New insights into the temporal prediction of landslides by a terrestrial SAR interferometry monitoring case study
- First Online:
Ten small rock slides (with a volume ranging from 101 to 103 m3) on a slope affected by working activities were detected, located, and timed using pictures collected by an automatic camera during 40 months of continuous monitoring with terrestrial SAR interferometry (TInSAR). These landslides were analyzed in detail by examining their pre-failure time series of displacement inferred from high-sampling frequency (approximately 5 min) TInSAR monitoring. In most of these cases, a typical creep behavior was observed with the displacement starting 370 to 12 h before the collapse. Additionally, an evident acceleration decrease of the displacement a few hours before the failure was observed in some rock/debris slides, thus suggesting the possibility of a mechanical feature of the slope that differs from the classical creep theory. The efficacy of the linear Fukuzono approach for the prediction of time of failure was tested by back-analyzing the ten landslides. Furthermore, a modified Fukuzono approach named average data Fukuzono (ADF) was implemented and applied to our dataset. Such an approach is able to improve forecasting effectiveness by reducing the error due to anomalies in the time series of displacement, like the acceleration decrease before failure. A prediction with a temporal accuracy of at least 2 h was obtained for all the analyzed rock/debris slides.
KeywordsTerrestrial SAR interferometry Landslide forecasting Displacement time series Average data Fukuzono Failure
Landslide prediction is a major step toward reducing the impact of natural disasters and risks related to human activities in mountainous and hilly areas. Landslides are complex natural phenomena involving volumes ranging from some cubic meters up to some hundreds of millions of cubic meters. Landslides do not behave as a perfectly rigid and brittle body; hence, they are always affected by deformation before the failure (Siddle et al. 2007). The trend of deformation vs. time is considered one key factor in allowing the prediction of time of failure. The amount of pre-failure deformation spans several orders of magnitude (from a few millimeters to several meters) depending on the type of material involved in the landslide, the slope geometry, and the landslide trigger. However, landslides are often characterized by a complex geometry and a combination of heterogeneous materials with different features, thus leading to a nonlinear behavior before the failure. Temporal variation of preconditioning and triggering factors such as climate, river activity, tectonics, and human activities further complicate the evolution of landslides over time.
Since the 1960s, several studies have been carried out with the aim to define rules and procedures to estimate the time of failure of landslides (i.e., Saito 1965; Fukuzono 1990; Voight 1988; Crosta and Agliardi 2002). Semi-empirical approaches based on the analysis of displacement (or derived values) time series have been developed and tested on both laboratory-scale and real events, sometimes obtaining good results in terms of landslide forecasting. All these approaches require the collection of displacement data before the failure by a suitable displacement monitoring technique. Back-analysis of large landslides, e.g., the 1963 Vajont landslide (Kilburn and Petley 2003), has demonstrated the efficacy of these approaches in predicting past landslides, but it also highlights the difficulties in predicting the failure time of present landslides (Crosta and Agliardi 2003). All these past studies suffer from the great complexity and infinite variety of landslides from the low quality of displacement data available in terms of temporal resolution in data collection, accuracy, and spatial resolution. In fact, the most recent studies still employ point-based techniques (e.g., inclinometers, extensometers) or monitoring systems with a low temporal frequency in data collection (days to hours) or low accuracy in displacement measurement, etc. In other words, little attention has been given to the importance of monitoring data and to related monitoring methods (Dunnicliff 1988) to improve the forecasting efficacy, even if over the last few years innovative and advanced monitoring techniques have been developed, especially in the remote sensing field (Mazzanti 2012).
Terrestrial SAR interferometry (TInSAR) (Mazzanti 2011; Luzi 2010), also known as ground-based SAR interferometry (GBInSAR), is one of the most recent and most powerful of these techniques for monitoring landslides. This technique has the following main advantages: (1) fully remote monitoring (no installation in dangerous areas is required); (2) widespread monitoring instead of the monitoring of single points; (3) simultaneous monitoring of a high number of points (up to hundreds of thousands); (4) high accuracy in surface displacement monitoring (up to decimal millimeters); and (5) a high data sampling rate (up to a few seconds). These improved capabilities play a fundamental role in the attempt to predict the time of landslide failures, thus allowing, for example, identification of the precise area affected by the movement (widespread view), cross-validation of the displacement time series by the enormous number of monitoring points, and the acquisition of dense and accurate time series.
A field experiment was carried out to evaluate the new opportunities offered by the TInSAR monitoring data for time of failure prediction purposes. Four years of continuous monitoring of an unstable slope by TInSAR with a data sampling rate of 5 min allowed the study of the pre-failure behavior of ten small-scale landslides identified on optical images. The relevant improvements in the forecasting methods for landslide time of failure are also investigated in detail.
Landslide failure prediction by displacement time series
This assumption leads to nonconservative forecasts, but the error is often small and may not be important to decision makers (Voight 1988).
The efficacy of this method in predicting landslide time of failure has been demonstrated by several authors, including Voight and Kennedy (1979), Voight (1989b), Cornelius and Voight (1990, 1994, 1995), Voight and Cornelius (1990, 1991), Rose and Hungr (2007), and Gigli et al. (2011).
Landslide monitoring by terrestrial SAR interferometry
Of these techniques, differential SAR interferometry (DInSAR) is among the newest and most powerful, and its features fit well with landslide monitoring requirements. DInSAR was developed for satellite applications in the early 1990s and was originally used to measure ground displacements at a regional scale (Curlander and McDonough 1991; Massonet and Fiegl 1998; Hanssen 2001). In the late 1990s and early 2000s, several innovation occurred in the SAR field such as the development of innovative approaches for satellite data processing based on data stacking (Ferretti et al. 2001) and the development of the first ground-based SAR equipment prototypes. Over the last 10 years, both satellite and terrestrial InSAR have been extensively used for landslide monitoring, thus demonstrating their efficacy (Pieraccini et al. 2002; Leva et al. 2003; Tarchi et al. 2003; Antonello et al. 2004; Hilley et al. 2004; Strozzi et al. 2005; Noferini et al. 2005; Casagli et al. 2006; Colesanti and Wasowski 2006; Farina et al. 2006; Herrera et al. 2009; Bozzano et al. 2010; Intrieri et al. 2012).
This means that considering a wavelength of approximately 17 mm and a rail length of 2 m (the features of the equipment used in this paper), the cross-range resolution is on the order of 0.4 and 4 m at a distance of 100 and 1,000 m, respectively.
The final result of TInSAR data acquisition is a bidimensional image made of pixels (up to some millions), whose footprint size increases with increasing instrument target distance.
Each pixel of the image is featured by a complex number, namely amplitude and phase values that are specific for a certain polarization, electromagnetic frequency, and incidence angle (Ulaby et al. 1982).
The signal to noise ratio (SNR) value is one of the parameters for assessing the backscattering features of a specific target (i.e., localized area) inside the investigated scenario. Both mean values or variability over time of this value on data stack are used.
Hence, the accuracy in the displacement measurement depends on the signal’s wavelength, the atmospheric conditions, and the sensing distance. Specifically, the lesser the distance and the more stable the atmospheric conditions, the higher the accuracy of the displacement measurements (from 0.01 mm in the lab to a few millimeters in the field).
It is worth noting that the interferometric analysis implies some limitations concerning the displacement measurement due to the cyclic behavior of the phase. As a matter of fact, the phase unwrapping (e.g., Goldstein et al. 1988) is a key part of the interferometric analysis for displacement measurement. Considering the wavelength of the herein used GBSAR equipment, the phase ambiguity is on the order of 4.5 mm (i.e., the λ/4 value) that can be assumed as 9 mm if we assume that inversion of displacement direction along the line of sight cannot occur.
The coherence value, which evaluates the correlation among the phases at each pixel, is a good estimator of the phase stability and, therefore, can be used for a preliminary evaluation of the expected displacement accuracy. This value ranges between 0 and 1, where 0 is the complete decorrelation of the phase and 1 is the complete correlation (Hanssen 2001). Hence, multitemporal analysis allows LOS displacement time histories of each pixel of the image to be obtained (Bozzano et al. 2011).
In summary, the following operational features make the TInSAR technique particularly suitable for the monitoring of landslides and for failure forecasting: (1) the ability to yield data and answers within a brief time (a few minutes); (2) efficacy under any weather and lighting conditions; (3) completely remote operability (it does not require the installation of sensors or targets on the monitored slope); (4) continuous distributed monitoring of the entire slope with a high pixel resolution; and (5) long-range monitoring (up to some kilometers).
Notwithstanding the numerous TInSAR successfully applications cited above, little attention has been devoted to the opportunities offered by this technique for landslide prediction purposes and still less for small-scale landslides.
The study area
Instrumental configuration of the TInSAR instrument
Number of scans in the SAR image
4.5 mrad (3.6 m at a distance of 800 m)
Inter-scan delay (waiting time between the end of one scan and the start of the next)
Measurement time interval
Synoptic table showing the physical features of the investigated landslides
15 days cumulative displacement before landslide failure
24 January 2009; h 14:32–15:32
18 February 2009; h 12:53–13:53
Gneiss colluvium + spritz beton
20 December 2009; h 16:46–17:46
Gneiss colluvium + spritz beton
Night between 15 and 16 January 2010
Gneiss colluvium + spritz beton
03 February 2009; h 16:58–17:58
Gneiss colluvium + spritz beton
Night between 10 and 11 February 2010
Mobilized and altered gneiss
12 February 2010; h 11:58–12:58
Gneiss colluvium + spritz beton
12 February 2010; h 11:58–13:58
Mobilized and altered gneiss
17 February 2010; h 12:57–13:57
Mobilized and altered gneiss
Night between 9 and 10 March 2010
Pre-failure behavior of small-scale landslides detected by continuous TInSAR monitoring
Synoptic table showing the pre-failure behavior of the investigated landslides
Time span deformation (min)
Total displacement (mm)
Maximum velocity (mm/h)
Maximum acceleration (mm/h2)
Delay between the acceleration peak and the collapse (min)
Delay between the last significant rainfall and the acceleration peak (min)
For each landslide, we computed the time series of displacement of those pixels that were considered the most representative of the overall landslide behavior. Specifically, pixels were chosen on the basis of the following criteria: (1) the highest quality TInSAR data available on the basis of SNR and coherence analysis; (2) pixels clearly located within the main landslide body (and not in the surrounding parts affected by tensional release); and (3) pixels located preferably in the topographic upper part of the landslides, as they are assumed to be less influenced by internal deformation.
- 1.The temporal evolution of displacement, i.e., the initiation of the movement (manually identified on the time series as the intersection point between horizontal trend lines and a clearly detectable inclined trend lines when a cumulative displacement of 2 mm is achieved), and the final collapse of each landslide were exactly dated and timed with a precision of a few minutes, thus obtaining values ranging from a couple of weeks to a few hours (Fig. 6).
- 2.The total landslide displacement from its onset to its collapse, thus achieving values ranging from a couple of centimeters up to 1 m (Fig. 7).
- 3.Maximum velocity, thus achieving values ranging from 8 to 64 mm/h (Fig. 8).
It is worth noting that all the landslides, apart from 2, 4, and 8, were characterized by the maximum velocity just before the collapse, thus showing a trend of velocity continuously increasing since the landslide start until the failure. The peak of acceleration was registered in all the landslides some hours before the collapse instead (Fig. 8).
In other words, the final stage of the landslide displacement was characterized by an increasing velocity but a decreasing acceleration rate. The peak acceleration of the landslides ranged from 1 to 82 mm/h2, depending on the landslide. The value of acceleration decrease (computed as the difference between the peak acceleration and the last value available before the failure) ranged from 2 to 154 mm/h2. Furthermore, the value of acceleration decrease is closely proportional to the peak acceleration value of the landslides. Specifically, the ratio between the value of acceleration decrease and the acceleration peak ranges from 0.8 to 2.1. It is worth noting that the achieved values of total displacement, velocity, and acceleration were very similar to those measured in the experiments by Fukuzono (1985) and Moriwaki et al. (2004), which is a further confirmation of the quality of the dataset presented herein, which is comparable to the data derived from a controlled experiment.
Nevertheless, it is worth noting that pre-failure acceleration decrease is something never observed before, whether in field or experimental landslides. This evidence, which could play a relevant role in time of failure prediction, was obtained thanks to the availability of accurate and high-sampling rate data.
Time of failure analysis of detected landslides
The large dataset of events occurred on the same slope (which means similar conditions and features of the landslides), and the detailed displacement data available represent a rare occasion to test the efficacy of semi-empirical approaches based on time series of displacements (or derived quantities such as velocity and acceleration). Furthermore, to the authors’ knowledge, semi-empirical approaches have previously been tested mainly on large-scale real landslides, except for the large-scale experiments by Fukuzono (1985) and Moriwaki et al. (2004). Hence, the presented dataset is quite new in this field of research.
For the other eight remaining landslides, the error in the predicted time of failure is very small. Specifically, for five landslides (1, 3, 6, 8, and 10), positive prediction error was found (i.e., predicted time of failure postponed with respect to the real landslide occurrence), while for two landslides (4 and 7), negative error (i.e., predicted time of failure anticipated respect to the real one) was found. Our attention was mainly concentrated on the positive errors as they are contrary to general precautionary principles.
To ameliorate these problems, a new approach named average data Fukuzono (ADF) was developed. ADF is a modified version of the Fukuzono method consisting of the average and moving average velocity computed from temporal consecutive data. In the first case, the data were averaged iteratively, starting from the first data collected. In the case of the moving average, the data were averaged by using half of the dataset moved iteratively by one single step until the last half before the failure.
Figure 11 shows the results achieved using the ADF approach for landslide 2. As can be seen, both the average and the moving average approaches (Fig. 11b, c, e, f) significantly reduce the prediction error achieved by using the standard Fukuzono method and reported in Fig. 11a, d.
The high temporal resolution (about 5 min) on data collection allowed us to characterize also the displacement of landslides that occurred in a short time span (few hours) with a high data sampling rate, thus allowing us to identify the last steps of the creep process before failure.
The spatially continuous monitoring (i.e., displacement maps instead of single points) with a high ground resolution (few meters) allowed us to measure the displacement also of small landslides (few tenths of square meters) and, in some cases, to cross-check the inferred displacement from adjacent pixels.
The high accuracy in the displacement measurement allowed us both to characterize landslides with a small pre-failure displacement (even a couple of centimeters) and to well define their acceleration behavior.
The capability of TInSAR to collect data under any weather and lighting conditions allowed us to get information about events that occurred on complex environmental conditions and during the night.
Furthermore, it is worth stressing that the opportunity of monitoring an entire slope with a high spatial continuity was a key feature that allowed us to detect and characterize landslides that occurred also in sectors that were not originally planned.
First of all, as TInSAR provides only surface displacement data, no information was available on the deep behavior of the investigated landslides. However, considering the shallow feature of the herein discussed landslides, such a limitation is not considered relevant for this study.
Phase ambiguity problem, especially in the final evolution phases before the failure when the displacement velocities are higher. As a matter of fact, considering the usual small size of the landslides, spatial unwrapping algorithms demonstrated to be not very effective and only a temporal phase unwrapping was used, supported by the authors’ expertise in the TInSAR data processing. In this regard, it is worth noting that, for most of the landslides, the velocity in the final stage is well below the phase ambiguity threshold (4.5 or 9 mm in 5 min, if we assume that inversion of direction of movement is not reasonable for a landslide), and it is more than reasonable to assume that the phase ambiguity problems have not happened. However, in some cases (such as numbers 4, 9, and 10), this occurrence cannot be excluded. Furthermore, no other data derived from different sensors have been collected.
Anyway, instead of what was stated above, all these landslides were characterized by a similar displacement pattern in terms of peak values of velocity and acceleration and time-dependent evolution. However, a peculiar acceleration decrease before failure was observed in several landslides in our dataset. Such a behavior represents an anomaly with respect to available data from the literature and, at present, cannot be explained by existing theoretical creep models that assume a continuous increasing of acceleration until failure. This pre-failure acceleration decrease has several implications in the application of the already-existing semi-empirical models based on the creep theory, as demonstrated by the analyses performed by the simple linear Fukuzono approach (Fig. 9). The pre-failure acceleration decrease leads to a delay in the failure time with respect to that previously predicted by using the linear Fukuzono approach. Such behavior cannot be fully understood by the available data, and it requires further study. Nevertheless, we cannot neglect that the creep theory assumes a deformation under the condition of constant stress. Such a condition is not considered in our experiments because rainfall occurred during the experiments; hence, we cannot exclude the possibility that this condition could be one of the reasons for the pre-failure acceleration decrease. However, it is worth noting that the constant strain condition is also not considered in actual large landslides, such as those studied by Crosta and Agliardi (2003). The effect of rainfall was observed in landslides 2 and 9, where multiple acceleration and acceleration decrease phases occurred before the failure. In fact, these landslides are characterized by a high prediction error using the linear Fukuzono approach. The other landslides are always predicted with an error lower than 3.5 h. A drastic reduction in the forecasting errors was achieved by the herein proposed average data Fukuzono (ADF) approach (Fig. 10). This method, which consists of the average and moving average of the displacement data over time, is very effective at reducing the forecasting error, especially for unsteady cases characterized by important decelerating phases (Fig. 11). Furthermore, the ADF approach seems to be very effective in reducing the effect of the final decelerating phase (Fig. 13).
However, it is worth noting that the ADF approach is very sensitive to the number of available data, i.e., detailed time series significantly increase its efficacy, thus confirming the importance of high-resolution monitoring data for landslide prediction.
With the aim of deepening our understanding of landslide pre-failure behavior and, therefore, landslide prediction capabilities, a long-term field experiment using TInSAR was carried out. By analyzing the pre-failure behavior of ten landslides that occurred during the experiment, we showed that high-accuracy and high-sampling rate (a few minutes) of surface displacement data can be successfully used to obtain a good prediction of a landslide’s time of failure. Furthermore, the widespread and high-resolution information retrieved by TInSAR displacement images allows investigation from a remote position and the assessment of the pre-failure behavior of small landslides. In fact, traditional contact monitoring methods may fail due to a low data sampling rate, low accuracy and, particularly, punctual and localized information that may lead to a lack of information or to misinterpretations, especially for small-sized landslides.
A landslide is a very complex process characterized by an infinite range of variable features. The ability to obtain detailed and accurate information about the behavior of landslides is crucial to predict their future evolution (i.e., failure). Existing semi-empirical models based on the creep theory are very effective at predicting time of failure in laboratory experiments, but they are effective for real landslides only in a few cases. However, we demonstrated that by using accurate and high-sampling rate data, these models may be effectively used in real practice. Furthermore, a high data sampling rate allows the effective use of the ADF approach, which is able to drastically improve the efficacy of creep-based prediction methods in the case of “unsteady landslides.” Hence, it can be said that by using TInSAR continuous monitoring and advanced data processing solution such as the ADF approach, we can achieve the same results previously achieved for “ideal” laboratory landslides for real field landslides.
This research was partially funded by MIUR grants—Prin 2009, Research title “Relationships between human activities and geologic instabilities by integrating monitoring data and geological models related to already studied case histories”. Principal investigator: Francesca Bozzano.
Open Access This article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.