Landslides are major hazards that may pose serious
threats to mountain communities. Even landslides in remote mountains could
have non-negligible impacts on populous regions by blocking large rivers and
forming dam-breached mega floods. Usually, there are slope deformations
before major landslides occur, and detecting precursors such as slope
movement before major landslides is important for preventing possible
disasters. In this work, we applied multi-temporal optical remote sensing
images (Landsat 7 and Sentinel-2) and an image correlation method to detect
subpixel slope deformations of a slope near the town of Mindu in the Tibet
Autonomous Region. This slope is located on the right bank of the Jinsha
River,
Landslides are major natural hazards in mountain regions and cause widespread disasters every year around the globe
(Petley, 2012; Zhang et al., 2020). Major landslides in remote mountain regions may pose serious threats to downstream communities by choking channels, which increases the risks of landslide-dammed-lake outburst floods
(Fan et al., 2020; Liu et al., 2019).
For example, a hillslope near the Baige village had two landslides, damming
the Jinsha River twice in 2018. The outburst floods caused widespread damage
along its route and affected areas as far as Yunnan Province,
Remote sensing techniques have been an efficient way to monitor slope
movement over large mountain regions (Du et al., 2020;
Handwerger et al., 2019). Optical passive and radar remote sensing provides the most frequently used
data to detect slope displacements. There are two kinds of mainstream
methods to derive slope movement. Synthetic-aperture radar (SAR) interferometry processing uses the
difference in phase images to derive subtle slope movement of a few
millimetres (Intrieri et al., 2018; Samsonov et al., 2020). However, large
ground displacements (e.g. of a few metres), dense vegetation or long time
intervals can lead to incoherence in phase images in this type of method
(Wasowski and Bovenga, 2014). Image correlation methods (also referred to as
pixel offset tracking used in SAR intensity images) constitute another type of
approach that uses SAR amplitude or optical images to correlate image patches
to measure slope movement and can derive subpixel ground displacements
from
In this work, using subpixel optical image correlation methods we report a landslide along the Jinsha River. Different from previous retrospective studies, the landslide in this work has not yet collapsed. We used multi-temporal Sentinel-2 images and found the slope is unstable and could pose a threat to downstream areas by blocking the Jinsha River.
The reported slope is
Topographic maps of the study area.
This area is tectonically active, and active faults run through this slope
from north to south. To the west of the faults are upper Palaeozoic strata
and to the east are Mesoproterozoic metamorphic rocks. Cracks and fissures
on the slope are visible from a 15 m resolution pan-sharpened false-colour
Landsat 7 image acquired in 2001 (Fig. 1b). These cracks and fissures may
be relics of historic earthquakes or precipitation. This part of the slope
has a percent slope of 45 % and a southeast aspect, with an azimuth
between 112.5 and 157.5
In this work, we mainly relied on Sentinel-2 optical images to derive slope movement. The European Space Agency's Sentinel-2 mission has two twin satellites in orbit, with a revisit time of less than 5 d. The Sentinel-2 optical imagery has 12 optical bands with wavelengths ranging from 440 to 2200 nm (Gascon et al., 2017). There are 4 bands with a spatial resolution of 10 m: blue, green, red and near-infrared bands. To derive slope movement, we used the red band because its wavelength is longer than those of other visible bands and is less influenced by the atmosphere. Compared to the near infrared, this band is less sensitive to vegetation and is more reliable for measuring slope deformation (Yang et al., 2019). We used the Level-1C product, which is orthorectified before distribution (Gascon et al., 2017).
This work used the COSI-Corr method, a correlation method for optical images to detect slope displacements (Leprince et al., 2007). To derive slope movement, two images in a roll should be used to form an image pair, including the base image and the target image. The base image is an earlier image, based on the image correlation algorithm (here we use the COSI-Corr) implemented to detect slope displacements in the target image (Leprince et al., 2007). For detailed parameters to use the COSI-Corr method, please refer to Yang et al. (2020).
In this work, we used three steps to detect slope displacements for the Mindu slope studied. For the first step, we used two image pairs (no. 1–no. 2) to find the stable and moving periods before and after November 2018. For the second step, we used 19 images in the stable period to estimate cumulative slope displacements in 5 target images in the moving period (image pair no. 3–no. 97). For the third step, we used another 9 images to derive displacements for every 2 adjacent images (image pair no. 98–no. 105).
In the first step, we used three Sentinel-2 images (on 13 November 2015, 12 November 2018 and 12 November 2019) to compose two image pairs (no. 1 and no. 2). The first image pair (no. 1) is composed of a Sentinel-2 image on 13 November 2015 and a Sentinel-2 image on 12 November 2018. Sentinel-2 images of the second pair (no. 2) were acquired on 12 November 2018 and on 12 November 2019.
List of the 19 base images in early 2018 and 9 targeted images in 2019. Base images were used to detect cumulative slope displacements in targeted images. Image pairs used in this step are no. 3–no. 97.
Detected slope displacements in image pair no. 1
Eight periods (image pair no. 98–no. 105) were used to derive the Mindu slope movement.
Detected image shifts (system error) in the stable zone. The EW (east–west) SD and NS (north–south) SD indicate uncertainties in the method, and the EW mean and NS mean were used to derive the final displacements in image pairs no. 1 and no. 2. The signal-to-noise ratio is denoted by snr.
By using the first two image pairs, we found the slope was stable from 13
November 2015 to 12 November 2018 and moved significantly from 12 November
2018 to 12 November 2019. Therefore, in the second step, we used two image
groups, a base image group in the stable period and a target image group in
the moving period, to detect cumulative slope displacements (Table 1). For
the base image group, there are 19 clear images without clouds in 2018. For
the target image group, we selected 5 images in 2019 (13 April, 17 July,
24 August, 5 October and 12 November) to detect cumulative displacements. In
all, there are 19
Misalignments between images can be estimated by selecting a stable zone (Bontemps et al., 2018; Lacroix et al., 2018; Yang et al., 2019). In this work, the stable zone was selected on the upper part of the landslide (red rectangles in Fig. 1b and c). Mean displacements estimated within the stable zone were used to correct image shifts. SDs of the displacements within the stable zone represent uncertainties, indicating the quality of the derived results for a given image pair. We selected this area because this stable zone is on the same slope as the landslide, which can minimize the influence of illumination and orthorectification errors.
In this work, we cross-validated measured slope displacements for 5 target images in 2019 identified in the second step. Uncertainties in the slope
displacements for a given target image were estimated from all 19 base
images in the stable periods. Standard deviations of these 19 measurements
were used to indicate their reliability. We further filtered out
displacements with moving directions that did not agree with the SRTM DEM-derived aspects. If there are 15
In Table 3, the EW mean and NS mean indicate the east–west (EW) and
north–south (NS) shifts in images in both image pairs calculated from the
stable zone. The EW SD and NS SD are SDs of displacements
in the stable zone to indicate image distortions. Low EW SD and NS SD
values indicate good performances during image orthorectifications. The
derived EW mean and NS mean were used to correct misalignments in image
pairs.
The base and target images for image pair no. 1 are from 13 November 2015 and
12 November 2018, respectively. The base and target images for image pair
no. 2 are from 12 November 2018 and 12 November 2019, respectively. The slope
remains stable in the first image pair, whereas detectable slope
displacements can be found in the second image pair (Fig. 2). The
durations of image pair no. 1 and pair no. 2 span 3 years and 1 year,
respectively. In Fig. 2a, we can see that the slope displacement from 2015
to 2018 was less than 2 m, whereas there was
As in Fig. 2, we can see that this slope remained stable from November 2015 to November 2018 and moved after November 2018. To derive time series of the Mindu slope displacements after November 2018, we used 19 base images in the stable period and 5 target images in 2019. All 19 base images are from early 2018, during which the slope was stable. Five selected target images were acquired on 13 April 2019, 17 July 2019, 24 August 2019, 5 October 2019 and 12 November 2019. For each target image in 2019, we calculated slope movement by using all base images. Therefore, there are 19 estimated slope displacements for each target image. We calculated the means and SDs of slope displacements for all target images (Fig. 3).
Means and standard deviations of the derived slope displacements in the five targeted images (Table 1). Detected means and SDs of slope displacement on 13 April 2019
Time series of the slope displacements. Image to the left shows
the slope displacements on 12 November 2019, and the map colour is shown as a
minimum–maximum linear-stretch type. Sub-panels
From Fig. 3, we can see that the mean displacements are a magnitude larger
than the SDs, which indicates that the displacements derived
between each target image and their base images agree with each other quite
well. Minor slope displacements were detected until April 2019 (maximum
3–4 m), whereas larger slope displacements can be observed in the later four target images (
We further selected six points on the slope to analyse time series of the
slope displacements in 2019 (Fig. 4). For most target images for the first
five points (p1–p5), most base images could derive
Slope displacements in different periods after the Baige floods (background images are Sentinel-2 data produced by ESA's Sentinel-2 satellites and downloaded via the GEE).
To analyse spatial deformation patterns in different periods, we selected nine
Sentinel-2 images forming eight image pairs (image pair no. 98–no. 105 in
Table 2, corresponding to eight periods in
Daily precipitation of the Baiyu meteorology station from October 2018 to February 2020.
Major landslides in mountains may dam river channels forming transient
lakes, the breach of which can result in catastrophic floods affecting downstream
communities (Dai et al., 2005; Fan et al., 2019; Liu et al., 2019). In this
work, we examined a hillslope near the town of Mindu along the Jinsha River. We
found the slope had significant movement from November 2018 to November
2019. Despite the area of the detected moving slope (715 577 m
High-spatial-resolution images from © Google Earth.
The image to the left was acquired on 30 March 2015 for the Mindu slope
In this work, we used the COSI-Corr method to derive slope displacements for the Mindu slope along the Jinsha River. The principle of this method is to use a sliding window to find pattern matches to derive displacements in image pairs (Leprince et al., 2007). Compared to the InSAR methods, this method is easier to understand and implement. In addition, image correlation methods favour larger displacements than InSAR methods. Limited by the wavelength of the SAR image, InSAR methods are versed in monitoring ground deformations on a millimetre to centimetre scale (Intrieri et al., 2018), whereas the capability of image correlation methods depends on spatial resolution of images. In general, image correlation methods are more reliable for deriving large ground displacements on a metre scale (Bradley et al., 2019; Lacroix et al., 2020). In this work, it might be quite challenging for InSAR methods to detect such large displacements. Long temporal intervals of a few months could lead to incoherence in SAR images (Li et al., 2019), whereas images (taken in the same season) with long temporal intervals of a few years can be used to derive reliable displacements given a stable land cover (Yang, 2020). Both methods can be affected by the atmosphere. Clear optical images without clouds should be used in image correlation methods. Although SAR images can penetrate thin clouds, the atmosphere could cause phase delay and lead to uncertainties in derived results (Li et al., 2019).
Both methods work well on bare land without vegetation, though dense vegetation could seriously affect InSAR methods (Intrieri et al., 2018). On the contrary, image correlation methods are less affected by vegetation cover as long as images in the pair are from the same season (Yang, 2020). As image correlation methods use pattern matches within an image pair, we speculate that vegetation density may not be a major challenge for the method. The Sentinel-2 images used in this work have four 10 m resolution optical bands (Gascon et al., 2017). In theory, any of these four bands may be used to derive slope displacements. But, an ideal band should not be sensitive to ground cover change unrelated to ground displacements, which could minimize background noise. In general, optical bands with shorter wavelengths are more prone to be affected by moisture in the atmosphere. Considering that the near-infrared band is very sensitive to vegetation, we used the red band in this work.
Both InSAR and image correlation methods can be impacted by complex terrains in mountain regions. Layover and shadow areas in SAR images should not be used in InSAR methods (Li et al., 2019). Similarly, shadows in optical images also influence derived results (Yang et al., 2020). To derive reliable results, optical images acquired during larger solar angles should be prioritized to minimize the influence of mountain shadows. Fortunately, there are algorithms that have been developed to restore information in mountain shadows in optical images (Shahtahmassebi et al., 2013), which may promote the efficacy of optical image correlation methods.
Many other factors may also influence the accuracy of slope deformation from image correlation methods, which include image orthorectification errors, different viewing angles during image acquisition and different illuminations in images (Stumpf et al., 2016; Yang et al., 2020). This work used the Sentinel-2 Level-1C product, which is orthorectified before distribution (Gascon et al., 2017). To correct for possible misregistration between the base and target images, we used a stable zone to calculate and correct image shifts. To reduce errors caused by different illuminations, all images used for the first two Sentinel-2 image pairs are from similar dates of different years.
The first two image pairs (no. 1 and no. 2) we mentioned above are composed of images of very similar acquisition dates in different years. Images of similar dates have similar zenith and elevation angles, which could minimize the influence of mountain shadows (Yang et al., 2020). To assess and reduce uncertainties in the second step, we first identified a stable period. Then, we used 19 base images in this stable period to derive cumulative displacements for a given target image in the moving period. The mean displacements from these 19 image pairs are expected to be more reliable than results from a single image pair. In addition, these 19 measurements can cross-validate each other and be used to estimate uncertainties by SD (Figs. 3 and 4).
There are a few strategies to suppress background noise in derived results, including selecting results with high signal-to-noise ratios (Lacroix et al., 2018; Yang et al., 2020) and integrating redundant information in time series of images (Bontemps et al., 2018). This work introduced a simple and efficient method by using the slope aspect to filter out slope movement that is different from the aspect. This is reasonable for this translational landslide as the mass moves downhill driven by gravity. This procedure could eliminate false slope movements and reserve true slope movement of the Mindu landslide. By integrating topographic information, this new procedure is expected to work well for ground movement in other regions that is consistent with slope configurations.
As we used orthorectified images, slope displacements derived in this work are horizontal movements. To derive ground movement along the slope, we need to consider local slope configurations. Because image correlation methods use sliding windows to detect similar patterns between the base and target images, precursors with horizontal rather than vertical ground movements can be detected. Landslides that have intact moving surfaces can be detectable by image correlation methods. For translational and rotational landslides, there are more horizontal than vertical ground movements, the former of which constitute the ideal landslide type to use in image correlation methods, whereas precursors of avalanches and rockfalls may be difficult to detect due to limited horizontal ground movement (Highland and Bobrowsky, 2013).
In addition, the smallest displacements that can be detected depend on the spatial resolution of optical images (Li et al., 2020; Stumpf et al., 2016). Although image correlation methods can detect subpixel ground movement, it is very challenging to detect moving surfaces that cover an area of a few pixels, as smaller window sizes could result in more background noise (Yang et al., 2020).
In this work, by using the COSI-Corr method and multi-temporal Sentinel-2 images, we found precursors of a major landslide along the Jinsha River in southwest China. Fissures on the slope probably existed before 2001, but the slope remained stable between November 2015 and November 2018. From November 2018 to August 2019, we detected significant slope displacements. The size of the activated part on the Mindu slope is similar to that of the 2018 Baige landslide, whereas the river width under the Mindu slope is half that of the Baige section. If this landslide continues to slide down and fails completely, it may block the Jinsha River leading to similar consequences to the Baige landslide.
By using an image correlation technique, we can track subpixel slope movement in optical remote sensing images. We also adopted an aspect constraint to pick out downslope movement and significantly reduced background noise. However, optical images, such as the Sentinel-2 images, can only detect slope movements of up to a few metres. To continuously monitor this slope, other data and methods (such as higher-spatial-resolution data or InSAR techniques) should be used. We also call for intensive monitoring of this slope and modelling of landslides that cause river blocking and subsequent flooding.
All Sentinel-2 images and the Landsat 8 image in this
work were downloaded from the GEE. The SRTM DEM and its derivative were
downloaded from the Geospatial Data Cloud website
(
LL and PS discovered the moving slope of this work. WY conducted analysis and drafted the manuscript.
The authors declare that they have no conflict of interest.
This article is part of the special issue “Remote sensing and Earth observation data in natural hazard and risk studies”. It is not associated with a conference.
Wentao Yang would like to express his gratitude to his large family for caring for his 2-year-old daughter while this work was underway.
This research has been supported by the National Science Foundation of China (grant no. 41807500) and the Second Tibetan Plateau Scientific Expedition and Research Program (STEP; grant no. 2019QZKK0606).
This paper was edited by Paolo Tarolli and reviewed by Mihai Ciprian Margarint, Luigi Lombardo, and one anonymous referee.