Between 2005 and 2014, non-event related landslides were responsible for 4506 deaths in the 40 most mountainous countries of the world, excluding tens of thousands of deaths linked to landslides triggered by earthquakes [1
]. Such natural disasters can often lead to catastrophic consequences, especially in mountainous areas where intense rainfall, high regional seismicity, weak geological materials, and steep slopes act as major triggers or predisposition factors. Therefore, it is essential to monitor landslides and to analyze related mechanisms in order to reduce and prevent the negative socioeconomic impacts linked to these natural hazards. The improvement of spatial and temporal resolution of satellite imagery, along with the development of novel techniques, makes it currently possible to cover large geographic areas and collect huge amounts of data periodically [2
]. These geodatabases can be used before the disaster to create susceptibility maps and monitor high risk zones, but also to identify low-risk rehabilitation zones post-landslide. In this context, remote sensing is a powerful tool for landslide disaster management, including risk evaluation and hazard assessment.
In Central Asia, landslide risks are enhanced by the intense tectonic activity, resulting from the ongoing collision between the Indian and the Eurasian plates [3
] (Figure 1
b). Kyrgyzstan, with 82% of its territory representing mountainous land area [1
], is highly prone to such natural hazards. Intense rainfall and rapid snowmelt during spring seasons, as well as the persistent seismic activity, further accentuates landslide risks in this country, particularly along the northern and eastern rim of the Fergana Basin (Figure 1
). In the Kyrgyz part of the Fergana Basin, near the town of Maily-Say, the first recorded massive activation of landslides occurred in 1954 and was followed by other events in 1958, 1969, 1979, 1988, 1993–1994, 1998, and 2003–2005 [5
]. These activations were most likely triggered by a combination of several meteorological factors (e.g., rainfall and snowfall). The most recent and most intense activation (among all recorded ones) occurred in spring 2017. In total, ~160 large (>100,000 m3
) and hundreds of small (<100,000 m3
) landslides have been triggered or reactivated during this major event, displacing an estimated total volume of more than 82 million m3
]. A recent study has shown the evolution of the annual number of landslides in the region of Mailuu-Suu [7
]. According to related results, the highest yearly rates (before 2017) of landslides were observed for the years 2010 and 2016. Within a study area of 12,000 km2
located at the eastern rim of the Fergana Basin, 230 landslides—former or new—are recorded per year, affecting an overall area of more than 2 km2
]. These results support those published by Havenith et al. [5
], which already highlighted an increase in the total area impacted by landslides in that valley since 1962. As previously demonstrated by Havenith et al. [6
], this recurrent activation of mass movements is mainly observed during the spring season, resulting from a combined action of geological and climatic factors [6
]. Due to the significant societal, economic, and environmental impacts related to mass movements around the city of Maily-Say, understanding the mechanisms underlying the reactivation of landslides has become a key question to prevent massive losses due to landslides in that region.
Over the past decades, numerous remote sensing techniques have been developed and employed in this context. Differential Synthetic Aperture Radar (SAR) Interferometry (D-InSAR), which is widely used to monitor movements related to the activities of volcanoes, earthquakes, glacier retreats, and landslides, allows measurement of the displacements along the line-of-sight (LOS) of a satellite [8
]. As this technique is based on the distance changes between the sensor and targets located at the Earth surface, slight ground displacements, such as slow landslides, can nowadays be measured over wide areas with an accuracy of around a centimeter [14
While D-InSAR is suitable for the detection of slow vertical displacements, Digital Elevation Models (DEMs) and optical images can be used to identify fast horizontal movements. DEMs, which depict the topography of a land surface, are obtained from radar images or Unmanned Aerial Vehicle (UAV) acquisitions. Using pre- and post-failure DEMs, the accumulation and depletion zones of a landslide can be delimited [18
]. Moreover, vegetation indexes, such as the Normalized Difference Vegetation Index (NDVI), are widely used to study landslides [19
]. By comparing two images taken at different time periods, the difference in NDVI, pixel by pixel, highlights changes in vegetation cover linked to fast mass movements [24
]. Thus, optical data are useful for change detection and for the identification of visible features in 2D (map view). Depending on the spatial resolution, optical satellite images are also commonly used for the geomorphic mapping of inaccessible mountainous areas.
In this study, we employ multiple distinct and complementary methods to understand the recent collapse of the Koytash landslide and the deformation affecting the neighboring Tektonik mass movement (see view shown in Figure 2
). We detect terrain elevation changes related to both fast and slow displacements of the ground surface by using UAV-based imagery combined with radar and optical remote sensing techniques. In addition, we conducted a meteorological analysis to identify the triggering conditions, which led to slope instabilities.
2. Study Area
Kyrgyzstan is located in the middle of Central Asia and crossed by the geohazard-prone mountain range of the Tien Shan, with peaks reaching altitudes of 7000 m. According to the Ministry of Emergency Situations of the Kyrgyz Republic, more than 7% of the total territory, representing an area of approximately 15,000 km2
, could potentially be affected by landslides. The most susceptible regions to slope failure are located along the northern and eastern rim of the Fergana Basin, which is also marked by the highest population density [6
]. In this area, mass movements are concentrated in a range of altitudes between 700 and 2000 m. These landslides, often characterized by deep and steep scarps, mobilize weakly consolidated sediments of Tertiary or Quaternary age, including loess deposits [7
]. The Mailuu-Suu Valley is a former uranium mining area in the Jalal-Abad region, southern Kyrgyzstan at the north-eastern border of the Fergana Basin. This region is particularly prone to landslide hazards and, during the last 50 years, has experienced severe landslide disasters in the vicinity of numerous nuclear waste tailings. The high intensity of natural hazards in that area is due, on one hand, to its location between the mountain range and the basin marked by a relatively wet climate and, on the other hand, to the presence of soft geological materials near the surface [23
] (see the geological map of Maily-Say in Figure A1
). This type of sediment is prone to gravitational mass movements initiated by intense precipitations or strong seismic ground shaking [6
]. Due to the presence of radioactive wastes produced by past mining activities, this region is exposed to a high risk of environmental contamination and demands particular attention (Figure 3
). As a result, landslides represent a major threat to the local population of the small town of Maily-Say and of the regions downstream, not only through the direct impact, but also through indirect effects that lead to the creation of landslide dams that may block the river and cause the formation of a lake. Such lakes represent a further threat for downstream populations due to the potential failure of the landslide dam, causing the release of huge amounts of water resulting in floods. In the case of Maily-Say, the lake could also possibly reach the level of the nuclear waste tailings and thus cause their instability and the related mobilization of radioactive material. In 2003, Vandenhove et al. [30
] identified 23 mine tailings and 13 nuclear landfills for a total volume of approximately 3 million m3
(see tailing locations in the geological map in Figure A1
). Due to its critical situation, Mailuu-Suu Valley was and still is the target area of several international risk assessment projects [4
During the 1990s, three of the largest landslides of the Mailuu-Suu Valley (i.e., Koytash, Tektonik, and Isolith; the first two are shown in Figure 2
), displaced more than 5 million m3
of material [32
]. These three landslides are located near the core of the central anticline and represent the greatest threat for the destabilization of some nuclear waste tailings [6
]. According to Havenith et al. [6
], the Mailuu-Suu landslides now extend over an area of approximately 10 km2
. Havenith et al. [5
] also showed that distant earthquakes (epicentral distance > 100 km), even of a magnitude less than 6, can induce slight displacements on landslide slopes in the Mailuu-Suu Valley.
The Koytash landslide (see location in Figure 3
) has been extensively studied for many years as it has been active since the 50’s [5
]. Two elements indicate that this landslide is characterized by a roto-translational movement: First, the presence of a steeply dipping scarp at the crest of the landslide, showing slickensides, indicates sliding rather than a simple detachment as would be expected for a translational landslide. Second, the landslide body slides along a paleogenic limestone basement with a slope steeper than the ground surface. Although some translational sliding occurs above the limestone along the internal paleogenic clay layers, the base of the landslide is more than 50 m deep, which requires a rotational component. Furthermore, some anti-dip slope scarps in the lower part of the landslide also suggest some rotational movements. These observations are based on Electrical Resistivity Tomography (ERT) measurements completed by one of the authors (H.B. Havenith) together with the GEOPRIBOR team (led by I. Torgoev, Kyrgyzstan) [36
]. Havenith et al. [5
] evaluated its average annual speed at 60 cm/year and its volume at 5 million m3
. In 2009, Schlögel [37
] estimated the average speed of the movement at 1–2 mm/day and recognized Koytash as a major threat due to the possibility of a river dam. This roto-translational landslide was massively reactivated on the 22nd of April 2017 blocking the Mailuu-Suu River. Due to the damming of the valley, a lake formed with a depth of ~15 m, almost reaching the base of the nearest tailing. A few weeks later, after the dredging of the dam to open a channel for the river, the dam partially failed and the lake progressively disappeared [38
]. The rupture in spring 2017 also destroyed a few houses at the foothill of Koytash; these houses had already been evacuated due to the landslide risks [38
Second, in addition to Koytash, we studied the large Tektonik landslide (perspective view in Figure 2
and map location in Figure 3
). Tektonik is a complex (multi-rotational) and flow-like landslide with a length of more than one kilometer in an E–W direction. This landslide, initiated on 4 July 1992, displaced 2.5 million m3
of material and caused significant environmental damages [6
]. The main escarpment is located at an altitude of around 1350 m, while its lower part, divided into two lobes, joins the Mailuu-Suu River at an altitude of around 1000 m. A secondary scarp is located just below the main escarpment. Its activity is controlled by many factors, such as the loess cover in the upper part and the natural groundwater conditions that make it unstable [39
]. The main risk for the surrounding population is related to its high river damming potential. This landslide first collapsed massively a few weeks after the earthquake of 15 May 1992, which struck the region with a magnitude of 6.2 [5
]. The epicenter of this earthquake was located only 30 km SSE from the town of Maily-Say. Although the aforementioned earthquake certainly contributed to the development of the Tektonik landslide in combination with the rise of the groundwater table after spring precipitations, one must consider that loss of stability over longer terms was caused by the presence of underground mining galleries [5
]. In 1994, 2002, and 2005, the massive movements of the Tektonik landslide repeatedly lead to the blocking of the Mailuu-Suu River, forming a dam and causing numerous upstream and later also downstream floods [5
]. In 2017, massive failure occurred on Koytash landslide, while the Tektonik landslide had only slightly moved in its upper part. The materials and methods used to show the respective behaviors of the two mass movements are detailed in the following section.
In Central Asia, numerous studies conducted on mass movements have previously demonstrated the high landslide hazard level of Kyrgyzstan [6
]. In 2015, as part of a large-scale geohazards analysis, Havenith et al. [27
] performed an extensive landslide susceptibility mapping of the Tien Shan mountain range. Previous studies identified the Mailuu-Suu Valley as one of the regions particularly prone to slope instabilities [3
]. Both studies also confirmed that the number of landslides is increasing in this region, with a predominance of the reactivation and enlargement of existing landslides. These mass movements are likely to grow or merge with neighboring landslides when reactivating.
In this study, we used different types of satellite data to investigate the recent reactivation of two of the largest landslides of the Mailuu-Suu Valley, Kyrgyzstan. The Koytash and Tektonik landslides represent a major threat for the local populations due to their recurrent past reactivations, the presence of numerous neighboring nuclear waste tailings, and the risk of damming the Mailuu-Suu River [34
]. Although former remote sensing studies in the Mailuu-Suu Valley used satellite data to identify landslides through change detection methods [4
], this study brings a new dimension by using modern techniques, such as InSAR analysis and UAV imagery, to measure displacement rates and detect deformation zones.
D-InSAR was used to detect slow displacement rates of Koytash and Tektonik one year before their reactivation. In 2015, Teshebaeva et al. [12
] performed a D-InSAR analysis to detect slow-moving deep-seated landslides near the town of Uzgen, 90 km southeast of our study area. Similarly, they highlighted a strong correlation between deformations peaks and precipitations and revealed the continuous activity affecting the slope of the landslides in their area of interest. While they used an L-band Advanced Land Observing Satellite Phased Array type L-band SAR (ALOS PALSAR) data set, we processed the newly available C-band Sentinel-1 satellite images. Although, L-band radar sensors have a longer wavelength (i.e., λ = 23.8 cm), which partially penetrates surface vegetation and allows the measurement of significant DLOS
over long periods, their temporal resolution is of 46 days [67
]. Therefore, due to the low vegetation cover in our study area, we chose the open access C-band SAR images that have only 12 days of repeat cycle (over the Mailuu-Suu area), reducing potential temporal decorrelation in comparison to the former study. Our analysis showed that the upper part of Koytash was the most active with negative DLOS
values showing an early stage of subsidence along the scarp. Furthermore, the differential interferograms suggest that the reactivations of Koytash and Tektonik were not sudden but resulted from a long-term deformation. This could partly be related to a geological phenomenon called creeping, characterized by a slow downward movement of the soil. It is interesting to note that, during the entire study period, the displacement magnitudes of Koytash were greater than those measured for Tektonik. This result justifies why only the upper part of Tektonik collapsed while Koytash was entirely reactivated. The calculation of VLOS
also led to this conclusion. When looking at the lower part of Koytash, a transition to very small positive values was detected in March 2017. This may be linked to the fact that the displacements are oriented perpendicularly to the LOS of the satellite, making them difficultly detectable. A few months after the collapse, in August 2017, we noticed that the movement had slowed down, implying a gradual stabilization of the landslide. These results suggest that it took a maximum of four months for the landslide to stabilize after being active for at least nine months prior to the failure. Accordingly, it would be justified to perform a back-analysis of the last decades by using ERS data to produce supplementary interferograms and thus estimate when Koytash’s movement was initiated.
One of the most constraining limitations encountered during our D-InSAR analysis was the 1D expression of DLOS
inducing the systematic underestimation of intensities. Therefore, depending on the field properties, it is only possible to track positive and negative displacements along the LOS of the satellite while deformation in the orthogonal direction is not detectable. In ascending orbital geometry, the observed displacements and velocities are positive when approaching the satellite and negative when moving away from the satellite. Although the multitemporal analysis with the FASTVEL algorithm requires at least 25 images, it is only able to detect mean velocity patterns (or hotspot areas as defined in [59
]) that are significantly changing over such a long temporal coverage. Moreover, the velocity rates and directions are evolving overtime in a complex and vegetated terrain meaning that only the relative spatial changes are comparable to detect hazardous areas. To fully apprehend the complex deformation behaviors and facilitate their interpretation, sophisticated multitemporal InSAR techniques with various sensors (multi-angle) [48
], ideally combined with in situ measurements, enable to break down the displacements into 3D in order to measure the true direction of the ground movement. Fuhrmann & Garthwaite [68
] demonstrated the advantages of combining LOS measurements from different viewing geometries. Resolving 3D surface displacements would provide an improved overview of the complex geomorphology and dynamic of landslides. It would also be interesting to exploit additional SAR processing chains such as the SBAS (Small BAseline Subset) [69
] or PSI multi-time series [62
] to go one step further in multitemporal InSAR analysis. Another frequently encountered error in D-InSAR analysis is related to phase unwrapping and often leads to the misinterpretation of results [72
]. In the case of Koytash and Tektonik, several parameters, such as the topography of the region and significant variations in the deformation rates, made the phase unwrapping a complex procedure. The topographic component, including both the roughness and vegetation cover of the landslide, can create artifacts (e.g., layover or foreshortening) responsible for sudden variations in slopes. Thus, an incorrectly unwrapped interferometric phase induces inaccurate displacement rates due to possible phase jumps (2π) [11
]. To overcome this limitation, in this study the unwrapping process was initiated from a reference point (DGPS) located in a stable and coherent area, outside the landslide boundaries. As demonstrated by Schlögel et al. [73
] and Teshebaeva et al. [12
], the use of L-band in combination to C-band SAR images could help obtain better results by reducing time decorrelation effects. L-band acquisitions may be more suitable for highly vegetated areas as well as larger displacements. Another limitation of D-InSAR is the loss of coherence when confronted to rapid displacements or changes in land cover (e.g., agriculture). Additionally, by increasing the frequency of SAR acquisitions, the temporal decorrelation decreases as well. In the case of Mailuu-Suu, the use of Sentinel-1A, with only 12 days of repeat cycle, enabled us to observe some precursors before the main reactivation. It is noteworthy that for certain well-covered areas, the combination of Sentinel-1A and Sentinel-1B allows shorter return periods of 6 days. However, the effectiveness of this technique is counterbalanced by its numerous limitations. The main sources of decorrelation are linked to meteorological conditions, spatial and temporal resolutions, magnitudes, velocities, and orientation of the movements, or even variations in soil properties.
Although radar analyses can only highlight slow displacements because the limit of detectability is a function of the wavelength of the sensor, abrupt topographic modifications as well as the evolution of the land cover can be underlined through change detection. This can be achieved using several techniques such as the difference in DEMs, NDVIs or with a multitemporal optical analysis. Variations in surface topography can be used to delimit the depletion and accumulation zones. Our results, based on TanDEM-X and UAV DEMs, revealed that the reactivation of Koytash was characterized by a combined rotational and translational movement. Indeed, soft rocks detached from the scarp (depletion zone) applied an intense, almost vertical, pressure on the material below which caused an uplift in the foot (accumulation zone). By contrast, due to the superficial deformation of the Tektonik landslide, we could assume that the displacement was more likely translational. The performance of this technique is based on the quality of the DEMs. Depending on their coordinate systems and their georeferencing, the comparison and the difference between the products will be more or less exploitable. In the case of the Mailuu-Suu Valley, DGPS points, collected in August 2017 during the field campaign, were used to adjust raw DEMs and to ensure the accuracy of the results. However, such corrections can sometimes lead to significant errors, which must be considered. Therefore, it is essential to test the overlapping compatibility of the input files. Moreover, even though DEMs created on the basis of drone acquisitions have highly superior resolution, they can only be collected on a local scale because they are extremely time-consuming. Nevertheless, high-resolution DEMs are widely used to estimate the volume of displaced geological material [18
Studying changes in land cover using NDVI is frequently used to survey the evolution of landslides and has been extensively applied to landslide monitoring studies [19
]. A NDVI analysis of the Mailuu-Suu region, based on five landslide inventories spanning the past 50 years (1962, 1984, 1996, 2002, and 2007), demonstrated the efficiency of using NDVI to detect new sliding activation [23
]. In this study, we used vegetation indices to highlight deformation zones and verify if our findings coincided with the D-InSAR results. On the basis of the difference of NDVI map, we observed bright areas (i.e., soil denudation) and darker zones (i.e., growing vegetation). As expected, bright areas were identified within the landslides’ boundaries, but also outside the collapsed zones. These might indicate the seasonal variability marked by a strong difference in the vegetation cover depending on the periods of acquisition of the two images (e.g., end of spring (June) and end of summer (October)). Normalization, pixel-by-pixel, of the images by the average NDVI value could possibly mitigate this seasonal effect. However, many parameters (e.g., humidity and anthropogenic changes in land use) must also be considered in order to distinguish the zones that are truly reactivated. Nonetheless, in the case of this study, NDVI was used as a complementary qualitative analysis to verify if the landslide boundaries coincided with important changes in vegetation.
The reactivation of Koytash and Tektonik was also observed in the multitemporal optical analysis showing the evolution of the landslides since 2007. Koytash’s recent reactivation can be considered as the most important failure event since the first movements were detected in the 1960s [76
]. This massive collapse led to the obstruction of the Mailuu-Suu River forming a dam and subsequently a lake. As the level of water rises, the lake becomes prone to contamination from an upstream uranium tailing and thus represents an important environmental risk for potential water pollution. In 2017, the level of the lake almost reached the toe of this nuclear waste tailing. The formation of this lake may also be a direct threat to the local populations due to related floods.
Considering the actual climatic context, this study highlighted the importance of understanding the major triggering role of meteorological factors affecting the (re)activation of landslides. To support the remote sensing analysis, and in order to better understand the influence of weather on natural hazards occurrence in Central Asia, we analyzed different potential triggers of the activation. The meteorological analysis performed in the frame of this study demonstrated the triggering role of rainfall combined with the rapid snow melt in the reactivation process as it had already been demonstrated for Kyrgyzstan [77
] but also for other regions [78
]. The significant concentration of rain as well as the sudden snowmelt contributed to the water saturation of the soil before Koytash’s failure. These predominant weather settings may have conditioned the deformation just before the reactivation, by increasing the pore pressure and thus the mobility of the sediments. Both the Koytash and Tektonik landslides are located in the bed of the Mailuu-Suu valley (Figure 3
a). Therefore, during rainfall and the melting of snow caps upstream, water flows through the valley following the Mailuu-Suu River. Furthermore, the landslides are located on steep slopes with overhanging mountains from where surface runoff also flows after the snow melts and intense rainfall. Due to the triggering role of precipitations, it would be justified to automate the daily calculation of meteorological statistics and develop an early warning system to alert the surrounding populations in case of an upcoming disaster. For example, in Africa, Monsieurs et al. [66
] were able to determine a rainfall threshold over which landslide susceptibility increases considerably.
With a perpetual increasing number of landslides, studies at regional scale become more and more necessary. With the current technological advances in satellite data, we are now able to process larger amounts of data and span wider areas. However, local and detailed analyses remain crucial when considering landslides such as Koytash and Tektonik that are recurrently reactivated and represent an important threat to surrounding populations. Due to the great density of landslides and the related risks, the Mailuu-Suu region is and will remain a critical area requiring continuous monitoring.
The aim of this study was to investigate the recent behavior of two landslides—Koytash and Tektonik—in the Mailuu-Suu Valley, Kyrgyzstan. This study was divided into three main objectives: a comparison of multitemporal DEMs, a D-InSAR analysis, and an analysis by optical remote sensing. To better understand the geo-environmental conditions causing the reactivation of the Koytash and Tektonik landslides, we also performed a complementary meteorological analysis. We showed that intense precipitations as well as the rapid melting of an important snowcap contributed to the activation of these two and many other landslides in this region in 2017. However, by creating displacement and velocity maps, the InSAR analysis revealed signs of sliding activity long before the last rupture of Koytash and Tektonik. This indicates that these landslides result from long-term deformations partly related to creeping. Indeed, InSAR allows the measurement of small magnitude displacements that are often not visible in the field. Although this method is not able to detect rapid movements (e.g., the collapse of landslides), InSAR can be used to monitor large-scale areas and identify long-term precursory deformations. Another major limitation is that InSAR measures displacements along the LOS that cannot be interpreted in 3D. Nonetheless, due to the open access availability of SAR acquisitions with short repeat cycles (e.g., 6 days for Sentinel 1A & B), InSAR is a powerful tool to monitor areas prone to landslides and is thus playing a key role in risk management.
Unlike InSAR, NDVI is a change detection method capable of highlighting the rapid evolution of vegetation. The difference of NDVI, between 2016 and 2017, identified several areas of landslide related land cover change, including the failure of Koytash and Tektonik. Nevertheless, anthropogenic changes in land use or seasonal variations in vegetation cover, depending on the acquisition date, may lead to misinterpretations of the results. Therefore, as shown in this study, NDVI can be used as a complementary qualitative analysis for landslide monitoring, capable of confirming deformations zones identified through InSAR. Similarly to NDVI, the multitemporal optical analysis highlighted deformation zones. The use of high-resolution optical images allows the delimitation of landslide boundaries, enabling precise surface calculations, which is not always possible with other remote sensing techniques. By mapping the evolution of the landslides in the Mailuu-Suu Valley, this optical analysis revealed that 30% of total landslide surface was reactivated in 2017. Additionally to the previous methods, we created a post-landslide DEM using UAV imagery, only possible at a local scale. The comparison of the latter with a pre-landslide TanDEM-X DEM, allowed us to determine a maximal accumulation of 30.13 m and a maximal depletion of 41.4 m. This evolution of the topography demonstrated that the collapse of the upper part of Koytash pushed on the underlying layers, which transferred the pressure downslope, resulting in an uplift due to the push-up effect and sliding material aggradation. When studying recent reactivations, post-landslide DEMs are rarely available, and it is therefore necessary to create a new DEM. In conclusion, the combination of radar and optical data revealed the recent landslide activity while the meteorological analysis helped us to identify the triggering factors responsible for the collapse of numerous landslides in Kyrgyzstan in spring 2017. The inherent limitations of the remote sensing methods used in this work, despite their considerable advantages, show that it is useful and essential to combine each of the techniques in order to obtain complete and conclusive results. It is therefore crucial to pursue the study of mass movements, as well as the development of novel cutting-edge techniques, to better understand their triggering conditions and, eventually, anticipate their rupture.