Exceptional Cluster of Simultaneous Shallow Landslides in Rwanda: Context, Triggering Factors, and Potential Warnings

Abstract

Rwanda, in eastern tropical Africa, is a small, densely populated country where climatic disasters are often the cause of considerable damage and deaths. Landslides are among the most frequent hazards, linked to the country’s peculiar configuration including high relief with steep slopes, humid tropical climate with heavy rainfall, intense deforestation over the past 60 years, and extensive use of the soil for agriculture. The Karongi region, in the west-central part of the country, was affected by an exceptional cluster of more than 700 landslides during a single night (6–7 May 2018) over an area of 100 km². We analyse the causes of this spectacular event based on field geological and geomorphology investigation and CHIRPS and ERA5-Land climate data. We demonstrate that (1) the notably steep slopes favoured soil instability; (2) the layered soil and especially the gravelly, porous C horizon allowed water storage and served as a detachment level for the landslides; (3) relatively low intensity, almost continuous rainfall over the previous two months lead to soil water-logging; and (4) acoustic waves from thunder or mechanical shaking by strong wind destabilized the water-logged soil through thixotropy triggering the landslides. This analysis should serve as a guide for forecasting landslide-triggering conditions in Rwanda.

1. Introduction

The worldwide growth in population leading to the constant occupation of poorly accessible lands (deserts, mountains, etc.) results in increased exposure to natural hazards [1]. This is further accentuated by global warming. While giant fires, floods, and droughts are becoming more and more devastating, landslides remain among the main hazards for local populations, especially in mountain areas [2]. With a surface of about 26,000 km² and a population of more than 13 million people, increasing about 2.3% per year [3], Rwanda is the most densely populated country in continental Africa (571 people per km²). This small land-locked country of eastern-equatorial Africa, also known as the “Land of the Thousand Hills” for its hilly to mountainous landscape, is located along the geodynamic structure of the East African Rift System (Figure 1a). To the West, the Congo–Nile ridge (or Kivu Ridge) bordering Lake Kivu results from the extensional tectonic activity along the eastern edge of the Kivu graben that was initiated during the late Miocene [4,5]. The altitude of the Kivu Ridge ranges between 1500 m and 3000 m, and the slopes locally reach up to 60°. To the northwest, the Virunga volcanic belt notably includes Mount Kalisimbi (4507 m, the summit of Rwanda) and the active Nyiragongo volcano in the Democratic Republic of Congo.

Because of its elevated topography (Figure 1b,c) and its geographic situation close to the Equator, Rwanda is subjected to a temperate tropical climate [6,7,8]. The Kivu Ridge acts as a topographic barrier for the moisture brought from the Indian Ocean by the easterly winds, leading to between 1300 and 1600 mm of orographic rainfall in western Rwanda (Figure 1d). The climate is characterized by two rainfall seasons, autumn (September to December) and late spring (March to May) separated by a less rainy period during two months (January and February) and a dry season (June to August). Since the 1960s, the constant increase in population associated with the scarcity of agricultural land has led to intensive deforestation and over-exploitation of the soil, especially in the Kivu region [9,10]. The steep topography, high rainfall, near-complete deforestation, intense degradation of the silt–clayey soil, and a major increase in population are all favouring a high level of landslide hazards. The landslide-related death toll has thus increased dramatically from 38 in the period 2011–2013 to 174 in the period 2013–2016 and 551 between 2016 and 2021 [11,12,13,14].

It is generally considered that rainfall is a main factor inducing shallow and deep landslide formation [15,16,17,18]. Short-time (a few hours to a few days) heavy rainfalls [19,20,21] and long-lasting (1 to 5 months) exceptional events [22,23] both play a role. However, the effect of those two rainfall patterns vary depending on the local geomorphology and soil type. Indeed, fine-grained soil with low permeability will require higher antecedent rainfall and permeable soil before the pore water pressure reaches its maximum [24]. In addition, long-term rainfall modifies soil properties, especially by forcing the migration of fine-grained material or dissolving and recrystallizing elements, modifying the porosity and permeability [25]. In that respect, daily rainfall (or even short-term, 1 to a few days rainfall) is not always a reliable parameter to predict landslides [17] as long-term continuous rainfall, allowing infiltration in poorly permeable soils, may be more efficient in reaching the pore water pressure threshold [18].

Research works on landslides in Rwanda were initiated in the late 1989s with some first descriptions of soil creeping and the impact of runoff water on slope stability [26,27]. These first studies were followed by a gap of several years, and only in 2014, did research start again, mainly based on remote sensing analysis. Most efforts consisted of building landslide occurrence and susceptibility maps in the country [28,29,30,31,32]. Although of main interest, those maps show remarkable differences mainly linked to the type of satellite images used, the data processing method, the restricted area investigated by the authors, and the limits in landslide size considered pertinent [28,29,33]. The mechanisms triggering and controlling the genesis of landslides in Rwanda have been less explored. Most authors consider that given the tropical humid climate, rainfall is the main driver [28,33,34,35,36]. Furthermore, no obvious correlation seems to exist with earthquakes generated by the Kivu Rift activity [37]. Few studies include field-based observations such as the nature and structure of the soil and substratum, the vegetation cover, the occurrence of previous landslides, and the infrastructures (roads, villages, etc.). However, these factors play a major role in the development of landslides. Thus, this study aims to integrate all these parameters in a holistic analysis of individual landslide events in Rwanda, in order to provide guidelines for hazard assessment.

The Karongi area (Figure 1) was selected for its exceptional cluster of ca. 750 landslides (counting performed in this study based on satellite images) that occurred in a 100 km² area over a single night from the 6th to the 7th of May 2018 (Figure 2). This event occurred in a highly populated area and killed 18 persons [38]. This death toll makes the Karongi 2018 event one of the deadliest among the recent landslide episodes. We combined field data, high-resolution satellite images, and climate models to describe the geological, morphological, pedological, and climate context of the landslides in order to understand their triggering factors and dynamics. We demonstrate that the Karongi event was favoured by a 2-month near-continuous rainfall that water-logged the thin soil resting on steep slopes. We suggest that the collapse of the slopes was ultimately triggered by a thunderstorm (although not so exceptional) that brought rainfall and also wind- and thunder-sourced vibrations, probably leading to thixotropy of the soil, triggering numerous landslides.

Figure 1. (a) Topographic map of Eastern Africa (NASA GTOPO30 DEM) showing the two branches of the rift (the Western Rift and the Eastern Rift) fringing the Lake Victoria high plateau. The white rectangle corresponds to Figure 1b. (b) Rwanda elevation map based on the Copernicus 30DEM showing the high topography on both sides of Lake Kivu. The eastern rift shoulder (the Congo–Nile ridge) locally reaches 3000 m in altitude (Figure 1c) and is strongly affected by landslides. The Karongi area discussed in this work is indicated by the white square. The grey rectangle corresponds to the topographic data illustrated in Figure 1c. (c) Topographic swap (profiles are sampled every 20 m and separated by 1 km) across the Lake Kivu rift system showing the high altitudes of the Congo–Nile ridge in Rwanda and the progressive eastward flattening towards eastern lowlands of Rwanda and the Lake Victoria plateau. The black line represents the mean value and the grey shaded area indicate the dispersion of the values. (d) Mean annual rainfall pattern calculated using CHIRPS [39] annual data from 1981 to 2022. Note the strong contrast between the relatively low rainfall (<1000 mm/yr) in eastern Rwanda and the high rainfall (>2000 mm/yr) in the eastern Congo Basin to the west. Mean rainfall on the Congo–Nile ridge ranges from 1250 to 1500 mm/yr.

Figure 1. (a) Topographic map of Eastern Africa (NASA GTOPO30 DEM) showing the two branches of the rift (the Western Rift and the Eastern Rift) fringing the Lake Victoria high plateau. The white rectangle corresponds to Figure 1b. (b) Rwanda elevation map based on the Copernicus 30DEM showing the high topography on both sides of Lake Kivu. The eastern rift shoulder (the Congo–Nile ridge) locally reaches 3000 m in altitude (Figure 1c) and is strongly affected by landslides. The Karongi area discussed in this work is indicated by the white square. The grey rectangle corresponds to the topographic data illustrated in Figure 1c. (c) Topographic swap (profiles are sampled every 20 m and separated by 1 km) across the Lake Kivu rift system showing the high altitudes of the Congo–Nile ridge in Rwanda and the progressive eastward flattening towards eastern lowlands of Rwanda and the Lake Victoria plateau. The black line represents the mean value and the grey shaded area indicate the dispersion of the values. (d) Mean annual rainfall pattern calculated using CHIRPS [39] annual data from 1981 to 2022. Note the strong contrast between the relatively low rainfall (<1000 mm/yr) in eastern Rwanda and the high rainfall (>2000 mm/yr) in the eastern Congo Basin to the west. Mean rainfall on the Congo–Nile ridge ranges from 1250 to 1500 mm/yr.

2. The Study Area: Geologic and Geographic Contexts

The Karongi area is located to the east of Lake Kivu, near the city of Karongi (Kibuye) (Figure 1). The regional geology belongs to the Meso-Proterozoic, strongly deformed tectonic units of the Karagwe Ankolean Belt that characterizes the whole of Rwanda [40]. The lithology is mainly composed of alternating phyllites, quartz-phyllites, and a few quartzitic units (Figure 3). This metamorphic basement is intruded by large Late Proterozoic S-type granitic batholites [40,41,42]. Since the Miocene, the tectonic activity linked to the opening of the East African Rift has led to the formation of the Kivu graben and the correlated uplift of its shoulders [43,44]. Surprisingly, though the area is seismically active, landslide events in the North Tanganyika–Kivu rift are not correlated with earthquakes [34,37]. In addition, only one significant earthquake of magnitude Mw > 4 has been recorded around the Karongi area in the last 50 years (24 September 2023, Mw = 4.5, depth = 10 km) [45].

Because of its location on the Kivu ridge, the Karongi area is mountainous, ranging in altitude between 1600 m and 2700 m above sea level (asl). The landscape is that of a succession of flat-topped hills separated by several 100 m deep steep-sloped valleys (Figure 4). The western part of the study area is limited by the large NNW-SSE Rwankuba ridge, and most of the main valleys trend N-S. The area receives an average annual rainfall of 1300 mm.

The Karongi region is intensely cultivated, and the land cover is composed of a mix of small forest patches and cultivated fields, some of them arranged into terraces (Figure 5). Forests are composed of very spaced trees, essentially eucalyptus, with very limited undergrowth vegetation. Many of these woods are located on very steep slopes and in places where the humic soil is very thin [48]. The farmed areas include tea plantations, generally cultivated on selected fertile soil, in areas of limited slope. The rest of the farmland is occupied by food crops such as beans, potatoes, bananas, soya beans, and peas, which are interchanged at least twice a year. The Karongi area is densely populated; for example, the Rwankuba sector, where most landslides are located, has a population density of 548.9 pop./km² [3] (see location in Figure 5). Houses are made essentially of mud bricks and generally distributed along gravel roads running on the flat top of mountain ridges. However, a significant number of houses are also built on hill flanks close to cultivated fields, leading to a significant landslide hazard.

3. Materials and Methods

The major factors influencing the formation and dynamics of landslides in Rwanda and surrounding regions are steep topography, high rainfall, and the nature and structure of the soil and the consequences of a recent major increase in populations (deforestation, soil degradation, infrastructure building). After conducting an exhaustive inventory of all the landslides corresponding to the 2018 event, we performed a detailed analysis of the geology, geomorphology, soil characteristics, land use, and climate in the Karongi area.

3.1. Landslides Inventory

Landslides were identified from high-resolution optical satellite images (CNES/Airbus) dated from October 2018 and available on Google Earth Pro, which were recognizable from the distinctive shape and colour of the strip and deposition areas (Figure 6). As mentioned above, the Karongi area is covered with fields and patches of forest. Landslides are easily recognizable from their oblong shape and their brownish colour corresponding to soil stripped from vegetation (Figure 6). For the smallest landslides, some confusion with freshly overturned fields may occur, although these generally display a characteristic rectangular shape. Landslides were delineated as polygons to their full extent, including the slide scarp and the spreading zone with deposited material. Fieldwork was conducted in November 2022, April 2023, and November 2023 to describe in detail the landslide morphologies, dynamics of movement, material displaced, underlying bedrock, and soil thickness and structure.

3.2. Topographic Data

Rwanda has acquired a 10 m resolution Digital Elevation Model (DEM) using stereo-restitution based on 25 cm spatial resolution ortho-rectified aerial images through the cooperation of the Rwanda National Resources Authority (RNRA), the Regional Centre for Mapping of Resources for Development (RCMRD), and the World Bank [49]. This DEM was used in this study to derive slope maps as the slope angle is a primary controlling factor for the occurrence of landslides, especially shallow ones [50]. All other parameters being equal, the probability of landslide occurrence increases with the topographic slope angle. In Rwanda, it has been observed that slopes greater than 25°, especially on phyllitic rocks, are often affected by mass movements and especially by muddy flows [27]. Landslide initiation slopes were calculated from the DEM, using the Raster Zonal Statistics tool available in QGIS software. In every polygon corresponding to landslides, the maximum slope was extracted, and since landslides in Karongi generally have a small area, this value was retained as the initiation slope of the considered landslide. The soil layer affected by landslides is generally thin and parallel to the bedrock (see soil description below); therefore, the slope angles before and after a landslide are very similar and can be considered identical within the error margin of the calculated slope.

3.3. Soil and Substratum Characteristics

Aside from slope, soil characteristics are also important factors in shallow landslide occurrence [51,52]. While the structure and composition of the soil is a key factor in determining landslide susceptibility, soil thickness especially controls the volume and type of discharge; the thicker the soil, the higher the discharge [53]. Soil characteristics in the Karongi area were assessed in the field at different places where complete soil profiles were exposed (i.e., when both the top humic layer and the basement were observed) either on natural scarps (especially the edges of landslides) or along road and path trenches. Major soil horizons were described to estimate their role in the occurrence of landslides in this area and to constrain the landslide detachment level. The soil thickness data recorded in the field (using a tape measure on exposed complete soil profiles) were interpolated to derive a soil thickness map using Inverse Distance Weighting (IDW) interpolation in QGIS software as follows: georeferenced measured thickness points were weighted during interpolation such that the influence of one point relative to others decreased with the distance from the unknown thickness point [54]. Field measurements of soil thickness were complemented by mapping rock outcrops (corresponding to 0 m soil thickness) on high resolution satellite images (provided by the GoogleEarth platform) in order to increase the accuracy of the interpolated maps. The lithology data obtained from the geological map were complemented by field observations and bedding and schistosity measurements to provide a detailed map of the basement lithology and structure (Figure 3).

3.4. Rainfall and Wind Data

To assess the potential contribution of climatic parameters to landslide initiation, and especially the role of antecedent rainfall [22], we carried out a detailed analysis of the regional rainfall and wind patterns. The rainfall analysis was based on 42 years (1981–2022) of daily rainfall open-access data obtained from the Climate Hazards Group InfraRed Precipitation with Station data (CHIRPS [39]). In the absence of data on the occurrence of previous landslides in the region, calculating an effective antecedent rainfall period [15] was not possible, and we thus focused on determining the specificities of 2018 in terms of rainfall patterns. The CHIRPS data were used to calculate long-term monthly and annual precipitation trends over the whole available period. The Rainfall Anomaly Index (RAI) [55] obtained using the CHIRPS data was used to compute the inter-annual (considering only the March–April period) variability in rainfall within the Karongi area to depict a potential difference between 2018 (year of the Karongi event) and the other 41 years of record in this area. Hourly rainfall, wind speed, and wind direction data from the European Centre for Medium-Range Weather Forecasts ERA5-Land reanalysis model [56] were used to locate and characterize rainstorm events within the month preceding the landslide event on the 7th of May 2018. Thunderstorms were mapped using both the rainfall and wind data, i.e., a thunderstorm corresponds to a combination of localized heavy rain and locally strong, centripetal winds (note that ERA5-Land only provides the horizontal wind component, although the vertical component can be high in such convective cells).

4. Results

4.1. Geology and Soil Structure in the Karongi Area

The lithology and structure of the geological basement can play a major role in the development of landslides. Lithology can participate in controlling the type, thickness, and structure of the overlying soil ([57] and references therein). Structures like bedding planes, fractures, or schistosity can affect the mechanical behaviour, facilitating ruptures when interacting with topographic slopes [58] or allowing for larger rainwater infiltration rates.

The metamorphic basement of the Karongi area is composed of various units of different lithologies [42]. In general, bedding trends NNW-SSE with dip angles varying from 25 to 75° (Figure 3). Schistosity is largely near parallel to the bedding except in a few strongly folded areas. Several N-S trending faults are reported on the geological map [42,46], for which age is underdetermined. However, although strongly altered zones secant to the bedding and schistosity that could represent shear zones are observed in the series, surface weathering makes it impossible to identify fault kinematics. To the north and on a ca. 1 km wide N-S strip east of the main Rwankuba ridge, the lithology is dominated by meta-sandstones inter-layered with schists, some of them graphitic. The Rwankuba ridge itself is armoured by a series of metre-to-decametre-thick meta-sandstone layers dipping some 50 to 60° eastward and forming a 60 to 70° steep cliff to the west. To the SW, the basement is composed of gneisses and micaschists intruded by a small granitoid pluton. Finally, to the SE of the study area, the basement is formed by phyllites inter-layered with quartzites.

The humid equatorial climate of Rwanda makes for strong soil weathering. Following a long Mesozoic and Cenozoic history of regional-scale surface planation and alteration [59,60,61], some several-metres-thick alteritic profiles formed during the Plio-Quaternary before being partially or completely reworked because of the topographic changes induced by the tectonic activity in the East African Rift system [62,63]. Outside of the landslide deposition areas, the measured soil thickness varies between 0 m and 3 m, mainly depending on the local slope (Figure 7). The maximum thickness values are found in the river valleys where fluvial and diffusion accumulation occur, while in most of the area, the soil thickness varies between 0.5 and 2 m. The soil structure seems to be the most complete on the flat-topped ridges; Figure 8 shows the following:

• The O horizon (litter layer) is very limited (0 to a few centimetres in thickness), except in wooded areas where it is formed by leaf accumulation.
• The A horizon (humic topsoil) is generally restricted to a few centimetres to decimetres, except in some very specific areas on top of the ridges where it can reach up to 1 m.
• The E and B horizons (leaching layer and subsoil) are difficult to distinguish and generally form a near-homogeneous layer. When developed on meta-sandstones and schists, it is composed of a reddish to light brown mixture of sand and clay, 0.5 to 2 m thick. A few dispersed pebbles and blocks of basement rocks are frequently observed within these horizons. In places where the substratum is formed by graphitic schists, this E + B horizon is mainly composed of brown clay.
• The C horizon (weathered parent material) is represented by a centimetre- to decimetre-thick layer of clasts, ranging in size from gravels to blocks, moderately rounded to angular. Those clasts are mainly quartzite and locally derived meta-sandstones. Some of them show iron coating, suggesting the existence of a strong lateritic phase, although no evidence of iron crust has been observed.
• Depending on its dominant lithology, the R horizon (parent material) can be strongly weathered, especially when composed of graphitic schists.

When present, the soil resting on steep slopes is very thin (generally less than 1 m), although still structured the same way as on the top of the hills (Figure 9). Finally, the soil occupying the bottom of the flattest valleys is more heterogeneous, formed by colluvium and/or, in some cases, fine-grained deposits marking ephemeral lakes or swamps created by damming from landslides.

4.2. Landslide Morphologies

Landslides in the studied area are generally of small area, ranging from 50 m² to 68,000 m², with a median surface of 1850 m² (including striped areas and downslope deposits). According to the updated Varnes classification [64,65], the observed May 2018 Karongi landslides were shallow mud–debris slides (avalanches and flows). As shown in Figure 2, the material is very liquid, indicating a high water content, and is deposited as a thin sheet of mud preserving the topography of the cultivated terraces. The median angle of the topographic slope for landslide initiation in May 2018 was 40°, with maximum slopes reaching 60° (Figure 10). Few landslides (about 3%) occurred on shallow slopes between 20 and 30°, probably favoured by the low viscosity of the material. Where observed, the detachment zone is always situated between the pebbly layer forming the C horizon and the underlying bedrock (Figure 8 and Figure 9).

The largest landslide, which occurred on the northern flank of the Nyamiringa River valley (see location in Figure 4 and refer to Figure 6 and Figure 9 for morphological characteristics), provides a good example of the morphology and kinematics of the Karongi landslides. Based on the images and the description made by local people, the landslide occurred as a low-viscosity mudflow that covered the landscape with a thin (< 1 m thick) sheet of mud and few larger elements of pebble to small blocks size made of locally derived basement lithologies (Figure 2). The landslide was triggered on a steep slope, locally > 50° developed on graphitic schists with schistosity dipping about 70°, near parallel to the slope (Figure 9). The top of the hill is capped by the 1 to 2 m thick characteristics of the soil described above. The surface of the striped area is covered with angular pebbles and small blocks of the underlying graphitic schists (corresponding to the C horizon described above), indicating that the material was mobilized above that layer (Figure 9c,d,e). Downslope, at the foot of the steepest slope, reddish, poorly sorted, sandy colluvium including floating blocks of local basement reach up to 3 m in thickness. These colluviums are affected by internal shear zones limiting rafts of material sliding down the slope and creating open cracks perpendicular to the slope (Figure 9f,g). Lower down, the deposits are progressively thinner, reaching about 1 m in the lowermost deposition area. The material is still poorly sorted but the matrix is much finer-grained (mostly fine sand, silt, and clay) and the proportion of blocks tends to increase. Those deposits are incised by narrow, regressive gullies (several metres deep in the upper part), some of them reaching the basement rock on the steepest slopes. The Nyamiringa River flows at the foot of the landslide, incising the basement rock and creating about a 5 m high knickpoint immediately downstream of the landslide. To the south of the river, ca. 10 m high terrace is composed of poorly sorted blocky material that we attribute to an old, large landslide that originated from the southern slopes of the valley, which probably blocked the Nyamiringa River creating a temporary dam (Figure 9h). Indeed, upstream of that gravity deposit, the valley floor is flat, filled with fine-grained material most probably deposited in an ephemeral dam–lake environment. Finally, the north-facing slope of the hill supporting the Burega school (Figure 6 and Figure 9) is again covered by a 1 to 2 m thick reddish soil characteristics of the area, deposited on phyllites with alternating quartzites. This slope was also affected by small landslides in 2018.

4.3. Long-Term Climate and Extreme Weather Conditions in Spring 2018

Based on CHIRPS data, the total yearly precipitation oscillated between 950 mm and 1600 mm, with an average of 1300 mm during the 42 years of the record (1981–2022) (Figure 11a). The rainiest year was 1987 (1525 mm), while the driest was 2008 (963 mm). The year of interest in this study, i.e., 2018 (1428 mm), was the fifth rainiest year (Figure 11a). The so-called ‘’long rainy season’’ (a period covering March to May) received the most precipitation (Figure 11b), but, practically, it rained abundantly during nine consecutive months from September to May (rainfall > 100 mm/month). The dry season extended from June to August (Figure 11b). March 2018 (229 mm) and April 2018 (233 mm) received a considerable excess of precipitation compared with the long-term average monthly rainfall of 156 mm (March) and 161 mm (April), summing to a total precipitation excess of 143 mm in the two months that preceded the landslide event. This is also shown by an outstanding RAI value (>5) (Figure 11c) for March–April 2018 compared with those of the other 41 years: the period March–April 2018 was the most humid in 42 years in the Karongi area. Previously, only 2013 (256 mm) and 1999 (240 mm) had a rainier March than 2018, and 1985 (275 mm) and 1986 (246 mm) had a rainier April. Only 1985 reached a combined March–April rainfall level (423 mm) close to that of 2018 (462 mm), although it was still lower.

Rainfall in the Karongi area was near-continuous in the period preceding the 7th of May 2018, with a total of 42 rainy days from the 1st of March to the 6th of May (a day is considered rainy when CHIRPS precipitation values exceed 1 mm), i.e., 3 days over 5 days in average. This was the highest number of rainy days during that period over the 42-year record (Figure 11d). The two days (the 5th and 6th of May) preceding the landslide event were the rainiest consecutive days in that period, with a total rainfall amount of 57.6 mm (Figure 11e).

Heavy rain events such as thunderstorms are prone to trigger landslides; indeed, local inhabitants indicated that the Karongi event occurred simultaneously with an impressive thunderstorm. However, as usual with oral reports, the appreciation of the strength of that storm, especially in terms of rainfall intensity, varied widely from nearly no to exceptionally heavy rain. These storms are poorly resolved in the daily CHIRPS data. Although the ERA5 and ERA5-Land models are known to misestimate the absolute rainfall amount associated with storms in tropical regions [66,67,68,69], their hourly resolution still allows for detecting and localizing them. The ERA5-Land data were used to determine the number, position, and relative strength of rainstorms in the Karongi region during the period from the 1st of March to the 6th of May 2018. (Figure 12). Five events combining high rainfall and strong wind were detected over the considered period (Figure 12c). Other days such as the 26th of April recorded high rainfall but no wind or inversely, such as the 7th of April, and are not considered thunderstorms. It is to be noted that the last event on the 6th of May displayed a low rainfall amount and a medium windspeed value relative to the others. The strongest event was recorded on the 11th of April, nearly one month before the Karongi landslide event.

5. Discussion

Individual landslides occur through a conjunction of slope instability and an external triggering event such as heavy rainfall, soil defrosting, or earthquakes. However, except when related to major rainfall events [70] or a strong earthquake [71,72], the formation of massive landslide clusters over relatively large areas such as in Karongi remains poorly documented and understood [25]. Based on the data presented above, we discuss the parameters that led to the simultaneous triggering of about 740 landslides of all sizes reported in our study area.

Heavy rainfall during the rainy seasons, especially in April and May, is the most commonly accepted landslide-triggering mechanism in Rwanda. However, our results suggest a more complex mechanism at the origin of the Karongi cluster. The analysis of the CHIRPS data indicates that March and April 2018 were not only exceptional in terms of the total amount of precipitation but also in terms of the number of rainy days (Figure 11). The soil structure in the Karongi area is mainly characterized by a low-permeability, clay-rich layer including the E and B horizons, resting on a C horizon composed essentially of a layer of gravel to blocks (Figure 8). This coarse horizon represents both a layer of low shear stress and a layer of water storage. Associated with the steep slopes of the Karongi region, this makes for very unstable soil cover that is prone to developing shallow landslides [73,74]. The high population density leads to intensive exploitation of land using non-mechanized processes. Except for the growing number of tea plots, crops on individual plots are rotated at least three times a year, even on the steep slopes. Soil overturning down to 50 cm or so increases the soil permeability by creating a fragmented surface (from overturning) and reducing the vegetation cover. The increased rainfall infiltration and the absence of a permanent root system (or its very limited efficiency in the areas covered with eucalyptus) reduces the soil stability [75,76,77].

Coulomb-type materials are a biphasic mixture of grains and fluid in which, following the Mohr–Coulomb theory, shear stress depends on the normal constrain τ_N. This constrain, which affects both the interstitial fluid and the contacts between the grains, can be formalized as follows:

τ_N = τ’_N + p

(1)

where p is the interstitial pressure and τ’_N is the normal effective constrain [78]. The onset of sliding occurs when the flowing-threshold constrain τ_C is reached:

τ_C = C + τ’_N tan (φ)

(2)

where φ is the internal frictional angle and C is the cohesion of the material.

Mudflows such as those characterizing the landslides in the Karongi district result from rapid liquefaction of clay-rich material following thixotropic behaviour [79,80,81,82]. When the thixotropic material initiates movement under a stress higher than the maximal equilibrium shear stress (i.e., τ_N > τ_C), its viscosity decreases sharply, leading to liquefaction. As shown in Figure 2, the mudflows in Karongi were extremely fluid and largely saturated with water. We suggest that the near-permanent, exceptionally high precipitation of March–April 2018, associated to the agriculture-related increase in infiltration, saturated the soil (Figure 13a). This increased the interstitial fluid pressure (p), thus reducing the normal effective constrain (τ’_N) following Equation (1), and reduced the cohesion [83], which in turn lowered the flowing-threshold constrain τ_C (Equation (2)). Because the porosity of the gravelly C horizon is higher, the fluid pressure in that layer (p₂) especially increased compared with that in the less porous clay-rich soil (p₁).

However, an external mechanism is needed in order for several hundreds of landslides to be triggered simultaneously on slopes with locally varying soil thicknesses, internal structures, and compositions. Aside from heavy rain, ground shaking from earthquakes is another major mechanism that triggers landslides [71,72]. However, no earthquake was recorded in the Karongi region or even in western Rwanda in late April or early May 2018. The analysis of the climatic conditions from ERA5-Land data highlighted a series of five thunderstorms between early March and the 6th of May, the latest being simultaneous with the landslide event, although not the largest (Figure 12). Soil vibration from the strong winds and the sound waves linked to thunder may be at the origin of the thixotropic liquefaction of the water–saturated soil. Small vibrations such as sound waves applied to a granular material induce a re-organization of the grains that leads to compaction [82,84] and, in a non-confined material, also allows for the activation of flow [85]. The characteristics of thunder-generated acoustic waves in the seismic record have long been recognized [86,87,88], and some studies suggested they could be used to describe the near-surface (down to several metres) ground velocity structure [89]. Measured vertical ground motion linked to thunderstorms is of the order of 0.1 to 2 microns/s and amplifies down from the surface while the radial peak-to-peak amplitudes decrease rapidly [88]. Similarly, wind energy is the source for a number of seismic signals including both horizontal and vertical motions [90,91]. Such signals are recordable down to ground depths of several 10 to 100 m [92]. In particular, seismic waves generated by the interaction of surface obstacles like trees are known to generate ground motions similar to or higher than small M 1.0–1.5 earthquakes [92,93,94]. Depending on the wind speed and the nature of the obstacles (including the nature of the connexion with the ground), recorded vertical movement can reach 10 microns/s [92].

The detailed parameters involved in the physical mechanisms leading to this exceptional landslide cluster remain to be determined. However, based on the data presented above, we propose that while the thunderstorms that preceded that of the 6th of May appeared generally stronger in terms of rainfall and wind speed, they occurred prior to the complete water saturation of the soil, and especially of the C horizon, where the fluid pressure remained below the equilibrium pressure (p_e2). Between the 25th of April and the 6th of May, the saturation reached a threshold that, under the mechanical solicitation of the sound waves and/or wind-induced vibrations, increased fluid pressure in both the C horizon and the water-logged, clay-rich soil above (Figure 13b), decreasing the cohesion (C) of the material and triggering the initial movement. Since clay-rich soil has a thixotropic behaviour, this movement initiated liquefaction and the catastrophic landslide (Figure 13c).

Individual mudflows of the same type as the Karongi landslides are frequently observed in western Rwanda, and their detailed analysis, similar to the one reported here, should help validate the proposed model, especially the mechanism leading to the initial thixotropic behaviour. Indeed, aside from natural acoustic or mechanical waves (wind, thunder, earthquakes), man-made tremors such as mine or road construction blasting may also be an initiation factor. In terms of hazard warning, an effort should also be made to recognize (and forecasting) the periods of peculiar rainfall conditions leading to progressive soil water-logging, especially in regions of thin, layered, clay-rich soils.

6. Conclusions

The cluster of ca. 750 simultaneous landslides that occurred in the Karongi district in May 2018 resulted from a combination of an exceptionally rainy two-month period before the event, a peculiar soil structure characterized by a coarse-grained C horizon allowing water storage, steep slopes, and a thunderstorm as the ultimate triggering mechanism. The proposed model remains to be further tested. Indeed, the detailed physical mechanisms underlying the proposed model should be further investigated, especially by determining the exact soil properties (cohesion, porosity, etc.) required to perform numerical stability tests. Understanding the parameters leading to the Karongi event is the first step toward prevention. Indeed, while heavy rain is usually accepted as the forcing parameter for landslides, the Karongi landslides are not linked to exceptionally heavy rainfall but to the progressive water saturation of the soil and the sudden triggering of thixotropy by acoustic waves during a thunderstorm. In this case, a hazard warning should have derived from the recognition of the medium-term (two-month) weather pattern rather than from the forecasting of a particularly strong storm. Nonetheless, this study highlights the necessity of a better assessment of the soil properties in the hilly region of Western Rwanda and comprehensive monitoring of the local weather conditions.