Introduction

Natural damming of rivers by landslides can have disastrous consequences, posing a major hazard in mountainous areas dissected by deep, narrow valleys (Fan et al. 2020). Some of the existing landslide-dammed lakes in Central Asia can be considered stable and relatively safe. On the other hand, other river blockages represent a very large hazard, as catastrophic outburst floods can occur (Strom 2010). The main concern is related to the occurrence of future large-magnitude earthquakes that could trigger material liquefaction and consequent collapse of the dam, which, finally, could unleash a catastrophic flood, whose destructive power would be exacerbated within steep and narrow valleys. Lake banks can also host slope instabilities: large-scale failure may occur, causing large waves that can overtop the dam (Barla and Paronuzzi 2013) or lead to complete or partial dam breach. One example of a large-scale failure that occurred in Italy in 1960 is the Vajont arch dam, which created a wave of 50 million m3 of water that flooded the Piave valley and town of Longarone (Mantovani and Vita-Finzi 2003). Long-term stability assessments of these dams and lake banks are of paramount importance, as they represent a significant threat to the communities living there.

Tajikistan is a region of Central Asia characterized by approximately 1300 lakes, mostly resulting from rockfalls: landslides have occurred in the Tajikistan area mainly due to very strong earthquakes and moraine deposit collapse (Ischuk 2006). The Usoi dam, which generated Lake Sarez on the eastern side of the country, was created by rockfalls and rock avalanches that were predominantly composed of crushed rocks with a wide grain size distribution and that are characterized by rapid movements and large volumes of fallen material of approximately 106 m3 (Fan et al. 2020). The Usoi landslide dam is the tallest in the world (approximately 600 m) among natural and artificial dams (Droz et al. 2008; Schuster and Alford 2004). Landslide dams and the related lakes usually have a limited lifetime, which can range from a few minutes to thousands of years, depending on several factors, such as the type of material, its geotechnical properties, its volume, the morphology of the area, and the characteristics of the river network (Tacconi Stefanelli et al. 2015, 2020). Very often, they fail in short periods due to their heterogeneous composition, marked by low density and easily liquefiable and fine-grained materials. Worldwide, there are many other notable examples of landslide dams in addition to the Usoi landslide, which is extremely long-lived (Schuster and Alford 2004): a famous dam is located in southwestern Iran, i.e., the Şīmareh landslide, which was generated after an earthquake that occurred at approximately 10,000 years B.P., and the landslide created two lakes, which have disappeared; in northern New Zealand, the Waikaremoana dam is approximately 2000 years old because of its resistance to erosion; a third dam is located at Lake Yashinkul on the Tegermach River in the Kyrgyz Republic, and it is characterized by rock blocks and falling debris that occurred between 1835 and 1966 (Schuster and Alford 2004).

Tajikistan is a mountainous area that is not easy to access or work in due to its topographic complexity. Therefore, this inaccessibility and the country’s economic and political difficulties have hampered the possibility of applying engineering solutions and studying the area with the necessary detail. Therefore, remote sensing plays an important role in risk assessment and monitoring of the region (Grebby et al. 2021; Schuster and Alford 2004). Earth observation (EO) data and techniques can represent valuable tools for assessing and measuring ground deformation at a wide scale, especially where the remoteness and inaccessibility of the sites make field surveys extremely difficult. In this scenario, SAR (synthetic aperture radar) techniques have a major role to play, as they have been successfully demonstrated to be highly valuable in mapping land motion (Crosetto et al. 2016 and reference therein), allowing us to measure surface deformations of wide areas with millimetre to centimetre accuracy and at a frequency varying between a few to several days with most recent satellite platforms. Interferometric synthetic aperture radar (InSAR) has proven to be a useful tool for studying and tracking ground surface deformation. Another useful method to study the displacement and movement of a mountainous area is based on optical satellite imagery. The numerical estimation of the displacement can be obtained through methods that correlate pairs of images acquired at different times, such as the Co-Registration of Optically Sensed Images and Correlation (COSI-Corr, Leprince et al. 2007) tool. In the literature, some examples of COSI-Corr and optical remote sensing applications to landslide cases are reported: Yang et al. (2019) provided an application to study the slope movement of the Baige landslide that blocked the Jinsha River (China) in 2018, while Hermle et al. (2021) employed PlanetScope images and a UAS (unmanned aerial system) to reveal the ground displacement and acceleration of complex mass movement in the alpine sector in Austria.

The aim of this article is to provide a detailed overview of the ground deformation of the wide area of Lake Sarez, including its surrounding banks and the Usoi dam, by adopting different remote sensing practices, such as InSAR using the Sentinel-1 dataset and an image correlation applied to SPOT-6 and SPOT-7 acquisitions. In this way, a synoptic and complete analysis of the ongoing displacements was retrieved, allowing us to reconstruct the temporal evolution (2014–2020) and to solve the spatial variability of the deformation affecting the Lake Sarez banks. In addition, since the volume of the landslide is still widely debated, a model of the geometry and the depth of the sliding surface of a potential landslide dam has been proposed by applying Carter and Bentley’s method (1985) derived from remote sensing data. The multiperspective analysis performed here may represent a solid base for the reduction and mitigation of geohazard risks, especially in impervious areas.

Study area

The area of Lake Sarez (Fig. 1) is a particularly relevant site, as it is affected by different hazards (seismic, landslide, and flood) with the potential for cascading effects: indeed, Tajikistan is located in one of the most seismically active regions of the world (Droz and Spasic-Gril 2002) within the Euro-Asian and Indian tectonic plate collision zone (Ischuk 2011). Located in the Rushan and Murgab districts of Gorno-Badakhshan Autonomous Oblast (Pamirs, Tajikistan) along the Murghab River, the Usoi dam (Fig. 1) is one of the most hazardous areas in the country. The impounded Lake Sarez, with a depth of 500 m, is the world’s deepest landslide-dammed lake (Costa and Schuster 1991). The area is characterized by a highland plateau at 3500–5000 m a.s.l., with a few warm periods and long severe winters and the presence of snow for most of the year.

Fig. 1
figure 1

Location of the Usoi dam and Lake Sarez within the Murghab River valley in eastern Tajikistan. The extension of the areas processed with Sentinel-1 and SPOT images is also indicated

Full size image

Lake Sarez (Fig. 2), which is 60 km long and has a stored volume of approximately 17 billion m3, originated on February 18, 1911, when a MW 7.7 earthquake generated a giant wedge failure of approximately 2.2 billion m3 of rock and debris that blocked the Murgab River and a tributary valley; this event formed the 560-m-high Usoi dam, impounded Lake Sarez, and created the smaller Lake Shadau. The blockage, named after the small village with 54 inhabitants that was buried, is 5 km long and 4 km wide and blocks the Murghab River at an elevation higher than 3000 m.

Fig. 2
figure 2

Above: SPOT-6 view of Lake Sarez and Usoi landslide dam in Tajikistan: on the left, view of the Usoi dam, with seepage in its body. Below: frontal view of the lake, with the name “right-bank landslide” (from Strom 2010), whose failure may cause a surge wave in the lake (photo from www.ey8mm.com)

Full size image

After the earthquake in 1911, the area drew immediate attention, and despite the inaccessibility and remoteness, some researchers started preliminary studies to estimate the volume, approximately 2.2 km3, and the mass, 6 × 109 tons. Based on these data, Preobrazhenskiy (1920) and Galitzin (1915) calculated the potential energy released by the landslide and concluded that it would be sufficient to produce the seismic amplitudes recorded at the Pulkova seismic station ∼ 3800 km away (Kulikova et al. 2016).

Geologically, the main part of the landslide body consists of terrigenous carbonate deposits, with quartzite, sandstone, and schist of the Carboniferous Sarez Formation, and the northern part is composed of Permian–Triassic marble and shale, with subordinate gypsum, anhydrite, and dolomite (Ischuk 2006). There is no information about the internal composition of the dam. Due to the landslide, the river has been completely blocked and lacked the capability to cut the landslide deposits to create a natural outlet. However, in approximately 1914, the water found its way through the landslide deposit in the uppermost and most permeable part of the dam, generating a spring approximately 140 m below the water level (Strom 2010). Since 1925, significant filtration from the dam was observed and has created a canyon in the landslide dam. The lake level is currently rising at an average of 0.2 m per year (Schuster and Alford 2004), and approximately 50 to 60 m3/s of water leaks through the dam body (Risley et al. 2006), rising to 85 m3/s during flood periods, when the water level increases (UN International Strategy for Disaster Reduction 2010).

Both the left and right sides of the bank have the same rock composition and belong to the Carboniferous Sarez Formation, with the left side being less complex from a tectonic point of view. Both sides are affected by slope movements (Fig. 2).

In the worst-case scenario that assumes the collapse of the dam (extremely unlikely), a catastrophic outburst flood from Lake Sarez would destroy the villages and infrastructure in the Amu Darya basin between the lake and the Aral Sea, endangering tens or possibly hundreds of thousands of people in the Murgab, Bartang, Panj, and Amu Darya valleys downstream across a distance of over 2000 km. The most endangered people would be those in the villages and towns along the lower Bartang River in Tajikistan (Barchidiv, Supomji, Shojan, and Rushan) and along the Panj River, which forms the Tajik–Afghan border, because these mountain valleys are narrow and there would be short warning times. An assessment of the flood scenario is presented by Schuster and Alford (2004). A general evaluation of the dam and slope stabilities with detailed geotechnical studies and using modern methods and equipment was necessary.

The Lake Sarez Risk Mitigation Project, funded by the World Bank, has been created and aims to reduce the landslide risk, setting up an early warning system with the idea of making people aware of the risk and the impact that an event could cause to ensure that inhabitants know how to behave during a possible emergency (Droz and Spasic-Gril 2006). For this reason, the dam and the lake banks are closely monitored with in situ instruments. The monitoring network and the early warning system are expected to protect the villages located along the Murgab and Bartang rivers and reduce the vulnerability of the population to natural disasters, including the potential outburst of Lake Sarez. Due to the enormous consequence that a possible event could entail, it is important to take into account the relevant data that have been obtained in the area until the present day, and the landslide has received significant attention from the scientific community (Papyrin 2001; Kazakov 2004; Ischuk 2006). Hanisch and Soder (2000) said that the movement calculated was about 1–2 cm in 9 months; Russian experts define the movement around 10 cm/year, and Ischuk (2006) said that probably this movement is related especially to the superficial part of the landslide. Droz et al. (2008) calculated the movement around 6–16 cm in 2 years (from 2004 to 2006) with the help of PSInSAR technique (Shakya 2020).

While the danger of a general Usoi dam failure caused by water pressure, internal erosion, or seepage was found to be low (Ischuk 2006), the hazard of an overtopping wave from new landslide masses falling into Lake Sarez is considered more relevant (Strom 2014 and reference therein).

In particular, a major concern is represented by the right-bank landslide, which is located approximately 4 km upstream of the dam (Fig. 2). This bank is part of the anticlinal complex of the Muskhol range, with an approximate E‒W direction. Geologically, the upper and lower parts of the bank present mainly sandstones, and the central part is also characterized by detrital shales. The main bedding is predominantly NE-dipping, and a slope of 10–30° leads to break lines in the bedrock. The right-bank landslide is formed by loose material such as silty, sandy, and blocky materials along with glacial till deposits that demonstrated cementation as a result of the challenging climatic circumstances. The carboniferous bedrock of the Sarez Formation, which is made of sandstones, schists, and slates, is outcropping. The landslip was divided into two pieces by Ischuck (2006): the NW section, which was less active and less likely to collapse, and the SE section, which was more active and more likely to collapse into the lake. Based on the geological conditions and slope features, Raetzo (2006) identified four sections: the south, middle, north, and top parts. The SE suggested by Ischuck (2006) and the south suggested by Raezto (2006) are congruent. Papyrin (2001) argued that there was also a section on the left bank side part of the Lake that could collapse in the lake, with smaller size (Shakya 2020).

The landslide has a width along the lake shore of approximately 1 km. The estimated volume ranges from 300 million to 2 billion m3 (Schuster and Alford 2004; State Committee on Emergencies 1997, 1999). The large estimated volume range exists both because the thickness of the landslide is very uncertain and because potential movement (single, large, monolithic landslide or smaller individual landslides) is still unclear.

The whole area is characterized by high seismic activity (Ambraseys and Bilham et al. 2012), potentially leading to landslide slumping into the lake, creating a large wave that could overtop and possibly breach the Usoi dam, creating a destructive flood downstream. On 7 December 2015, an earthquake with a magnitude of 7.2 occurred in the Rushon district of the Gorno-Badakhshan Autonomous Oblast in Tajikistan at 12:50 Tajikistan local time, approximately 10–20 km from Sarez Lake and Usoi dam in Tajikistan. This earthquake is comparable and similar to the 1911 event (Kulikova et al. 2016). After this earthquake, four aftershocks were recorded near the lake that created many difficulties and damages, especially for the people who lived there (it killed two people, and more than a thousand people became homeless), but the landslide was not significantly affected by the event (Metzger et al. 2017). Along the valley of the Murghab River (a headwater tributary of the Amu Darya River basin, also known as Bartang from the junction with the Ghunda River immediately below Sarez Lake), there are several villages along both sides of the valley.

Methods

To completely depict the deformation scenario of Lake Sarez, an integrated satellite analysis was applied to extract as much information as possible about the slope instability affecting the area of interest. The necessity of a synergistic approach derives from the complexity of the displacement pattern that is expected by the landslides in this area (Raetzo et al. 2006). The purpose of the analysis of satellite radar data was twofold:

  1. 1.

    Detect, record, and map any slow deformation phenomena (from mm up to several cm per year) that potentially affect the surroundings of Lake Sarez, using advanced InSAR techniques.

  2. 2.

    Detect and measure any surface changes produced faster displacement (from metres to hundreds of metres per year), using the correlation of pairs of optical images

SAR

Relying on active radar sensors, interferometric applications, which is an overarching term referring to the exploitation of the SAR signals of at least two complex-valued SAR images (Bamler and Hartl 1998), currently represent the most consolidated approach to measure and quantify ground deformation induced by landslide occurrence (Scaioni et al. 2014), as they have the unique ability to obtain measurements anytime, regardless of the time of day or season.

Satellite InSAR has also been widely used since the 1990s (Achache et al. 1996; Fruneau et al. 1996) to measure the spatial extent and magnitude of surface deformation associated with mass movements (Solari et al. 2020). Multitemporal approaches (MTInSAR) have been demonstrated to be highly valuable in analysing a wide range of geological and geomorphological phenomena, including landslide-related events at different stages (Tofani et al. 2013), and it has become one of the most widely adopted and reliable methods for the remote detection of landslides. The ability to retrieve numerous MPs (measurement points) and to track subcentimetre deformations of landslides has made MTInSAR a supporting tool for the detection and mapping of actively deforming slopes (Cigna et al. 2013), the characterization of landslide mechanisms, the zonation of sectors with different velocities and behaviours within the landslide area (Berti et al. 2013), the modelling of large slope instabilities (Berardino et al. 2003), and the monitoring of the evolution of a single event (Meng et al. 2021), which eventually lead to catastrophic failures (Carlà et al. 2019).

Satellite-based InSAR can be used to detect and map landslides in remote areas and in mountainous terrain or in general where the deployment of ground-based instruments is not logistically feasible and where in situ activities are challenging. This aspect is important for mapping landslides without prior knowledge of their location (Bekaert et al. 2020) and applies particularly to areas with permafrost (Meng et al. 2022) or seasonally frozen ground (Hao et al. 2019), whose vulnerability to geohazards, including landslides, is expected to increase with rapid warming (Yao et al. 2019; Cui and Jia 2015). In the last 2 decades, the use of interferometric applications has been fostered by the launches of several satellite platforms hosting sensors working within specific bands of the microwave domain, corresponding to different wavelengths (λ). The most commonly used bands in SAR applications are the C-band (5–6 GHz, ∼ 5.6 cm wavelength), X-band (8–12 GHz, ∼3.1 cm wavelength), and L-band (1–2 GHz, ∼23 cm wavelength). An exhaustive list of SAR sensors is presented by Wasowski and Bovenga (2014).

The geometry of the acquisition of the radar sensors allows two different velocities, the ascending (VA) and descending (VD) velocities, which provide two ground velocity components, i.e., vertical (VV) and east–west (VE), assuming a negligible north‒south component. Using the ascending and descending raster maps and considering the line of sight (LOS), it is possible to estimate the vertical and E‒W velocities by applying the following formulas:

$${V}_{A}={V}_{V}\;{\text{cos}}\;{\theta }_{A}+{V}_{E}\;{\text{sin}}\;{\alpha }_{A}\;{\text{sin}}\;{\theta }_{A}$$
$${V}_{D}={V}_{V}\;{\text{cos}}\;{\theta }_{D}+{V}_{E}\;{\text{sin}}\;{\alpha }_{D}\;{\text{sin}}\;{\theta }_{D}$$

with \({\alpha }_{A}\), \({\alpha }_{D}\), \({\theta }_{A}\), and \({\theta }_{D}\) representing the ascending, descending, and LOS azimuths and the look angles. The horizontal (Vh) and vertical (Vv) components can be obtained by combining the descending and ascending velocities and data through the following equations:

$${V}_{h}= \frac{{V}_{A}\;{\text{cos}}\;{\theta }_{D}-{V}_{D}\;{\text{cos}}\;{\theta }_{A}}{{\text{sin}}\;{\theta }_{A}-{\text{cos}}\;{\theta }_{D}-{\text{sin}}\;{\theta }_{D}\;{\text{cos}}\;{\theta }_{A}}$$
$${V}_{v}= \frac{{V}_{A}\;{\text{sin}}\;{\theta }_{D}-{V}_{D}\;{\text{sin}}\;{\theta }_{A}}{{\text{cos}}\;{\theta }_{A}-{\text{sin}}\;{\theta }_{D}-{\text{cos}}\;{\theta }_{D}\;{\text{sin}}\;{\theta }_{A}}$$

SqueeSAR

SqueeSAR (Ferretti et al. 2011) is a high-resolution multi-interferogram approach based on SAR imagery capable of identifying both permanent scatterers (PS) and distributed scatterers (DSs) to monitor and study ground displacement in a given area without altering the traditional process of permanent scatterer SAR interferometry (PSInSAR), despite their different behaviours from a statistical point of view. It could be applied both in urban and extra-urban areas also where PS data are not enough. The main assumption of the SqueeSAR technique is to identify sets of pixels sharing the same radar backscattered signal, which basically corresponds to statistically homogeneous pixels. This step is achieved through a Kolmogorov-Smirnov (KS) test which is based on the amplitude value of a co-registered (overlapped) and calibrated stack of SAR images. Thus, DSs (pixels with commonly low level of reflectivity) are identified according to the following steps: (i) analysis of single pixels of the image; (ii) creation of a window around the pixel; (iii) comparison of adjacent pixels with the KS test; (iv) processing and analysis of statistically homogeneous pixels, while pixels with different distribution functions are discarded; (v) DSs identified within statistically homogeneous areas are processed using the traditional PSInSAR algorithm (Ferretti et al. 2001) for the estimation of the deformation maps and the construction of displacement time series of each measurement point (MP). This approach is characterized by two important and innovative concepts: the first is the DespecKS algorithm, which is used to identify statistically homogeneous pixel families (SHP) to evaluate the statistical correlation of DS by taking into account PS information, and the second is the phase triangulation algorithm (PTA), which is a link between DS and PS and helps determine the DS value. The technique uses and processes multitemporal SAR satellite images of the area and returns high-quality displacement measurements. This method uses a grid of pixels that allows data acquisition within a few millimetres of accuracy for surface displacement. The technique measures the projection of the displacement vector along the satellite line of sight (LOS), and by using different results from different acquisition geometries in the same time interval and area, 2-D data can be measured along the east‒west and vertical directions. Both the descending and ascending orbits must be visible from the target, although this is not easy to find in real scenarios. To overcome this problem, the area is divided into small cells, and a pseudo-PS is created to average the measurement points within the same cell (Bischoff et al. 2020). Then, for each cell, the results and displacement estimates are combined. The flow chart of this method is in Fig. 3.

Fig. 3
figure 3

Flow chart of the COSI-Corr (above) and the SqueeSAR (below) methods

Full size image

Similar to any radar-based approach, the SqueeSAR technique measures the component of deformation parallel to the LOS, so it is not easy to retrieve complete information about slope movements. For the E–W-oriented slope, it is possible to combine the ascending and descending components to obtain a complete overview of the area. When this is not possible because the slope is not oriented E‒W, the VLOS (velocity along the LOS) is projected along the steepest slope (Notti et al. 2014). The velocity along the slope, called Vslope, is calculated in relation to the VLOS and the C-index.

$$V_\textit{slope}=\frac{v_\textit{LOS}}C$$

The C-index is a parameter that expresses the percentage of movement registered by an SAR sensor and is defined as

$$C=\left({\text{cos}}\;S\times {\text{sin}}\;(A- \frac{\pi }{2})\times N)+(-1\;{\text{cos}}\;S\times {\text{cos}}\;(A-\frac{\pi }{2}\right)\times E)+({\text{sin}}\;S\times H)$$

where N, E, and H are the line of sight (LOS) versors of the satellite, which are different in the descendent and ascendent geometry; S is the slope angle; and A is the slope aspect (Confuorto et al. 2022).

For the analysis of the ground deformation of the Usoi dam and Lake Sarez, the ESA (European Space Agency) Sentinel-1 constellation was used. Launched in April 2014, the Sentinel-1 sensors opened new possibilities for InSAR applications. The mission is composed of two twin satellites, Sentinel-1A and Sentinel-1B, which share the same orbital plane and offer an effective revisiting time of a few days (six-day repeat cycle in Europe and other specific areas; twelve days globally), which is extremely suitable for interferometric applications. Therefore, considering the previous SAR satellites, Sentinel-1 data exhibit regional-scale mapping capability, systematic and regular SAR observations and rapid product delivery (typically in less than 3 h from data acquisition). Sentinel-1 SAR products are freely accessible, thus providing the scientific community, as well as public and private companies, with consistent archives of openly available radar data suitable for monitoring applications (Raspini et al. 2018). A total of 91 C-band Sentinel-1 acquisitions were processed with the SqueeSAR approach, covering the period from 11 March 2016–19 February 2020.

COSI-Corr

COSI-Corr (Leprince et al. 2007) is an ENVI-integrated module designed to obtain accurate geometrical processing of both satellite and aerial optical images through orthorectification, co-registration, and correlation and has been developed to identify ground deformation from multitemporal acquisitions. In particular, it is used for slow landslides (Lacroix et al. 2018), glacier flows (Jawak et al. 2018), and coseismic deformation (Zinke et al. 2019) and can also be applied to change detections that require an accurate registration of the images (Ayoub et al. 2009). The COSI-Corr method is usually used for the analysis of seismic ground deformation and glacier movement and has the possibility to study displacement at the subpixel level (Zinke et al. 2014). The technique consists of comparing two input images, one before and one after the movement, and after identifying the stable areas using the ground control points (GCPs), the software provides two displacement layers for each pair of images, the north‒south (N/S) and the east‒west (E/W) layers. A quality parameter, i.e., the signal/noise ratio (SNR), is also produced. For the N/S displacement, a positive value indicates the north direction, and a negative value indicates the south direction; in the E/W direction, the east is represented by a positive value, and the west is represented by a negative value. It is important to underline that the generation of GCPs is achieved without reliance on external data. This is accomplished by employing a shaded image of the digital elevation model (DEM) as the initial ortho-rectified master. Subsequently, the first ortho-rectified image generated serves as the master reference for subsequent slave images (Leprince et al. 2007). To correlate the images and to provide the displacement, there are two types of correlators: frequency and statistical. The first is based on the Fourier algorithm, it is more precise and accurate, and it is also more sensitive to noise, so it is useful for producing images of good quality. The statistical correlator, instead, is coarser, and it is applied in noisy images that have already been studied with the frequency correlator but produced bad results (Yang et al. 2019). In the frequency correlator, it is possible to obtain different results matching different combinations of initial–final window sizes: with the increase in the window sizes, the noise and the uncertainties decrease, and the results become more reliable. The flow chart of this technique is in Fig. 3. Four high-resolution (1.5 m) SPOT-6 and SPOT-7 optical images were chosen to identify the displacement of the area. The images are from September/October 2015, 2017, 2019, and 2021, because in the remaining months, the area is usually covered by snow.

Results

Both sides of Lake Sarez are affected by slope instabilities. The two main landslides that are the objects of in-depth analyses have different extents: the right-bank landslide (named RB in Fig. 4), which is located on the northern flank of Lake Sarez, has an area of 5.34 km2, while the landslide located on the southern flank (named LB in Fig. 4) has an area of 1.63 km2. Many findings of the application of SqueeSAR and image correlation techniques applied to the Sentinel-1 and SPOT-6/7 satellites, respectively, are described in the following sections.

Fig. 4
figure 4

Ground deformation maps for Lake Sarez along ascending (above) and descending (below) geometries. PS and DS were detected and classified according to their mean annual LOS velocities. RB, right-bank landslide; LB, left-bank landslide; UL, Usoi landslide; UD, Usoi dam. The purple line is the boundary of the main landslides

Full size image

InSAR

Ground deformation maps in ascending and descending geometries produced with the SqueeSAR processing of the Sentinel-1A acquisitions are shown in Fig. 4. PS and DS were detected and classified according to their mean annual LOS velocities. Given the lack of vegetation, Lake Sarez is a suitable site for InSAR processing, with only snow cover limiting its effectiveness. With almost 300,000 points for each acquisition geometry, these maps include a wealth of information that can be exploited to scan wide areas, spot instabilities, and reconstruct their deformation histories back to 2016.

While the Usoi dam (UD in Fig. 4) shows moderate to low deformation rates (approximately 8000 points in each geometry, with velocities ranging between − 15 mm/year and 15 mm/year), PS data show clear patterns of active displacement along the sides of Lake Sarez, both in ascending and descending geometries, confirming literature information.

For the right-bank landslide (Fig. 4), ca. 4500 PSs are available for the ascending geometry (with a PS density of approximately 1000 PS/km2), and 6100 PSs are available for the descending geometry (with a PS density of approximately 900 PS/km2). Given the high density of PSs, the east‒west and vertical components were also resolved with high detail, resulting in 1400 MPs for the east‒west and vertical components after the application of a regular grid of 50 m (Fig. 5).

Fig. 5
figure 5

PS data for the right-bank (RB) and left-bank (LB) landslides for vertical and east‒west components. Points selected and identified as TS are related to the time series of Fig. 5, while the cross-sections are shown in Fig. 6

Full size image

For the left-bank landslide (Fig. 4), ca. 1600 PSs are available for the ascending geometry (with a PS density of approximately 1000 PS/km2), and 1550 PSs are available for the descending geometry (with a PS density of approximately 950 PS/km2). Additionally, in this case, the east‒west and vertical components were resolved with high confidence, resulting in 450 MPs after resampling on a regular grid of 50 m (Fig. 5).

A simple visual analysis performed on the basis of four velocity components offers a clear indication of the spatial pattern of deformation exhibited by the right-bank landslide, which in particular seems to assume roto-translational movement, with a predominant vertical pattern in the main scarp and high values of the east‒west component in the body.

In addition to the simple use of mean annual velocities, landslide analysis can frequently take advantage of the information provided by the displacement time series, which represents the most advanced product of any multitemporal InSAR processing, providing the deformation history of the measured point by acquisition, with millimetric precision. Displacement time series are necessary for studying the kinematics of a phenomenon because they show possible seasonal trends, nonlinear movements, ground acceleration, and any change that may have occurred during the monitoring period.

Five MPs have been chosen in the upper, middle, and downside parts of the two landslides to obtain a general overview of the movement (Fig. 5): three MPs in the “right-bank” side, TS1, TS2, and TS3, located in the axial part of the body along the B-B’ section, and two in the left side of the landslide, i.e., TS4 and TS5, located in the eastern part of the body, where the highest movement occurs. The time series covers a period between March 2016 and February 2020, without showing a particular and significant change in the trend, with the movement almost linear for both the vertical and horizontal components (Fig. 6). The vertical components of TS4 and TS5 from the left side of the landslide show a slight variation from the linear trend, highlighting a small increase in the displacement from July 2017 with a high velocity: approximately 150 mm/year for TS4 and 240 mm/year for TS5.

Fig. 6
figure 6

Time series for both the right-bank landslide (RB) (TS1, TS2, and TS3) and left-bank landslide (LB) (TS4 and TS5) in the vertical and horizontal components

Full size image

Two displacement cross-sections (black lines of Fig. 5) of the RB landslide have been shown to analyse how the kinematics change spatially. The sections include one parallel (B-B′) and one perpendicular (A-A′) to the landslide movement direction, and they consider data from March 2016 and February 2020. In the east‒west component, B-B′ starts at the top of the landslide and ends in the lake, showing values of displacement of the scarp area and the body similar to the foot of the landslide (Fig. 7), i.e., values resulting from rotational movement. A homogeneous pattern can also be observed in the A-A′ section, where the movement appears uniform along the whole landslide width. The maximum displacement is approximately 400 mm for both the A-A′ and B-B′ sections. In the vertical component along the B-B′ section, it is possible to observe a movement from the scarp up to the lake banks, which highlights a higher displacement value within the main scarp sector instead of the foot of the landslide; along the perpendicular cross-section (A-A′), there is a generally uniform displacement with an increase in the upper east part of the landslide. The maximum displacement that is obtained is approximately 500 mm for both A-A′ and B-B′. The comparison between the E‒W and vertical components in the B-B′ cross-section clearly shows the roto-translational movement of the landslide, with a higher vertical component and a lower E‒W displacement in the main scarp, and vice versa along the body.

Fig. 7
figure 7

A-A′ (left) and B-B′ (right) displacement cross-sections for both the east‒west (top) and vertical (bottom) components of the right-bank landslide

Full size image

COSI-Corr

The four SPOT-6 and SPOT-7 images from 2015 to 2021 chosen for the analysis of the Lake Sarez area formed three image pairs using two temporally adjacent images (2015–2017, 2017–2019, and 2019–2021); moreover, an image pair encompassing the 2015–2021 interval, covering the entire period, was used to study and derive the ground deformation. The four image pairs were processed at different combinations of initial–final window size values (in pixels), which provided different image measurement qualities (from 8–8 pixels to 128–128 pixels). A small window size, such as the initial–final 8–8 pixels, has low measurement qualities in the image and represents uncertainty in studying the landslide displacement because the low measurement qualities occur randomly; however, by increasing the sizes, until the initial–final window of 128–128 pixels, clearer spatial patterns are created by reducing the background noise, and the points accurately depict the profile and the shape of the landslide. Figure 8 shows the analysis of the SPOT images using the 128–128-pixel combination that identified a movement that is exclusively concentrated in the left-bank landslide. The values of the movements are approximately 8–10 m in 6 years (approximately 1.5 m/year) for the E‒W components, and the movement is prominent between 2015 and 2017. From 2017 to 2021, the landslide shows a small movement, which is especially focused on the right sector of the landslide, with maximum values of approximately 0.5 m/year. In the N‒S component, the movement is high, approximately 12–18 m in 6 years (2–3 m/year), and in this case, most of the movement is between 2015 and 2017. Since 2019, no movement in the landslide was observed. In the N‒S component, both in 2015–2017 and 2015–2021, it is possible to see that most of the movement is located in the right sector of the landslide, characterized by a more intense colour.

Fig. 8
figure 8

Representation of COSI-Corr analysis of left-bank (LB) landslide surrounded in black for the four pairs of images that were analysed: 2015–2017, 2017–2019, 2019–2021, and 2015–2021; they were analysed in the east‒west and north‒south components

Full size image

Moreover, by observing in detail the results provided by the COSI-Corr, a small movement within the landslide mapped on the right side of the lake is visible (highlighted in Fig. 8 with a red circle), especially in the increased initial–final window size of 128–128 using the cumulated 2015–2021 pair; it is also visible in 2015–2017, with velocity values of approximately 6–8 m in 6 years along the east‒west component. In the north‒south component, this movement is very small. This area coincides precisely with a sector within the landslide body where InSAR data are missing.

Discussion

The global coverage, wide-area mapping capabilities, and regular acquisition planning ensured by satellites make it possible to scan wide areas and to identify unstable zones, especially where remoteness, vastness, and climatic conditions make it difficult to perform field activities and target slopes. Because it is a virtually inaccessible area, the precise and spatially dense information delivered by interferometric and optical data can be highly valuable for Lake Sarez, with a particular focus on the right-bank and left-bank landslides located in the western part of the lake.

Table 1 includes a summary of the velocity values recorded by the adopted monitoring technique (InSAR and image correlation) for the two landslides. For InSAR measurements, both VLOS and Vslope are reported, with the latter, as expected, having higher and more representative values due to its correction with the C-index, which takes into account the orientation of the slope and the direction of the satellite.

Table 1 Values of velocities V and Vslope from InSAR data and VCOSI-Corr from optical analysis along the various components: descending, ascending, E‒W, vertical, and N‒S
Full size table

According to Raetzo (2006), right-bank landslides can be divided into four main sectors: the upper, northern, central, and southern parts, which are characterized by different slope processes and geological conditions. The upper part is characterized by more superficial landslides and rock glaciers with gravitational processes affecting the main landslide in the areas below; in this area, it is not easy to identify the instability of the bedrock. The northern part is nonactive, with the last movements probably attributable to before the last glaciation. The topographic conditions are different from those below, with reduced slopes and inclinations. The central part is composed of sandstones and presents few fractures and a lower degree of dislocation. The southern part is the most active and is characterized by a well-defined disintegration niche involving loose material and bedrock, with an inclination of 40–60°, which leads to frequent rockfalls within the lake.

Then, taking into account the four sectors identified by Raetzo et al. (2006), most of the movement is focused in the southern sector for both the VLOS and Vslope of the landslide for the ascending geometry. Along the descending geometry, the high values of VLOS and Vslope are located in the same sectors of the landslide, as shown in Fig. 9.

Fig. 9
figure 9

Representation of the Vslope values for both the RB and LB landslides in descending and ascending geometries

Full size image

For the left-bank landslide, there are no sectors identified in the literature, but both the VLOS and the Vslope present the highest movements in the eastern part for both the ascending and descending geometries, which is also confirmed by the COSI-Corr data.

The values identified by the COSI-Corr method are higher than those of the VLOS and Vslope, which probably depends on the different approaches that the methods use during data processing. For the right-bank landslide, the COSI-Corr velocity values are available only for a small sector, which is not visible through InSAR data, probably because the other sector of the landslide has lower movement, which is not significant with this kind of technique and with the pixel resolution of SPOT images.

In fact, there are some shadow areas that are not visible using InSAR analysis due to a combination of the geomorphology of the area and the acquisition geometry of the satellite and, sometimes, due to very fast movements, which do not allow the interferometric technique to detect a coherent signal because of its inability to analyse such rapid deformations. Moreover, COSI-Corr data enables the detection of the north–south component of the deformation, notoriously a blind sight direction for the right-looking configuration of SAR satellites. This combination of optical and InSAR data, therefore, turned out to be very useful and interesting because it provides a clear and complementary overview of the movement of the two landslides.

The COSI-Corr data mainly show movements on the “left-bank” side, especially during 2015–2017, which is not fully covered by the InSAR data (data were available from 2016 to 2020). Data show higher velocities in the eastern part of the landslide, especially in the north‒south optical component, as shown in Fig. 8. InSAR does not allow the analysis of movement in this direction, thus confirming the suitability of optical imagery, which can easily integrate and complement radar-based approaches. The east‒west component confirms this specific movement using both COSI-Corr and InSAR approaches. The optical analysis, which shows an interesting overview of the left-side landslide, cannot be used to understand the movement in the right slope since no movement is shown. This could be explained because the right-bank landslide may have low velocities of movement that are not applicable through the use of optical data. To confirm this, a small sector on the right-bank landslide is detected exclusively using the COSI-Corr analysis, showing higher velocity values compared to the other parts.

In any case, both the interferometric and optical data agree on the definition of the broader kinematic picture of the right-bank landslide that is not experiencing accelerations due to pre-failure deformation (tertiary creep; Intrieri et al. 2019) but that is affected by an ongoing linear displacement, as highlighted by the time series. The landslide does not show specific accelerations linked to seasonal processes or the presence of snow, but it is homogeneous in each part of the landslide. For the left-side landslide, there is a small increase in the displacement that is visible in the time series around June 2017–January 2018 in the vertical component; this increase is not visible through the COSI-Corr technique, probably because the movement is small and it is not significant from an optical point of view due to the limited number of images available for the area (and the consequent large time span between each of them).

By studying the data collected from the temporal series, it is likely to conclude that the roto-translational movement of the RB landslide is likely characteristic of the slope and represents its own movement and it is not connected to the earthquakes that happened in the region. The satellite precipitation data obtained from ERA5-LandFootnote 1 have been analysed for the entire area of interest and cover the entire period analysed by the InSAR and optical methods, from 2015 to 2021. They confirm that there is no seasonal movement because the area is characterized by infrequent and not intense precipitation, the monthly average precipitation from 2015 to 2021 is around 40 mm and the highest monthly precipitation is around 120 mm in spring. So, the slope movement could not be associated with rains, but the presence of snow for most of the year can melt and permeate the ground, creating movement and having an important role and impact on the slope. For the left-bank landslide, COSI-Corr data highlight most of the movement in approximately 2015–2017, which could be associated with the earthquake that occurred in the area in December 2015.

InSAR data have also been used to postulate the geometry and depth of the sliding surface of both right-bank and left-bank landslides using a method originally developed by Carter and Bentley (1985), improved by Cruden (1986), and validated with the use of satellite interferometric data by Intrieri et al. (2020), who also dubbed it the vector inclination method (VIM). This method assumes that the direction of the superficial ground reflects the geometry of the sliding surface, which is generally true in the case of landslides with no strong internal deformations along the vertical axis.

Notably, the VIM is affected only by the direction of the superficial movement and not by the amount of displacement, which means that variations in the modulus of the displacement vector with depth are allowed as long as the direction is relatively unchanged. Through a geometric process, the whole sliding surface can be reconstructed along a cross-section to provide a kinematically possible solution. The method is based on three assumptions: a point on a landslide surface will move in a parallel slope direction of the sliding surface beneath; the mass will move as a rigid body; there is only one sliding surface. The procedure consists in drawing the cross-section of the landslide which intersects the measurement points. Starting from the back scarp, the normal lines to two consecutive movement vectors are obtained, and their intersection becomes the rotation centre of the sliding surface passing through the first vector and ending at the second one. This process is then repeated for each movement vector. Even though the surface is obtained as a series of circular sections, the results may well represent planar shapes as well, as a result of circular sectors with a long curvature radius. The application to InSAR displacement data, which are notoriously referred to as the sensor’s line of sight, implies that the vectors must be decomposed into the vertical and horizontal components and that the real direction of movement on the horizontal plane must be assumed to be generally parallel to the slope direction. The major issues in the applications of this method arise due to the noise of interferometric measurements (especially for slow-moving and N‒S-oriented landslides) and when the MPs are not well distributed along the cross-section or are not dense enough, which is not the case for the right-bank slide. While the VIM works best with a calibration derived from independent information on the landslide’s thickness (such as an inclinometric measurement), in a case such as this, which is characterized by extreme difficulty in carrying out field investigations and monitoring, this method probably represents one of the most practical ways to form a data-based hypothesis on the geometry and depth of the sliding surface.

The results of the VIM on the right-bank slide along the C-C′ cross-section (Figs. 10 and 11) show the presence of a slightly compound surface, with a major translational mechanism and a depth ranging from 70 to 140 m in the central part of the landslide along the middle longitudinal axis, corresponding to 110 m of equivalent thickness (i.e., the value that provides the same cross-section surface if multiplied by the length of the sliding surface). Since the satellite data can also provide information on the area of the landslide (5.34·107 m2), the emerged volume of the landslide can be estimated to be approximately 1.4·109 m3, thus narrowing down the 0.3·109–2·109 m3 range from the state-of-the-art technique. Interestingly, the displacement data at the foot of the landslide suggest its continuation below the lake level. While it is impossible to make precise assessments with no knowledge of the topography or the interferometric data below Lake Sarez, it can be conservatively assumed that the sliding surface ends with a circular shape (as suggested by some uplifting MPs on the right side of the toe) and that the topographic profile, which is slightly concave, can either continue by keeping the same average slope or with the slope of its most terminal part (see the two black dashed lines at the bottom of the slope in Fig. 10). In both cases, the landslide, which already visibly narrows the lake by some 500 m, appears to extend for another 275–325 m below the water level, whereas the lake in that location has a width ranging from 1 to 2 km. The total volume of the landslide considering this submerged part would then rise to approximately 1.5·109 m3 (Fig. 12).

Fig. 10
figure 10

Landslide outline at surface level and submerged cross-section C-C′ for the RB landslide and cross-section D-D′ for the LB landslide

Full size image
Fig. 11
figure 11

Reconstruction of the sliding surface (red dotted line) of the right-bank slide using the VIM along the C-C′ cross-section. The two black dashed lines represent possible interpolations of the topography below the lake level (in cyan). The red arrows indicate the displacement vectors of the MPs projected along the cross-section

Full size image
Fig. 12
figure 12

Reconstruction of the sliding surface (red dotted line) of the left-bank slide using the VIM along the D-D′ cross-section. The red arrows indicate the displacement vectors of the MPs projected along the cross-section

Full size image

Similarly, the VIM has been applied to the left-bank landslide, which exhibits a nearly constant dip marking a translational sliding surface (Figs. 10 and  12). Since this surface is steeper than the topography, the landslide depth increases downhill, up to 125 m (69 m of equivalent thickness). There is no trace of more horizontal or slower displacement at the base of the slope, which suggests that the landslide continues underwater, although the absence of any changes in the inclination of the movement vectors prevents us from making data-based hypotheses on the submerged part, thus making any estimation of the total volume an undefined underestimation, while the volume of the emerged part is estimated at approximately 2.1·108 m3.

Provided information on the extent, volume, and rates of displacement of both landslides is highly valuable, as it allows one to make a complete geometric and kinematic characterization of the most relevant slope instability affecting Lake Sarez and whose failure may cause a surge wave in the lake (Strom 2010, 2014 and reference therein). Following the definition provided by Hungr (1997) of the landslide intensity as “a set of spatially distributed parameters determining the destructiveness of a landslide,” the practical assessment of intensity is often quite difficult since it is highly site-dependent (Corominas et al. 2014), and many parameters, both geometric and kinematic, should be accounted for, depending on the landslide type and propagation mechanism. For slow-moving landslides (large slides and rockslides, such as those affecting the Lake Sarez area), intensity is generally expressed in terms of the displacement rate (Mansour et al. 2011; Frattini et al. 2013) and geometric parameters (Guzzetti et al. 2005). Defining the landslide intensity is therefore crucial for a proper evaluation of vulnerability, for quantitative risk analysis, and for the assessment of risk scenarios. However, these outcomes are not sufficient to establish an early warning procedure and for the kinematic and mechanical modelling of the landslide, which require additional information and data; nonetheless, the satellite-based data can support the abovementioned practices, providing insightful data to define state of activity, deformation rates, and geometry of the deformed area.

Conclusions

Tajikistan is a mountainous and largely inaccessible area with complex topography, but it hosts several sites that have drawn the attention of geoscientists interested in landslide analysis. One such case is represented by the slope sectors surrounding the famous Lake Sarez, formed by a landslide dam event, which has been analysed in this paper. In particular, two large landslides, one on the right bank and the other on the left bank of the lake, have been the target of this study, as they could generate disastrous consequences in case of failure within the lake. The collapse of the landslides could indeed generate an anomalous wave overtopping it, endangering the people living along the lake and the river downstream. Despite this serious threat, knowledge about the kinematics and size of these landslides is still scarce due to logistic constraints that make surveying difficult.

Therefore, the objective of this work was to use two different Earth observation methods to obtain a complete overview of the area and advance the state-of-the-art techniques of the site: the InSAR technique exploited with the SqueeSAR approach using Sentinel-1 data and the optical imagery approach with the support of ENVI and COSI-Corr software using SPOT-6 and SPOT-7 images. InSAR data showed movements in both landslides, with a maximum velocity of 447 mm/year for the right-bank side landslide and up to 771 mm/year for the left-side landslide. For the right-bank side landslide, there are few data obtained from the COSI-Corr analysis, probably due to the low velocities of the movement. On the other hand, the optical technique provided interesting results for the left-side landslide, especially in the 2015–2017 period, which is partially covered by InSAR data. The velocities obtained with COSI-Corr are higher than the velocities obtained by the InSAR data due to the different approaches used and the different measured components. Although only four SPOT images of the area were obtained, the data provided a clear and precise overview of the movement of the landslide, whose displacement was also confirmed by the InSAR results. For the right-bank side landslides, the geometry and depth of the sliding surface were hypothesized based on InSAR data. This latter aspect may represent a starting point for modelling the volume and failure mechanism of the landslide from a risk assessment perspective.

This approach holds promise for broader applications in different case studies due to the complementarity of two different methods, such as COSI-Corr and InSAR, based on the synergistic exploitation of radar and optical data. Their combination proves valuable improvements in monitoring and studying a wide spectrum of landslide velocities, especially where installation of ground sensors is difficult or unfeasible. Moreover, for what concern landslide dam-related scenarios, given the pressing need for precise studies and analyses about their kinematics geometry, findings of this paper offer a powerful tool for landslide investigation, allowing the reconstruction of the sliding surface and a preliminary assessment of landslide volume. The correlation between these two techniques not only offers a current solution but also lays the groundwork for future assessments and evaluations of landslide dam evolution, especially in inaccessible or impervious areas.