Introduction

Landslides are one of the most common geological hazards in the world, and the number of landslides and the amount of serious economic and human losses due to landslides continues to grow (Petley 2012; Froude and Petley 2018; Wei et al. 2019). In recent years, many catastrophic landslides have occurred in southwestern China, especially in Sichuan Province, and resulting huge losses caused widespread concern. For example, the landslides induced by the 2008 Wenchuan earthquake lead to thousands of casualties (Yin et al. 2009; Cui et al. 2011; Huang et al. 2012), and the 2017 Jiuzhaigou earthquake killed dozens of people (Fang et al. 2018; Sun et al. 2018). In this region, the number of landslides increases each year for various reasons, which makes landslide change detection research very necessary for developing rescue strategies after landslide failure (Fan et al. 2017).

Research on the change detection of landslides has attracted widespread attention from researchers around the world. At present, there are numerous methods of studying this important issue, such as field surveys and optical satellite image interpretation, multi-temporal digital elevation models (DEMs) from terrestrial laser scanning (TLS), or unmanned aerial vehicles (UAVs), numerous simulation, synthetic aperture radar (SAR) techniques, and normalized difference vegetation index (NDVI) time series from different platforms.

Field surveys and optical satellite images provide detailed information of land surface change characteristics, but the former work requires a great deal of manpower, financial support, and time. And the optical satellite images are significantly influenced by shadows and the landslide change information cannot be highlighted on the pixel scale (Joyce et al. 2009; Bontemps et al. 2018).

TLS and UAV can be used to obtain multi-temporal DEMs to determine how the landslide topography has changed represented by multi-temporal images (Jaboyedoff et al. 2012; Casagli et al. 2017). However, the equipment of TLS is expensive, only one side of a block can be observed, and dense vegetation needs to be removed manually (Kerle 2002; De Maio et al. 2015). For UAV, products are usually digital surface models (DSMs), not DEMs, which are obtained on places covered by vegetation (Hu et al. 2018; Rossi et al. 2018)

Different types of simulation software, like DAN–W, MassMov2D, particle flow code (PFC), MatDEM, the material point method (MPM), have been applied to studying parameters of velocity, energy, and land surface changes post landslide failure (Scaringi et al. 2018; Yang et al. 2019b; Wei et al. 2019). However, obtaining the micro-strength parameters of a rock-soil mass is tedious and time-consuming, and the result can have a low efficiency, large errors, and strong blindness (Wei et al. 2019). Furthermore, the high-quality topography data required for the modeling are difficult to obtain.

The SAR technique has also been used to study land surface changes (Zhao et al. 2012; Li et al. 2019). The phase information of the SAR can be used to obtain the time series deformation information before the landslide failure, and the post-failure landslide change intensity can be analyzed using the amplitude information (Qu et al. 2016; Li et al. 2019). However, the SAR technique is not suitable for landslides facing N-S (Casagli et al. 2017) and the data are significantly influenced by the vegetation area where landslides are easily triggered (Wasowski and Bovenga 2014). In addition, the SAR workflow is difficult to handle and the free access database for a chosen SAR satellite can be small.

NDVI is a common by-product of optical remote sensing images, and it is an effective indicator used to identify surface changes caused by landslides (Mondini et al. 2011; Yang et al. 2019a). Yang et al. (2013) used an NDVI time series to identify landslides change induced by the 2008 Wenchuan Earthquake. Lacroix et al. (2018) studied the ability of free access Sentinel-2 optical images to detect landslide changes before failure and determined the ability is influenced by the size of the landslide. Yang et al. (2019a) used the NDVI and Sentinel-2 images to track the landslide surface changes before the Jingsha River landslide failure and demonstrated that it has a huge potential. Moreover, the NDVI can reduce the impact of shadows on the results (Fiorucci et al. 2019). However, few studies have used the NDVI, especially with Sentinel-2 as a data base, to detect and examine the ability of the NDVI to identify the post-failure landslide change processes.

The goals of this study are to validate a simple but effective method of using a time series of Sentinel-2 images and the by-product NDVI to detect post-failure landslide changes over a 3-month period. We took the Huangnibazi landslide in Sichuan Province of China as an example to validate and expand Sentinel-2 satellite applications to detect post-failure landslide change.

Study area and data

Huangnibazi landslide

The medium-sized Huangnibazi landslide occurred in Li County, Sichuan Province, southwestern China, on the eastern side of the Qinghai-Tibetan Plateau, where the plateau transitions to the Sichuan Basin (Fig. 1a). This region has experienced strong earthquakes, such as the 2017 Jiuzhaigou earthquake (Mw 7.0) and the 2008 Wenchuan earthquake (Mw 7.9) (Yin et al. 2009; Sun et al. 2018; Huang et al. 2009). The failure type of the Huangnbazi landslide was creep-sliding and fracturing, which may have been reactivated and triggered by the effect of the Jiuzhaigou earthquake (Mw 7.0) (Shao 2018).

Fig. 1
figure 1

a Location of the Huangnibazi landslide (red dot), b enlarged topographic map of the study area, and c remote-sensing image of the study area from Google Earth

Full size image

The landslide occurred 500 m above the G317 road in deeply incised valley terrain (Fig. 1b). It took 3 months to reach a stable condition after the first signs of deformation on 8 August 2017 and buried the nearby houses and destroyed the villagers’ orchards (Fig. 1c). The shape of it looks like an arm-connected hand with a wide front and narrow back (Fig. 2a). There are bedding plans (Fig. 2b, f) and cracks on the rear edge (Fig. 2e) can be found. The undisturbed land was covered by very sparse vegetation (Fig. 2d) limited by the local dry-hot valley type climatic conditions (Li et al. 2019). The bedrock of the landslide mainly consists of Silurian and Devonian phyllite, and the landslide’s body is silt and gravel formed by weathering of the phyllite. The average slope of the landslide was about 35°, and the top of the slope was nearly upright bedrock. Its volume was estimated to be about 1,230,000 m3.

Fig. 2
figure 2

a Post-failure orthoimage of the Huangnibazi landslide acquired by a UAV. The red dashed lines outline the sliding boundary on 31 October 2017, shown in the right bottom corner. b Bedding plane on left side, c field investigation using real-time kinematic global positioning system (RTK-GPS), d landslide deposits, e cracks in rear edge, f bedding in left scarp, and g a house buried on 30 August 2017

Full size image

Data

We downloaded ten free-cloud images of Sentinel-2 Level-1C product free through the Sentinels Scientific Data Hub as database (Table 1). At this level, the registration accuracy of these multi-time series images is better than 0.3 pixels (Gascon et al. 2017). These images were cropped to an area of 0.63 km2 (0.86 km × 0.73 km) in order to better illustrate the landslide’s evolution and to obtain an accurate NDVI classification. Combined with the field work and UAV images, the results from the Sentinel-2 images were subjected to rigorous testing.

Table 1 Metadata of the free-cloud Sentinel-2 satellite images used in this study
Full size table

Methods

In this study, multi-temporal optical images taken by Sentinel-2 were used to detect post-failure landslide changes. First, free cloud images from Sentinel-2 were acquired, and the true color time series images were composited to gain a preliminarily understanding of the evolution stages of the landslide and to allow us to group the images into three categories and produce a red-green-blue (RGB) composite image. Then, four procedures were developed to detect landslide changes using the NDVI: (i) the NDVI of every Sentinel-2 image was calculated and divided into five categories using the natural break method; (ii) the time series NDVI values were obtained after examining the boundary differences recovered from the UAV orthoimages for similar times; (iii) the post-failure landslide change intensity was determined by counting the number of landslide scars in every image; and (iv) an RGB composite image was produced to detect post-failure landslide changes, and cross comparison and validation was conducted to test the effectiveness of the method by comparing our results to the results of previous studies. All of this work was conducted using the ENVI software version 5.2 and the GIS environment of the ESRI ArcGIS package. The workflow of our study is shown in Fig. 3.

Fig. 3
figure 3

Workflow of post-failure landslide change detection and analysis using Sentinel-2

Full size image

The detail data handling process were clearly described by Yang et al. (2019a). In this paper, we extend the definition of landslide scars and the method to the period after landslide failure. And to examine the detected landslide scars, we compared the landslide boundary difference of the UAV orthoimages taken on 31 October 2017 and extracted the landslide scars from the Sentinel-2 images taken on 06 November 2017 because of the short time interval between these images. Then we extracted the landslide scars of all ten Sentinel-2 images and produced an RGB composite to illustrate the slope movement based on three different change stages found in the time-series true color Sentinel-2 images. Finally, in this way, the vegetation depressed means of the newly emerged landslide scars were represented by warm colors (i.e., red), and the perished original landslide scars were represented by cold colors (i.e., blue).

Results

Landslide evolution revealed by the time-series true color images

To qualitatively investigate the landslide’s evolution, the time series true color images (Fig. 4) were synthesized using three 10 m spatial resolution bands: blue (B2 at 490 nm), green (B3 at 560 nm), and red (B4 at 665 nm).

Fig. 4
figure 4

Time-series true color Sentinel-2 images covering the Huangnibazi landslide, with dates of acquisition shown in the lower-left corners

Full size image

In the first image of Fig. 4a, obvious cracks could be detected on the slope surface and divided into two regions. Region A (orange dashed circle) exhibits a local slide and region B shows cracks are threaded together. In Fig. 4b, the landslide boundary is significantly different, and the loose soil is more broadly exposed due to the mass movement. Two grass-covered land areas on the landslide were not damaged. The upper small part C only changed slightly throughout the entire sliding period. While the lower part D can be divided into part E with sparse vegetation and part F with denser vegetation (Fig. 4c). Then, the landslide experienced a front and lateral extension stage (Fig. 4d–f) throughout November 2017. In the following days, no obvious boundary or texture changes in the landslide were identified reflecting that the landslide gradually reached a steady state. In conclusion, the landslide can be divided into three change stages: (i) a startup and acceleration stage (August–October 2017); (ii) a front and lateral edge expansion stage (November 2017); and (iii) a stabilization stage (December 2017 to January 2018).

Landslide changes revealed by the time series NDVI

NDVI time series classification

Using the natural break method, we determined that the classification images (Fig. 5) contain three major types: uniform bare land surface with no vegetation (class 1); a circumambient class (class 2), which contains evergreen grass vegetation influenced by class 1 for the 10 m spatial resolution of the Sentinel-2 image; and the three other evergreen shrubs with different types and vegetation densities. We can know that class 1 includes landslide scars, houses, and roads. In this type, the roads are the linear shapes in the bottom right corner, and the landslide scars are irregular.

Fig. 5
figure 5

Natural breaks method classification of the NDVI surrounding the Huangnibazi landslide. The area corresponding to part E in Fig. 4c is missing in Fig. 4c–j due to the sparse vegetation density. The red line represents the final landslide boundary identified during the field investigations

Full size image

Figure 5 shows all of the NDVI images and the red line is the final landslide boundary according to the in situ survey. From these images, we determined the change intensities and lateral spread processes of the different periods similar to Fig. 4. One difference in the landslide change trend is that the corresponding part E in Fig. 4c was missing in Fig. 5c–j. After checking with UAV images, we found that the vegetation density of part E was too low to be found by NDVI of the Sentinel-2 images on 10 m/pixel. However, the clear background contrast of landslide scars with the other classes is an obvious advantage of the method. This method makes it easier for stakeholders to identify landslide change and to gather information about the surrounding coverage conditions of a landslide.

Landslide scar extraction

To extract the landslide scars from each Sentinel-2 image, the NDVI thresholds of class 1 were determined using the natural breaks method, including the maximum and minimum threshold values. Figure 6 shows the maximum and minimum thresholds (dashed lines) of the class 1 areas in the 10 Sentinel-2 images in chronological order at irregular time intervals. From these results, it can be seen that the NDVI of the landslide scars remained at a low level (< 0.25). Meteorological factors like clouds could cause bias in the NDVI threshold values of the landslide scars. In addition, seasonality caused the NDVI values to decrease closer to winter (Yang et al. 2019a).

Fig. 6
figure 6

NDVI time series of the class 1 areas and typical point thresholds. If the value was larger than the maximum value, the pixel was treated as a non-landslide scar, and if the value was below the thresholds, the pixel was treated as a landslide scar. The typical decreasing trend of the NDVI pixels indicates that the landslide’s main scarp is expanding

Full size image

Landslide scar validation

By comparing the NDVI and UAV orthoimages, we found that minor differences existed between the landslide boundaries. Careful examination (Table 2 and Fig. 7) revealed that the omission errors are larger than the commission errors in the landslide and in part F. Part E was missed due to the low vegetation density caused by extension due to mass movement. This means that the Sentinel-2 images can be effectively used to identify landslide scars (95.38% for the landslide and 97.54% for part F), but it does miss some true scars (8.79% for the landslide and 25.19% for part F). Part C has a lower precision (63.27%) than the others areas because the right boundary was small cracks that could not be identified by Sentinel-2 on 10 m/pixel. These omissions mainly occurred on the uppermost part of the landslide mainly because the boundary of them was heavy extension cracks (Fig. 2e).

Table 2 Accuracy of the landslide, part C and F mapped using the NDVI and Sentinel-2
Full size table
Fig. 7
figure 7

Verification of the landslide scar extracted from the Sentinel-2 image taken on 06 November 2017. The black lines show the boundaries of landslide, parts C, E, and F were recovered from UAU Orthoimage taken on 31 October 2017

Full size image

Landslide scar count change analysis

By comparing the numbers of landslide scar pixels, we found that all of these numbers increased as time progressed with a slight fluctuation after 11 November 2017 (Fig. 8). For the roads, houses, and large boulders, which were classified as class 1 as well, the number of class 1 classifications was bigger than the number for the landslide scars. At the beginning, the number of landslide scars increased significantly from 13 August to 17 October 2017, after which the landslide scars only changed slightly. The three landslide change stages identified in the true color images also reflect in Fig. 8. The minor fluctuation during the stabilization stage may have resulted from changes caused by human activities such as field research.

Fig. 8
figure 8

Time series of the pixel numbers of class 1, the landslide scar, and change stages

Full size image

Landslide changes indicated by landslide scars counts and NDVI

To display the landslide change in pixel units, we made an RGB composite image (Fig. 9) by putting the count percentage images for 2017/12 to 2018/01 in the red channel, those for 2017/11 in the green channel, and those for 2017/08 to 2017/10 in the blue channel based on the change progress stages found above. So according to the composition time series principle, warm colors (i.e., red) indicate increasing counts and newly emerged features during the change period, whereas cold colors (i.e., blue) indicate a decreasing trend and the disappearance of previously existing landslide scars. The white pixels indicate that the pixels were constantly treated as landslide scars. The black-colored pixels were landslide scars that appeared during all of the periods and were stable. Gray pixels were not detected as landslide scars in any period.

Fig. 9
figure 9

RGB composite image of the landslide scars in the three periods: 2017/12 to 2018/01, 2017/11, and 2017/08 to 2017/10. The gray pixels represent the removed non-class 1, and the white pixels are those constantly marked as landslide scars. (The road is also in white, but it can be easily excluded). The points A (red), B (black), C (blue), and D (yellow) are the four points selected for the NDVI time series shown in Fig. 6

Full size image

From the RGB composite image, it can be seen that the colors on the landslide are distinctly different from the surrounding ground features. We further analyzed the NDVI time series of the four different color points shown in the RGB composite images (Fig. 6). Point 1 in Fig. 9 is a warm red, and its NDVI time series shows a typically decreasing NDVI (Fig. 6), which indicates that the vegetation was destroyed by landslide changes. Point 2 in Fig. 9 is black, and its NDVI time series is consistently larger than the landslide scar NDVI threshold (Fig. 6), which indicates that the pixels are in a steady state. Point 3 in Fig. 9 is blue, and its NDVI time series shows a typically increasing NDVI (Fig. 6), which indicates that the intact vegetation on the upper side moved downward, replacing the landslide scars on the lower side. This is indicative of the retrogressive landslide mode of landslide evolution. Point 4 in Fig. 9 is yellow, and its NDVI time series has an obvious turning point, which indicates that the pixels changes into landslide scars and the vegetation was destroyed by landslide runout from 7 September 7 to 17 October 2017. A few wrong NDVI values (e.g., point 2) were detected may be because the pixel value was influenced by the adjacent pixels of other types for the spatial resolution and the vegetation phenology.

Discussion

Compared with other free access satellites, Sentinel-2 has the advantages of a 10 m spatial resolution in the near-infrared, red, green, and blue bands, as well as a short revisitation interval. And they can be used as a resource to detect precursory motions and to monitor the period before a landslide occurs, giving it great potential value as a monitoring system (Lacroix et al. 2018; Yang et al. 2019a). However, its potential for post-failure landslide change detection has been less well-explored. In this study, we confirmed a simple but effective method of using all of the available Sentinel-2 optical images and the by-product NDVI to detect post-failure landslide changes. First, we analyzed the landslide’s evolution using true color Sentinel-2 images. Then, the post-failure landslide change was successfully detected by examining the changes in the landslide scars and the NDVI. We validated this work using UAV orthoimages acquired over a similar time period. This work demonstrates the potential of using Sentinel-2 time series images to detect post-failure landslide changes in medium-size landslides.

It should be noted that several factors may inevitably influence the results of the method used. First, no comparison work has been conducted on the influence of the geometry of the detected landslide, especially investigations of how the linear shape with widths smaller than the smallest unit can give rise to adjacent pixel influences. Considering the 10 m spatial resolution of Sentinel-2, we think this method may be more suitable for large and giant landslides. Second, the seasonality and the atmospheric influences on the effectiveness of the method cannot be ignored. For example, clouds may reduce the amount of available data and block regular rule investigation. Though a mask can be used to weaken the clouds’ influence, the accuracy still needs to improve (Coluzzi et al. 2018). Therefore, more advanced models or algorithms should be developed to solving this issue. Third, vegetation phenology influences the NDVI ranges and thresholds of the different classifications (Yang et al. 2019a). To overcome this factor, more available data are needed to summarize the phenology influence that can weaken the data provided by satellites (Deng and Zhu 2018). Finally, the method used to detect landslide movement and deformation must accommodate vegetation, and it is useless in the situation of a second sliding event in the existing landslide deposits that do not destroy the vegetation.

Despite these limitations, this study did accomplish post-failure landslide change detection using the vegetation index, which is treated as noise by other remote sensing techniques, in a simple way. Compared with the previous research methods applied to the Huangnibazi landslide (Shao 2018; Wang et al. 2019; Li et al. 2019; Xie et al. 2019), the method presented in this study has significant advantages in detecting landslide changes, as summarized in Table 3. Therefore, Sentinel-2 images have a great potential for use in detecting post-failure landslide changes.

Table 3 Comparison of the existing methods used to analyze the changes in the Huangnibazi landslide
Full size table

Conclusions

In this study, we explored the potential of using multi-temporal Sentinel-2 images to detect landslide changes after the Huangnibazi failure. Using ten images of this area, we confirmed that the vegetation index NDVI can be used to detect post-failure landslide changes in a simple method and an easy workflow. Using the time series true color Sentinel-2 images, we qualitatively divided the Huangnibazi landslide into a startup and acceleration stage (August–October 2017), a front and lateral edge expansion stage (November 2017), and a stabilization stage (December 2017 to January 2018). The change in the number of landslide scar pixels also reflected these three change stages in a quantitative manner. Based on the three different stages, an RGB composite was constructed to illustrate the different change periods of pixels on different color. Most importantly, the changes in the pixel units were illustrated by using cold colors to represent the retrogressive landslide mode. Using the database of the satellite operation, a more robust model could be constructed to reduce the influences of the clouds, vegetation phenology, and atmosphere on the landslide detection capacity and sub-pixel tracking ability could even be achieved. By incorporating other methods, small- or large-scale changes in the signal of a landslide on a regional scale can be used to identify pre- and post-failure changes in landslides. All of these findings reflect the huge potential of using the Sentinel-2 database for similar studies.