CL soil types are those mostly influenced by the 2014 rainfall event, since the higher initial moisture conditions enhance the formation of a wide wetting front within the soil profile.
In conclusion, the evidence that the triggering mechanism may change only due to variations of the initial water content should be considered also for future research activities regarding the physically based modelling of shallow landslides. Specifically, future improvements may concern a better evaluation of the role of other parameters (such as soil thickness and slope) in the triggering process. With respect to the case study described in this paper, starting from field evidences, further experimental tests should be performed, also considering the other soil types outcropping on Monte Mario. In this way, it will also be possible to verify if the current inferences may be extended to the entire area. In this fashion, physical laboratory modelling can also be viewed as a supporting tool for numerical models aimed at temporal and spatial prediction of shallow landslides occurrence over large areas. In fact, the rationale is to define the relation between rainfall, soil moisture, and triggering mechanisms for known laboratory boundary conditions and then to extend this relation at the investigated site through monitoring data collected in the field.

Data Availability
The data used to support the findings of this study are included within the article.

Conflicts of Interest
The authors declare that they have no conflicts of interest.