Landslide detection, monitoring and prediction with remote-sensing techniques

Landslides are widespread occurrences that can become catastrophic when they occur near settlements and infrastructure. Detection, monitoring and prediction are fundamental to managing landslide risks and often rely on remote-sensing techniques (RSTs) that include the observation of Earth from space, laser scanning and ground-based interferometry. In this Technical Review, we describe the use of RSTs in landslide analysis and management. Satellite RSTs are used to detect and measure landslide displacement, providing a synoptic view over various spatiotemporal scales. Ground-based sensors (including ground-based interferometric radar, Doppler radar and lidar) monitor smaller areas, but combine accuracy, high acquisition frequency and configuration flexibility, and are therefore increasingly used in real-time monitoring and early warning of landslides. Each RST has advantages and limitations that depend on the application (detection, monitoring or prediction), the size of the area of concern, the type of landslide, deformation pattern and risks posed by landslide. The integration of various technologies is, therefore, often best. More effective landslide risk management requires greater leveraging of big data, more strategic use of monitoring resources and better communication with residents of landslide-prone areas. Detection, monitoring and prediction are essential to managing landslide risks. This Technical Review examines the use of remote-sensing technology in tracking landslides and mitigating disaster.


Technical review
we present future directions for research and development to best improve their usage in landslide risk management.

Space-borne techniques
Since the launch of Landsat-1 in 1972, RSTs were recognized as a promising tool for landslide study 6,19 as they provide systematic, large extent information of different parameters of the land surface, enabling investigation of active disruptive processes at different spatial and temporal scales 7,8 . Currently, the study of landslides with various geometries, styles and kinematics is enabled by the availability of sensors operating at different wavelengths and frequencies, which offer different spatial, temporal and spectral resolutions 20 . Modern landslide analysis also benefits from imagery collection within the multispectral 21,22 and microwave 10,23 domains of the electromagnetic spectrum. For example, both very-high-resolution (VHR) multispectral and synthetic aperture radar (SAR) satellite acquisitions have been exploited to compile landslide databases 24 and to measure and quantify ground deformation induced by the occurrence of landslides 25 . Characteristics of spaceborne multispectral and radar sensors and basic processing principles are described in this section.

Multispectral satellite sensors
Satellite images acquired in the visible, near-infrared and shortwave infrared spectrum have been largely used for producing landslide inventory maps, often with a regional scale of analysis 26 . In most cases, multispectral images are used for mapping terrain features indicative of the occurrence of landslides, such as surface landforms, geological structure, land-use practices, distribution of vegetation 24 and temporal changes of these factors 27 .
Multispectral satellite platforms (such as Landsat, ASTER, SPOT and Sentinel-2) host passive instrument payloads capable of detecting and measuring reflected or emitted electromagnetic radiation from the surface of Earth 28 (Fig. 1a). They observe Earth perpendicularly to the orbital track following sun-synchronous trajectories, to ensure that the angle of sunlight on the surface of Earth is consistently maintained. Acquisition time, selected to minimize the potential impact of shadows and to guarantee a suitable level of illumination, is usually in the morning (the mean local solar time at the descending node is 10:30 for Sentinel-2). The acquisition time of the day is quite close for most of the missions (Landsat, SPOT and WorldView-3); to make it so different, archives and existing and historical missions are comparable and therefore can be integrated. This feature is important, for example, when trying to reconstruct the past activity of a landslide as it minimizes the needs to apply corrections for different shadow angles and light conditions.
There are several considerations in selecting satellite data for use in landslide research. For instance, the revisiting time is highly variable between satellites, ranging from a daily basis (WorldView and Pléiades constellations) to a few days (QuickBird-2 and GeoEye-1) to weeks (16 days by the Landsat family and 26 days by . A substantial improvement in temporal observation frequency is provided by the adoption of virtual constellation of mid-resolution multispectral satellites 29 . Jointly, Landsat-8 and Landsat-9 (a repeat cycle of 8 days) and Sentinel-2a and Sentinel-2b (a repeat cycle of 5 days) provide a global

Introduction
Landslides are ubiquitous in sloping environments and can be driven by tectonic 1 , climatic 2 or human 3 activities (Box 1). Landslide disasters arise when the hazardous movement of soil and rocks directly or indirectly impacts vulnerable human settlements and infrastructure, sometimes causing widespread damage. Between 2004 and 2016 alone, an estimated 55,997 people were killed worldwide in 4,862 distinct landslide events (not counting those caused by earthquake), with Asia being the most affected continent 4 .
Landslide risks could be exacerbated by global changes such as increases in the frequency or magnitude of rainfall and permafrost degradation 5 . As such, international organizations and agencies promote landslide risk reduction policies that include understanding risks and enhancing disaster preparedness for effective response. The 2020 Landslide Kyoto Commitment, for example, aims to provide the tools, information, platforms, technical expertise and incentives to promote landslide risk reduction on a global scale.
To achieve these goals, it is vital to detect the presence, monitor the velocity and predict the stability conditions of unstable slopes and the evolution of these features with time. However, the complex and widespread nature of landslides necessitates accurate measurements, often with a spatial coverage capable of targeting one or more slopes and an acquisition frequency high enough to catch substantial changes in the landslide (depending on the type of landslide; Box 1). In situ monitoring techniques are generally not suitable for widearea monitoring, as each single sensor provides information only concerning its surroundings, and can require the operators to enter hazardous or inaccessible areas. Remote-sensing techniques (RSTs) have therefore become an indispensable tool in landslide investigation 6 , as they offer a systematic, synoptic and cost-effective view of the ground surface at various scales. The use of RSTs has enabled a greater understanding of the complex interaction between different geological and geomorphological phenomena underlying disruptive processes 7,8 .
Broadly, RST includes tools that do not require human contact with the target and can be space-borne or ground-based, both of which are used in landslide science. For example, satellite RSTs are extensively used in landslide monitoring, with their use growing since the 2000s owing to continuous technological advancements, such as the development of algorithms to substantially improve the quality of monitoring data 9 and the launch of satellites with increasingly higher acquisition frequency and spatial resolution 10,11 . Ground-based laser scanning is widely used for slope instability detection, mapping and monitoring [12][13][14] . Some techniques can be applied from either platform. For example, since the 1990s, satellite-based sensors have used interferometry, which can measure how much a target is moving with sub-millimetre accuracy. Since the 2000s, interferometry has also been used applying ground-based apparatuses, allowing for more tailored applications. Since the advent of RSTs, the capability to rapidly map landslides over wide areas 15 , to collect information that provides mechanistic and kinematic insight 16,17 and to gather big data for issuing early warnings 18 has considerably increased.
In this Technical Review, the potential and application of RSTs for landslide detection, monitoring and time prediction are discussed. We mainly focus on multispectral and interferometric satellites, ground-based interferometry, laser scanning and Doppler radar, which have been increasingly used among the RSTs developed since the early 2000s. RST function and development in landslide science are described, followed by an overview of their application, advantages and limitations in detection, monitoring and forecasting. Finally, Technical review on the particular spectral band, generally increasing from the visible and the near infrared to the shortwave infrared. The launch of VHR satellite systems, which usually cover the visible bands and include panchromatic domain (blending the information of blue, green and red bands), led to the acquisition of images with 0.25 m up to 1 m ground sampling distance. For VHR satellites (for example, GeoEye-1 and GeoEye-2, QuickBird-2, Cartosat, WorldView-3, Pléiades-1 and Pléiades-2 and IKONOS), panchromatic imaging capability at nadir is about four times finer than that of the multispectral band 31 . There are many pansharpening algorithms 32 for merging low spatial resolution multispectral images with higher spatial resolution panchromatic bands to generate a single high-resolution colour image. However, data fusion of different spatial and spectral resolution can impact on the characteristics of the image useful for landslide detection and mapping, such as linear features, textures, contrast and colour 33 .
The current availability of different satellite platforms increases the possibility to timely acquire cloud-free multispectral imagery following a major landslide event or during a period of high deformation, when monitoring landslide activity and predicting its possible failure must be rapidly carried out. Nevertheless, weather conditions (especially clouds and snow cover) and daylight can still hamper the acquisition of suitable images, depending on the climate zone and the season. In general, mountain areas and areas with humid climates and rainy seasons are potentially the most unfavourable regions and times for optimal images. Unfortunately, these are often the times and places when landslides are more likely to occur, representing a major limitation to this technology.

SAR satellite sensors
Unlike passive remote sensing, active sensors such as SAR instruments 34 provide their own energy source for illumination. As a result, they are able to obtain measurements anytime, regardless of the time of day or season, using bands of the microwave domain 35 . This feature is critical for continuous landslide monitoring and prediction.
A SAR image is composed of pixels corresponding to a ground area characterized by amplitude and phase values. Although amplitude depends on the reflectance of the illuminated objects, phase value of a single SAR image partly depends on the sensor-target distance -when a point on the ground moves, the distance between the sensor and the point changes and the phase value recorded by the sensor is affected accordingly. This feature is the key element in interferometric applications, a wide term referring to the exploitation of the SAR signals of at least two SAR images 36 . The spatial resolution of SAR images depends on the sensor used and its acquisition mode. The temporal resolution depends on the revisiting time of the satellite and ranges from several weeks (Radarsat constellation) to a few days (Sentinel-1 constellation). A suitable approach to exploit phase variation between two consecutive radar images acquired over the same target area is the differential interferometric SAR (DInSAR). After topography compensation, DInSAR enables the measurement of surface motions from several dozens of centimetres to 1 cm or less (depending on the wavelength) with suitable resolution over areas of tens or hundreds of squared kilometres 37 .
Atmospheric phase distortions 38 and geometrical and temporal decorrelation 39 reduce (or even compromise) the quality and reliability of DInSAR results. To overcome these limitations, interferometric SAR (InSAR)-based information can be enhanced through multitemporal interferometric SAR (MTInSAR) approaches, based on the analysis of long stacks of coregistered SAR imagery 9,40-43 . MTInSAR is designed to identify a grid of measurement points (MPs) corresponding to a single pixel or a group of few pixels exhibiting a stable radar signature over the entire observation period (Fig. 1b). MPs usually correspond to man-made objects, rocky outcrops and bare soil, which are characterized by high reflectivity and stability (movements not higher than few tens of centimetres per year). For each MP, it is possible to estimate displacement time series along the satellite line of sight (LOS) and a set of quality parameters. LOS deformation rate can be estimated with

Landslide classification
Landslide is an all-inclusive term referring to downward and outward movement of slope-forming materials under the influence of gravity 157 , which can ubiquitously act on both natural, and engineered and excavated slopes 158 . Landslides can occur in slopes on the verge of instability and can be triggered by natural processes (rainfall, snowmelt, erosion by stream, volcanic activities and earthquake shaking), human disturbances (agriculture, construction, mining, cuts or variation of reservoir levels) or any combination of these factors.
Landslides can be initiated by both intense, short-duration rainstorm events (hours to days) and long-lasting precipitation (weeks to months), leading to longer timescale seasonal fluctuations 159 . The first events commonly trigger shallow slides, whereas the second episodes are normally associated with deeper and larger slope movements (5-20-m depth). Deep-seated and shallow landslides intrinsically differ in their volume, extent, hydrological triggering 160 , occurrence 161 and the risk posed.
One of the most widely used classifications is based on the type of movement (falls, topples, slides, spreads, flows and complex) and the type of material involved (rock, debris and earth) 162 , with a velocity scale (later updated 86 ) completing the scheme (see the table). An updated classification proposes a matrix with 32 landslide types 163 . However, classification schemes are not perfect, as the type of movement and material varies in space and time, and an almost continuous gradation of types exists.   remote-sensing products. a, QuickBird imagery in the multispectral very-high-resolution analysis of the Giampilieri area, Italy, which had more than 600 landslides following a storm. The image is represented in false colours, in which the red areas are those with a high reflectivity in the near-infrared spectrum (that is, vegetated areas), the cyan areas are bare soil or cities and the yellow areas are the pre-event and post-event differences, indicating landslides (from ref. 11 ). b, Deformation map of the Val d'Aosta region, Italy, elaborated using multitemporal interferometric synthetic aperture radar Sentinel-1 data acquired along the ascending orbit of the satellite. c, Displacement map of the landslide of Pomarico, Italy, from a ground-based interferometer and projected over a multispectral image. Yellow pixels were stable over 30 months; the purple shows movement towards the sensor owing to a landslide. d, Product of a Doppler radar applied on a rock avalanche threatening a street in Ruinon, Italy, on 12 September 2020. The dashed yellow area is the Doppler radar field of view. The colour bar indicates the local intensity of the detected rockfall using an arbitrary scale that is a proxy of the momentum of the landslide. The full green line denoted as ROI (region of interest) includes the area of potential movement. If the collapsing material enters the area called DZ (danger zone), delimited by the full yellow line, then a warning is issued. e, As in d, but for 1 October 2020. f, Comparison of two lidar scans of the Marano landslide, an earth flow in Italy, that were acquired 5 days apart. The cool colours represent earth advance and the warm colours denote a retreat. Part a is reprinted from ref. 11 , CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/).

Technical review
Analysis of amplitude information is an effective alternative to InSAR for mapping surface deformation in the case of temporal decorrelation 45 or loss of coherence 46 . Amplitude-based methods exploit the correlation of the speckle pattern of two images (pre-landslide and postlandslide failure, for instance). If the deformation introduces geometric distortions without substantially affecting the SAR image reflectivity, by tracking amplitude features in multitemporal SAR data, movement rates (from tens of centimetres to tens of metres) exceeding those observables with InSAR can be recorded. Despite a lower precision (of the order of 1/50th to 1/20th of the image pixel) 47 , speckle tracking offers all the three geometrical components (x, y and z) of the displacement vector, having images acquired along ascending and descending orbits.
Analysis of phase and amplitude collected by radar sensors is a powerful tool for landslide investigation, especially over large areas, allowing for the retrieval of information about distinct displacement regimes, ranging from small to high deformation rates. A general drawback of space-borne applications is that they are not customizable; acquisition parameters such as acquisition frequency are fixed and typically of the order of a few days, and shadows -affecting steep slopes, narrow valley and certain orientations in particular -reduce the amount of ground surface reflecting the signal. For example, for spaceborne InSAR, north-south facing slopes are unfavourable as supposed movements along such slopes would be roughly perpendicular to the LOS and therefore undetectable. As a result, specific locations might not be optimally studied with satellite platforms and could require a more tailored approach using ground-based sensors.

Ground-based techniques
Although satellites scan wide areas and spot unstable zones, terrestrial remote-sensing technologies monitor single slope instabilities for short-to-long-term landslide management. By exploiting radar, laser and multispectral signals, ground-based technologies can be versatile, accurate and precise, with high spatial and temporal resolution. For example, ground-based InSAR (GBInSAR) can take less than 1 min (up to 10-20 s for some apparatuses). This (at least) tenfold improvement means that landslides up to the boundary between moderate and rapid velocities can be monitored (Box 1). This section describes ground-based interferometry, Doppler radar and lidar in landslide monitoring.

Ground-based interferometry
Ground-based interferometry was born in mid-1990s following the observation that the technique developed on satellite platforms would have been useful to measure slope deformations 48 and man-made structures such as dams 49 . Ground-based apparatuses are developed to use both synthetic (such as GBInSAR) and real aperture radar. In the latter case, dish antennas are used to make acquisitions over small (a few squared metres) areas and are then rotated and pointed at the adjacent resolution area until the whole scenario of interest is scanned; therefore, the acquisition time is a function of the extension of the monitored area. The acquisition time for GBInSAR is related to the length of the synthetic aperture, which is usually obtained by moving back and forth the antennas along a linear rail (generally 2-3 m long). Both systems are capable of near real-time slope-scale monitoring, with potential for early warning, as they produce nearly spatially continuous displacement data (displacement maps) (Fig. 1c), with deformation rates ranging from a few millimetres per year up to several metres per day. Similarly, they both can achieve sub-millimetre precision and a spatial resolution in the order of 1-10 m. They also perform an acquisition in less than 2 min (at least considering a 30-45° horizontal field of view), operate fully remotely and ensure 24-h operativity in nearly all-weather conditions (although strong wind, rainfall and high air temperature and humidity gradients can generate noise). These features are key to providing landslide warnings.
The flourishing of various applications and the continuous technological evolution of such systems demonstrate that ground-based interferometry is a cutting-edge approach for landslide monitoring and prediction. Uses have been successfully trialled dating back to the early 2000s, including monitoring and early warning of landslides threatening public safety 16,50 , unstable slopes in open-pit mines 51,52 , the flanks of volcanic edifices 53 , the residual slope risk after collapse 54 and assessing the effectiveness of stabilization works 17 .

Doppler radar
Some types of rapid landslides, such as rockfalls and debris flows, often lack notable precursors. Therefore, risk mitigation strategies are usually focused on active or passive defence structural works and can use tools to detect the ongoing collapse; one such tool is a sophisticated radar system that exploits the Doppler effect principle. These devices enable effective real-time monitoring, so that alerts are issued during the event itself. As the moving mass of earth is often situated along a slope at a higher altitude than the elements exposed to risk, this real-time monitoring can provide a lead time of some tens of seconds, which could be sufficient to automatically turn off a traffic light and close a street or trigger a siren to clear the area and allow people to reach a safe location.
The first successful attempts to combine rapid mass movement detection and road closure with medium-range Doppler radar techniques were made on snow avalanches 55 . After that, a long range, wideangle, radar-based avalanche detection system that is integrated into a fully operational warning system has been developed to detect avalanches and to close threatened roads without human intervention 56 . Since the late 2010s, the Doppler radar technique has been successfully adapted and implemented on debris flow 57,58 and rockfall monitoring and early warning systems [59][60][61] .
The operating principle 62 consists in the emission of an electromagnetic signal in the microwave domain by a frequency-modulated continuous wave radar sensor. The signal reflected by the surrounding scenario is picked up by multiple receiving antennas and analysed to determine the occurrence of rapidly moving objects along the observed slope. If the object dimension and velocity are large enough, a deviation in the signal frequency, or Doppler effect, is induced and detected. Owing to its ability to provide 2D maps of rapid events, including the measurement of their velocity, duration and trajectories, Doppler radar represents an operative tool to detect and track rapid landslides (faster than 1 m s −1 ) and to issue alarms.
Detection warnings can be customized according to appropriate spatial criteria and need to be filtered using specific algorithms to discriminate true events from false alarms (such as those owing to atmospheric disturbances or to tree top oscillation). The system is, therefore, able to measure in real time the landslide speed and to precisely geolocalize its propagation trajectory (Fig. 1d,e). It is possible to detect the presence of bodies larger than some cubic metres that move faster than a few metres per second at about 1-2 km from the sensor. Several simultaneously occurring events can also be tracked, provided that they are at least 10 m far from each other. Alarm notifications (SMS and e-mail) are sent by the control unit associated with the radar sensor; the system can also be connected to safety devices such as traffic lights, sirens or self-closing bars. This solution can be particularly convenient in the case of roads at the base of rocky walls or detrital slopes affected by widespread and repeated instability phenomena.

Lidar
Lidar techniques are used for monitoring unstable slopes as they provide detailed digital models of the observed surface over which other spatial information can be projected (such as interferometric displacement maps), information on the volume of an event and, if multiple scans taken at different times are compared, a displacement field can be detected and monitored (Fig. 1f). They enable 3D landslide analysis by using a narrow laser beam to scan a certain object according to a pre-defined scanning pattern. The direct output of the laser scan is a cloud of points, the positions of which are expressed in polar coordinates referred to the centre of the device. For a specific angular resolution, the resulting point density is therefore a function of the laser shooting distance. Point clouds can be analysed to generate the 3D model of the landslide. Instruments used for ground surveying and monitoring are mainly based on the time-of-flight measurement for distance calculations and have maximum ranges from a few hundred metres up to a few kilometres 13 .
Lidar systems can be used from a fixed position on the ground (terrestrial lidar), or they can be mounted on mobile platforms, such as aeroplanes (aerial lidar), drones, cars and boats (mobile lidar). As the distances are referred to a fixed reference system that is centred on the instrument, terrestrial lidar does not require positioning correction systems. This feature reduces the elaboration time and prevents the need for a global navigation satellite system to record coordinates. Most commercial lidar sensors record from one to up four echoes of the reflected pulse, which makes it possible to distinguish different return pulses deriving from objects placed at different ranges (vegetation and ground) for the same emitted pulse. In addition, full waveform lidar systems are able to digitize the complete return waveform 63 , giving access to an almost unlimited number of returns per shot and high multiple-target resolution, thus increasing details on canopy, sub-canopy structures, ground vegetation and ground morphology.
The spatial resolution of aerial lidar is a function of the altitude and flight speed, ranging from a few centimetres for altitudes of about 250 m to a few decimetres for flight altitudes higher than 1,000 m. Data accuracy depends on the accuracy of the devices used for movement compensation. Thus, aerial lidar is particularly suitable when a large area needs to be covered, or if the investigated slope is too far to be observed from a terrestrial position. For an optimal panoramic viewpoint, terrestrial lidar provides higher accuracy and resolution, especially on steep and overhanging slopes.

Applications
Both space-borne and ground-based RSTs are key in the three main steps required in the instrumental management of landslides hazard (detection, monitoring and prediction). The RSTs described in the previous section are evaluated here under the lenses of these applications. There is not a single technique outperforming the others in all the fields, and performances depend on technical parameters, such as spatial resolution, extension of the investigated area, operating distance, acquisition frequency, accuracy and detectable velocity, that vary with each sensor (Fig. 2). The choice of the most appropriate instrument (or instruments) also depends on the type of application in relation to the type of landslide, as any landslide type is generally associated with a characteristic range of possible volumes, velocities and other features such as type of trigger, geomechanical properties and topography, which make some tools more suitable than others (Table 1).

Detection
Landslide detection consists in the collection of data and information on the presence, distribution and intensity of landslides and the identification of elements affected by the deformative process. Multispectral imagery is mostly used in the detection of already occurred landslides, which can be apparent on the ground but still needs to be identified, mapped and classified, for example, to provide support to postemergency activities such as damage assessment and reconstruction. The use of multispectral imagery has benefitted from increasing computing capacity, improvement in analysis and visualization software and the flourishing of cloud platforms for processing. The increased availability of HR and VHR sensors has also had an important role in overcoming the low spatial resolution of many existing sensors 64 . Classification models applied to VHR panchromatic and HR multispectral satellite images have proven to be valuable for detection and mapping rapid rainfall-induced landslides over large areas 65 . Automatic methods for landslide mapping can be grouped into two main classes: whether they perform a pixel-by-pixel analysis to detect colour or luminosity variations between pre-event and post-event images (pixel-based approach), or whether the basic unit of this analysis is conducted on a group of pixels defining meaningful shapes (object-based approach).
When detection is not made on slowly moving landslides (not already collapsed areas), it is often carried out by measuring the ground displacement with InSAR techniques. For example, satellite DInSAR has been widely used since the 1990s 66,67 to measure the spatial extent and the magnitude of surface deformation associated with mass movements [68][69][70][71] . Satellite-based InSAR can be used to detect and map landslides in remote areas, in mountainous terrain and/or, in general, where deployment of ground-based tools is not logistically feasible and in situ activities are challenging. This aspect is important in mapping landslides without previous knowledge of their location 72 , and particularly for areas with permafrost 73 or seasonally frozen ground 74 , whose susceptibility to geohazards, including landslides, is expected to increase with rapid warming 75,76 . Indeed, since the early 2000s 23 , InSAR has been profitably used to study landslides, becoming a consolidated tool used at different scales -from national 77 to regional 78 , basin 79 , single slope 80 and building 81 -as well as in different phases of landslide response and Civil Protection practice 82 . The ability to make numerous MPs over the landslide body allows for the detection and mapping of the actively deforming slopes 83 , the characterization of landslide mechanism 84 , the zonation of sectors with different velocities and behaviours within the landslide area 85 and the modelling of large slope instability 69 .
MTInSAR, in particular, is highly effective in the analysis of landslide-related events at different stages 26 , although it is restricted to 'coherent' landslides, which have little internal deformation and with extremely slow and very slow movements 86 (Box 1). Over natural, rural and mountainous areas, MTInSAR results are impacted by effects that reduce the intensity and stability of the microwave signal, especially the fast and chaotic movements of the vegetation. In these settings, exploitation of long wavelength (such as L-band) can lead to the generation of more reliable displacement maps 87 and the retrieval of higher spatial density of MPs 88 than short wavelength (such as C-band), owing to its ability to penetrate through a tree canopy surface. However, processing and interpreting SAR data can be challenging 89 in alpine environments, which are characterized by high topographic gradients and persistent snow cover. SAR amplitude-based and pixel-offset techniques 90,91 are a valid method to map the deformation induced by rapid landslides in a vegetated area and where rapid changes impair the InSAR approach 92 .

Technical review
Owing to the wide spatial coverage and cost-effectiveness, satellite applications (both radar and multispectral-based) are suitable in scanning wide areas and spotting localized deformations, especially in hazardous zones in which field activities can be challenging or unfeasible. Existing satellite SAR and multispectral sensors are theoretically appropriate for landslide detection and mapping in a wide range of settings in terms of wavelength, ground resolution and temporal repetitiveness. Rather than relying on a single sensor, the adoption of different spectral intervals (visible, infrared, C-band, X-band and L-band) is encouraged, as it mitigates intrinsic limitations of any one particular sensor. Moreover, the synergic use of different techniques can extend observation capabilities, allowing for the detection and mapping of different deformation regimes; although slow-moving landslides can be analysed using both DInSAR and MTInSAR approaches, rapid landslides can be successfully detected only with interferograms, that is, from the comparison of just two acquisitions. In the case of very fast displacements (the maximum detectable velocity ranges from several centimetres to a few hundreds of centimetres, depending on wavelength and revisiting time of the sensor), reliable information on surface displacement can be obtained using only approaches based on multispectral change detection and amplitude tracking.
Ground-based techniques are also increasingly helpful in landslide detection. At the local scale, high accuracy and flexibility make ground-based techniques extremely useful to retrieve basic information regarding the rate and the extent of deformation of single (or a group of) landslide (or landslides). Ground-based interferometry, for example, can be used for rapid mapping by performing 2D acquisitions to cover areas beyond those that were initially considered of interest and potentially to detect further unstable portions of a slope.
Lidar is widely used in landslide characterization and detection 93 . For example, the visual inspection of high-resolution digital elevation models provided by lidar often greatly improves landslide mapping capabilities 94,95 as interesting features, such as fractures, strata, debris deposits, overhanging rocks and so on, can be recognized on a 3D reconstruction. The high resolution of point clouds acquired by terrestrial lidars allows for a complete geomechanical characterization of rock masses, especially in demanding environments [96][97][98] . Finally, the comparison of two point clouds taken at different times allows for landslide volume estimation and identification of rockfall source areas 99,100 .

Monitoring
Following its identification, it is fundamental to measure the key parameters of a landslide, most notably its surface displacement over time.
Multispectral images are suitable in monitoring landslide motion 101  b Scene extent (km 2 ) c Operative range (m)

Fig. 2 | Technical comparison between different remote-sensing techniques.
Every described technique has a range of operability. a, Spatial resolution. b, Extent of the area that can be surveyed. c, Distance from the target. d, Frequency of the acquisitions. e, Accuracy. f, Velocity of the target. Each technique has its own strong and weak points, which means that none is suitable for every application and type of landslide. InSAR, interferometric synthetic aperture radar.

Technical review
despite obvious limitations related to cloud cover and, until late 1990s, to insufficient spatial resolution 6 . Effectiveness of multispectral imaging monitoring increased with the advent of VHR satellite imagery (IKONOS and QuickBird) and with the launch of the Sentinel-2 mission 102 , which provides regular and frequent acquisitions with global coverage. Image correlation can be used to quantify landslide evolution 103,104 . This method uses moving windows to calculate the correlation between texture similarities in pairs of diachronic images and is the most popular and effective method to derive sub-pixel displacements, for example, using aerial photographs and QuickBird satellite images 105 , Pléiades satellites 106 or Landsat-8 acquisitions 107 on different types of landslide. Slope displacement time series using Sentinel-2 images 108 or the SPOT1 and SPOT5-Pléiades archive 109 have also been obtained. Other than landslide kinematics, successive retrogressive and advancing phases can be successfully reconstructed 110 as well, with short revisiting time satellites.
There are methods to detect and monitor slow-moving landslides (ranging from millimetres per day to few centimetres per day 111 ), but most image correlation methods are only suitable for monitoring landslides moving at metre-to-decametre rates per year. These methods tend to have an accuracy between one-eighth and one-half of a pixel size, depending on the sensor 106 , with Pléiades providing the lowest uncertainties 112 . The availability of large, continuous image datasets reduces the uncertainty, as full data redundancy contributes to creating a more robust time series of displacement 109 . Moreover, for a given reference image, using images acquired during a similar time of the year further reduces displacement uncertainties 108 related to the conditions of acquisition and background noise. Different satellite sources with similar spectral and spatial properties (Landsat-7, Landsat-8 and Sentinel-2) can be combined to create dense time series 113 .
Concerning satellite SAR, displacement time series are the most advanced product of any multitemporal interferometric processing and provide the temporal pattern of deformation of each MP for each acquisition over the observed period. When supported by other instrumentation and field surveys, visual 114 , semi-automatic 115 and automatic analysis 116 of time series can support the characterization and monitoring of landslide kinematics 117 , zonation of sectors 118 , the identification of post-seismic 119 and rainfall-induced 120 velocity changes, seasonal variations 121 and the monitoring of remedial works performance 122 . Furthermore, the shorter than a week revisiting time of satellites such as Sentinel-1, TerraSAR-X or COSMO-SkyMed makes landslide monitoring at the regional scale feasible with a systematic regularity 123 .
Although satellite interferometry is clearly useful, interferometric landslide monitoring requires the presence of suitable targets in the area to be imaged by the radar sensors, such as buildings and artificial objects, which can provide good phase stability for interferometric applications; conversely, the response from bare soil, crops and vegetated areas can be insufficient or even non-existent in non-urban areas. Artificial targets 124 , in the form of active and passive corner reflectors 125 , can be used for both deformation monitoring studies in crops and bare soil areas 126 and accuracy assessment of InSAR measurements 127 .
Although interferometric landslide monitoring works well for slower landslide rates, ground-based radar interferometry has near real-time acquisition frequency and so is able to measure movement rates up to several metres per day 128 . Therefore, this technique is suitable for monitoring most landslides during the phases preceding the possible catastrophic failure, such as earth flows 17 , earth slides 16 , large rockslides 129 and unstable volcanic flanks 130 . Very or extremely rapid movements (Box 1) such as rockfalls are too fast to be measured with radar interferometry, but can be successfully monitored by Doppler radar, which can also be coupled with early warning systems, for example, to close endangered streets (Fig. 3).
Like radar-based monitoring, multitemporal lidar surveys can be used to monitor different types of landslide processes. The introduction of automated terrestrial laser scanners has increased the data acquisition frequency. However, automatic data processing is required, especially for early warning applications 131 , limiting its current application for monitoring.
Overall, satellite methods are suitable for systematic tracking of ground deformation phenomena at the regional scale. Accurate, highfrequency data from sensors installed in situ enable the understanding of specific details of the deformative process and, eventually, the identification of precursory signs leading to landslide failure.

Multispectral
The range of the electromagnetic spectrum from 0.4 μm to about 12.5 μm.

Multispectral band
A specific wavelength range across the electromagnetic spectrum that roughly corresponds to the blue, green and red visible light up to near and shortwave infrared that are captured by multispectral sensors.

Radar
A sensor that emits signals within specific bands of the microwave domain, corresponding to different wavelengths.

Revisiting time
The time interval between two consecutive observations of an area.

Synthetic aperture radar
A technique that moves the antennas of a radar interferometer to obtain data with a spatial resolution equal to that of a sensor with antennas as large as the entire trajectory length.

Prediction
The ultimate goal of landslide monitoring is prediction (or forecast): the ability to timely, accurately and efficiently estimate one (or more) parameters connected to the future failure of a landslide 132 . Failure usually means a catastrophic evolution of the landslide, leading to a strong increase in velocity and runout 132 , often associated with a total or partial disintegration of the material involved and formation of macrocracks. The precise requirements for timeliness, accuracy and efficiency vary from case to case, as they depend on the scale of the forecast (large areas require longer lead time), the type of elements at risk (people, critical infrastructures and cultural heritage), the type of countermeasures needed (town evacuations, street closures and clearance of a sector of an open-pit mine) and the level of preparation of the people responding to the prediction, for example. The prediction of the time of failure (ToF) of a landslide relies on ground displacement data (or the derivative velocity and acceleration), rainfall amount or, less commonly, variables such as acoustic or seismic emissions related to progressive fracturing 133 . Of these three approaches, ground displacement data are the most widely used at the slope scale 134-137 because displacement represents a direct measurement of the stability conditions. However, rainfall is preferred when forecasting at the regional scale, in which diffused and near real-time displacement measurements are typically not available. Predictions based on other parameters, all measurable at the slope scale only, are less common as it is more difficult to relate to the possible collapse.
The general assumption behind ToF forecasting based on displacement is that the final rupture is typically preceded by a power-law acceleration called tertiary creep, which can be modelled to extrapolate the time of collapse. For this reason, almost all the ToF prediction methods only rely on displacement data acquired during the tertiary creep stage [138][139][140] (Box 2).
The advance with which predictions are possible generally depends on the size of the landslide and on the type of the material. Landslides on the scale of millions of cubic metres can display early signs of tertiary creep tens of days before the collapse 141,142 , allowing for a practically feasible prediction 1 week or more before the failure. Stiff and brittle rocks tend to collapse abruptly 143 , limiting warning time, although rapidly measuring instruments such as GBInSAR are now able to provide a few hours of lead time, even for smaller slides of a few tens of thousands of cubic metres 144 . The failure mechanism also seems to affect the predictability of a landslide -masses in which a sliding surface is developing through the propagation of an underground crack tend to respond well to tertiary creep modelling 145 . However, the interference of external factors (such as human activities) and complex geological conditions makes it impossible to be conclusive on a precise relation between failure mechanism and failure prediction.
The ability of an instrument to monitor versus to forecast is determined by whether it can make new acquisitions (and the necessary elaborations) at a frequency compatible with the evolution of the landslide. A high acquisition rate serves to follow every step of the kinematic evolution of a landslide, particularly its accelerations, to model what could come next, especially when a paroxysmal collapse can occur (ToF prediction in strict sense). However, even without a proper prediction model capable of providing a date or an hour of probable collapse, the incoming collapse can become evident by a strong increase (sometimes tenfold or higher) in the displacement velocity, so a high acquisition frequency is important also to issue a warning based on some apparent mobilization or empirical threshold.

Technical review
Nevertheless, a prediction approach based on models -even empirical models -generally requires some data elaborations not compatible with all landslide velocities. For example, the Doppler radar is proficient at early warning against rockfalls and debris flows, as it gives a signal as these very or extremely rapid landslides have already started to catastrophically move down the slope. By contrast, groundbased interferometry is more suitable for ToF prediction; considering the time needed to process and interpret data, a delay of tens of minutes is possible before an alert can be issued. Nevertheless, this lead time can still be considered a very good early warning performance for those landslides that experience slow or moderate velocity during the tertiary creep (such as rockslides, rock avalanches, deep-seated earth slides or particularly large rockfalls; Box 2). Moreover, when the movement rates of these landslides increase shortly before collapsing, a ToF prediction can already been made during the preceding acceleration phase (Fig. 4).
Ground-based techniques are the most widely used in landslide prediction, but the launch of the Sentinel-1 twin satellites 146 has enhanced the use of InSAR in the landslide analysis and presents new possibilities for failure forecast. In particular, these satellites provide a relatively short revisiting time (a 6-day repeat cycle in Europe and some other areas, globally 12 days), regional-scale mapping capability (because of the large 250-km swath) and a systematic and regular acquisition plan. Post-events analyses on a large catastrophic rockslide in China have shown a clear acceleration weeks before the failure, which could have enabled an accurate ToF prediction and to prevent many human losses 142,147 .
Satellite multispectral data, systematically acquired over large areas with short revisiting time, also could be useful for large landslide prediction and warning, at least for some typologies of sliding phenomena. For example, a 9-month time series of displacement for a landslide in the French Alps was observed using the 5-day revisiting time of the Sentinel-2 satellites, capturing a sudden reactivation leading to a failure 109 . Similarly, a large Sentinel-2 archive and a pixel off-track method were exploited to derive possible precursors before a landslide in the southern Himalayas 148 .

Successful landslide monitoring and forecasting
Gallivaggio Sanctuary (near Chiavenna, Italy) is a cultural heritage site that was threatened by an unstable rock mass hanging from a 500-m high, sub-vertical granitic wall. A ground-based interferometric synthetic aperture radar (GBInSAR) was installed in 2016, which detected a 5,000 m 3 rock mass experiencing continuous deformation and threatening both the sanctuary and the nearby road 164 . Traffic and access to the sanctuary were regulated by a sequence of alert thresholds, and restrictive orders were issued according to the measured velocities. On 29 May 2019, the rock mass collapsed, as shown in a frame from a 29 May 2018 footage recording the failure of the landslide (see the figure), and left a 15-m wide scar on the rock mass 164 . The time of failure was accurately predicted several hours in advance, so despite damage to both the road and the buildings, the timely evacuation of the area prevented any loss of life. An early warning system based on displacement velocity thresholds was also implemented in Veslemannen, Norway, which comprised colour-coded hazard levels (green, yellow, orange and red, with red implying evacuation). The monitoring included GBInSAR, a network of extensometers, and time-lapse cameras, observing a 54,000 m 3 active portion of the larger Mannen unstable rock slope 152 . The landslide experienced a gradual acceleration over the years, triggering as many as 16 evacuations from 2014 to 2019, including the last one before the actual slope failure. The last warning was accompanied by displacement velocities high enough to be at the limits of GBInSAR capabilities and eventually, 1 h before the collapse, a complete loss of coherence occurred in the GBInSAR data. Continuous and transparent communication with local population and authorities was crucial to public trust in the warning system, which effectively prevented any human loss.
The figure is adapted from ref. 164

Summary and future perspectives
Remote sensing is critical in the detection, monitoring and prediction of landslides. New and improved technologies now collect an array of information on different types of slope instabilities at multiple scales. In general, owing to new algorithms, and sensors with a higher acquisition frequency and spatial resolution, we are now able to monitor fast landslides that were beyond instrumental capabilities until the late 2010s and to better detect and contour slope instabilities in remote and vegetated areas. The current bottleneck in landslide monitoring and prediction is not only the ability to collect landslide data but also to fully exploit it. For example, satellites continuously acquire and archive images that could detect, monitor and, in some cases, predict people-threatening landslides, but that are often analysed only after catastrophic events 142 . Ground-based interferometry produces displacement maps consisting of thousands of pixels; however, the common practice is to arbitrarily select only a few control points instead of using the full spatial coverage provided by the radar 149 . A partial solution to these data underexploitation could be a more open sharing policy within the scientific community and in a more intense exchange among industry, local administrators and researchers. In general, the more that data circulate, the more likely they will be used.
In addition to better exploiting existing data, the geographical scope of landslide monitoring must expand. Additional research and monitoring are needed in developing areas where population and/or landslide susceptibility are expected to increase and in low-income countries that might have less capability to adopt prevention policies to mitigate landslide impact. The European Ground Motion Service by Copernicus Land Monitoring Service 150 , whose first baseline release was in mid-2022 and is the largest wide-area MTInSAR deformation monitoring system, is an example of broad-scale monitoring. On the basis of Sentinel-1 SAR data processed at full resolution, the European Ground Motion Service provides free seamless information regarding natural and anthropogenic ground motion phenomena over Europe, fundamental for landslide mapping and monitoring, including less investigated areas. In the near future, these ground displacement data will increase the need for algorithms capable of spotting prone-to-failure areas.
The integration of different techniques should be a focus of landslide science in the next 5-10 years. With the launch planned for January 2023, the NISAR mission will be the first dual-frequency fully polarimetric SAR satellite, operating in both L and S (3.2 GHz, ~9.3 cm wavelength) microwave bands 151 . Integration of data from multiple bands will provide deeper insight into the dynamics of landslides, especially in vegetated areas 72 . Displacement maps produced by ground-based interferometers bear with them geometrical distortions that make them difficult to interpret without projecting them over a georeferenced map or a 3D reconstruction of the environment. Today, these projection procedures are standard post-processing ( Fig. 1e) but require time and effort. Therefore, the next generation of ground-based interferometers should include an integrated laser scanner to directly produce 3D images of the scenario displaying the point-by-point displacement; the comparison of multiple laser scans could also enable to correct noisy interferometric measurements and to encompass larger displacement rates without the limitation of the LOS. Similarly, Doppler and ground-based interferometry should be integrated in a single tool capable of covering a large range (around 10 orders of magnitude) of landslide velocities (Fig. 4).
However, in landslide monitoring and management, it is also critical that there is an appropriate match between the cost of instrumenting an area and its actual use. For instance, when an integrated and expensive monitoring network is set into place to deal with a large landslide that has recently shown signs of reactivation, the high cost could reduce the length of time when this system can be kept working owing to long maintenance cost. Using less costly instruments for a long-term monitoring campaign on one or more landslides could ultimately increase the probability that any resulting collapse is monitored. The green and yellow areas show the velocity range within which ground-based interferometry and Doppler radar can operate, respectively. Some types of landslides, such as rockslides and deep-seated earth slides, can experience sudden exponential accelerations (represented by the arrows transitioning from red to blue) just days or hours before failure, thus exhibiting movement rates beyond ground-based interferometric capabilities; however, the trend measured during the pre-paroxysmal phase (in red) can be enough to extrapolate the prediction of the time of failure.

Technical review
To increase the efficacy of these systems, we recommend that forecasting methods based on tertiary creep be incorporated in early warning systems. They currently are not, mostly owing to noise and to the difficulty in discriminating normal acceleration phases from the final acceleration preceding the catastrophic collapse (as exemplified in Box 2 (ref. 152 )). However, moments preceding failure are ultimately clearly distinguishable from other (seasonal) reactivations, as the episode leading to failure usually reaches higher velocity and acceleration values 153 . Formalizing these differences to increase the confidence and the advance necessary to discriminate false alarms from true alarms during emergencies is a much-needed step forward for decision-makers.
Finally, landslide management incorporates technological and social aspects, so risk communication and perception are clearly important. Good communication strategies 154,155 and technologically advanced monitoring and early warning systems are not always paired in practice, so even the most accurate predictions made with advanced RSTs will not be effective if the warnings do not reach the population. Raising risk awareness is a cost-effective means to reduce risk by diminishing the exposure to a hazard with unsafe behaviours and increasing the understandability and efficacy of landslide warnings 156 .