See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/383561961

Early warning of landslides based on statistical analysis of landslide motion

characteristics and AI Earth Cloud InSAR processing system: a case study of
the Zhenxiong landslide in...

Article in Landslides · August 2024

DOI: 10.1007/s10346-024-02350-5

CITATIONS READS

3 144

5 authors, including:

Bingquan Li Li Yongsheng
China University of Geosciences 61 PUBLICATIONS 561 CITATIONS
24 PUBLICATIONS 240 CITATIONS
SEE PROFILE
SEE PROFILE

All content following this page was uploaded by Bingquan Li on 30 September 2024.

The user has requested enhancement of the downloaded file.

 Technical Note

Landslides Bingquan Li · Yongsheng Li · Ruiqing Niu · Tengfei Xue · Huizhi Duan

DOI 10.1007/s10346-024-02350-5
Received: 28 March 2024
Accepted: 9 August 2024 Early warning of landslides based on statistical
© Springer-Verlag GmbH Germany,
part of Springer Nature 2024 analysis of landslide motion characteristics
and AI Earth Cloud InSAR processing system:
a case study of the Zhenxiong landslide
in Yunnan Province, China

Abstract Landslides, as a common natural disaster, pose a signifi- Displacement is the most indicative parameter of landslide
cant threat to human society and the natural environment, includ- evolution in both space and time (Saito 1965; Zhou et al. 2018;
ing loss of life, economic damage, and environmental destruction. Ding et al. 2021). On this basis, phenomenological early warning
Effective landslide early warning is key to reducing these negative models have been extensively developed (Salee et al. 2022; Zhang
impacts. However, current warning methods face two major chal- et al. 2023). These models, grounded in phenomenological crite-
lenges: one is the reliance on static threshold judgments, which not ria derived from the statistical characteristics of external triggers
only easily leads to false and missed alarms but also cannot adapt (such as rainfall and reservoir water levels) or internal geologi-
to complex and changing natural conditions. The second is the lack cal parameters associated with landslides, are favoured for their
of ground data support in areas with complex terrain, which greatly convenience and cost-effectiveness (Peres and Cancelliere 2021).
limits the application range and accuracy of traditional warning However, the correlation between external triggers and landslide
methods. To overcome these challenges, this study designed an occurrence remains uncertain in some cases, affecting their practi-
efficient processing system for Interferometric Synthetic Aper- cal effectiveness (Gariano et al. 2015). In contrast, utilizing inter-
ture Radar (InSAR) based on the (Artificial Intelligence) AI Earth nal geological parameters as warning criteria can more accurately
Cloud platform, integrated with the Comprehensive Standardized reflect the progression state of the landslides (Intrieri et al. 2019).
Deformation Index (CSDI) approach, to provide an early warning Despite some successful applications based on the above models,
analysis for the Zhenxiong landslide in Yunnan Province, China on the complexity and dynamism of intrinsic landslide characteristics,
January 22, 2024. Utilizing the cloud platform for rapid generation along with the variability of environmental factors and monitor-
of deformation rates and selection of characteristic deformation ing data errors, make it difficult to apply specific landslide early
points to reflect landslide trends, and applying the CSDI method warning models and their corresponding thresholds to other cases
for time-displacement curve analysis, enabled a fast and accurate (Carlà et al. 2017; Valletta et al. 2023).
landslide early warning. The research results show that the method To reduce false and missed alarms, various indicators for land-
proposed in this study can effectively warn of landslide events, sig- slide early warning have been proposed, including the displace-
nificantly improving the accuracy and practicality of the warning. ment tangential angular (Xu et al. 2011), the displacement speed
By combining InSAR technology with the CSDI model, this study ratio (DSR, Wang et al. 2014), the dimensionless displacement trend
not only addresses the challenges faced by traditional methods but (Cascini et al. 2022), the load-unload response ratio method (He
also provides new insights and solutions in the field of landslide et al. 2017), and the normalized velocity trends (Segalini et al. 2018).
early warning, demonstrating the great potential of technological However, these methods primarily focus on the acceleration phase
innovation in natural disaster management. of the deformation, neglecting the historical information of the
uniform deformation phase. This approach amplifies the features
Keywords Early warning · AI Cloud platform · InSAR · CSDI of the displacement curve, leading to potential judgments in early
warning results. To overcome these challenges, Zhang et al. (2024)
Introduction introduced a method based on the probability distribution features
Landslide disasters, among the most frequent, widespread, and of landslide velocity, which can adaptively adjust the thresholds,
devastating natural catastrophes globally, pose significant threats thereby reducing the false alarm rate of landslide early warnings.
to human life, property safety, and the stability of major engineer- However, this method solely relies on ground monitoring data and
ing facilities (Pecoraro et al. 2019; Casagli et al. 2023). The complex has not been applied using InSAR data.
structures and diverse formation mechanisms of landslides make Landslides frequently occur in complex mountainous areas,
achieving effective disaster prevention and mitigation challeng- making traditional ground surveys and monitoring insufficient
ing by relying solely on costly landslide control measures (Hussain for disaster prevention and mitigation. As a well-established
et al. 2023). It is crucial to deformation-based early warning systems monitoring technology, InSAR has been widely applied to moni-
with reliable monitoring data to prevent and mitigate geological tor landslides, earthquakes, and mining areas due to its ability to
hazards associated with landslides. This represents a vital direc- produce precise deformation measurements from a distance (Kang
tion for future development (Hussain et al. 2023; Fang et al. 2023). et al. 2017; Carlà et al. 2018; He et al. 2021, 2023, 2024; Li et al. 2022a,

Landslides
 Technical Note
b; Liu et al. 2024). Despite significant advancements in SAR data conditions of the study area, along with the SAR data set used.
processing, challenges such as large data volumes, lengthy process- The “Data processing strategy” section describes the methods and
ing times, and issues with image alignment persist. A variety of processing procedures employed in this study. The “Results and
software products for SAR image processing are currently avail- analysis” section presents the deformation rates of the landslide
able, including popular commercial packages such as SARSCAPE, and the early warning results, while the “Discussion” section ana-
SARproz, and GAMMA, as well as open-source alternatives such as lyzes the processing efficiency of the InSAR data processing sys-
SNAP, ISCE, DORIS, GMTSAR, StaMPS, PyRate, GIAnT, and Mintpy. tem, the impact of different mean normalization methods on the
While these open-source tools have facilitated the development of early warning outcomes and the triggering factors of the landslide.
InSAR technology, processing often requires substantial computa- Finally, the “Conclusions” section offers concluding remarks.
tional resources and complex parameter settings.
Differential Interferometric Synthetic Aperture Radar (D-InSAR) Study area and datasets
is a crucial component of InSAR technology (Massonnet and
Feigl 1998; Mora et al. 2003). However, its applicability for long- Geologic and climatic conditions of the study area
term monitoring is limited by factors such as decorrelation and At 5:51 am On January 22, 2024, a landslide disaster impacted Hex-
atmospheric disturbances. To address these limitations, techniques ing and Heping, two groups of villagers in Liangshui Village, Tang-
like the multi-temporal InSAR, including the Permanent Scatterer fang Township, Zhenxiong County, Zhaotong City, Yunnan Province
(PS) method suited for stable urban radar scatter characteristics (Fig. 1). The landslide mass had a volume of approximately 160,000
(Ferretti et al. 2001), and the Small Baseline Subset (SBAS) method ­m3, with an elevation difference of about 250 m. The landslide’s
(Berardino et al. 2002; Lanari et al. 2004), which mitigates errors due velocity reached 40–50 m/s, extending over a distance of 388 m.
to temporal decorrelation in D-InSAR and enhances the accuracy and This event buried 18 households with 44 deaths and 2 injuries. The
reliability of deformation monitoring, have been introduced. With landslide occurred in a mountainous area of mid-to-high altitude
the exponential growth of new wide-swath SAR satellite data, time- characterized by structural erosion and denudation along the
series analysis of massive datasets necessitates greater data storage northern edge of the Wumeng Mountains on the Yunnan Plateau.
and computational resource requirements. Personal computers or The area experiences a plateau monsoon climate, with an average
small workstations are currently inadequate for processing large annual precipitation of 914.6 mm, and is prone to cold weather
volumes of data, necessitating the use of large data server systems during the winter. Hexing and Heping villages are situated in a
and networked computing clusters. Therefore, high-performance basin of fault depression basin on the Yunnan Plateau, with the
computing (HPC) for Sentinel-1 data has become a current research southwestern side of the area featuring slopes of mountainous ter-
focus in the InSAR community (Yagüe-Martínez et al. 2016). HPC rain at the edge of the basin’s fault depression, where numerous
typically refers to computational systems and environments utilizing landslide deposits have developed. These villages were constructed
many processors or organizing multiple computers within a cluster. on ancient landslide deposits.
Institutions such as COMET in the UK and e-GEOS and TRE in Italy
utilize HPC environments for operational InSAR services (Costantini
et al. 2017). However, building the complex software and hardware Dataset description
environments required for powerful computing capabilities demands
significant human, material, and financial resources, which are This paper uses 57 ascending C-band SLC images from Sentinel-
beyond the reach of most InSAR researchers and production units. 1A acquired from January 5, 2022, to January 19, 2024, in the IW
Compared to the traditional HPC approaches, establishing a free mode with VV polarization (as seen in Fig. 1). Topographical
InSAR data processing system on existing cloud platforms offers a phase effects are removed using a Copernicus Digital Elevation
viable solution. Model (DEM) with a spatial resolution of approximately 30 m.
InSAR requires multiple observations over time for the study The temporal baseline combination generated according to the
area, involving researchers in the time-consuming and complex small baseline principle is shown in Fig. 2. Each SAR image and
process of data downloading, processing, and analysis to obtain the subsequent two images form an interference pair, and 111 inter-
landslide deformation rates and displacement time series. This ferograms are generated.
lengthy process is not conducive to timely landslide early warning.
Therefore, there is also an urgent need for an integrated cloud plat- Data processing strategy
form that can manage vast amounts of data and streamline these
processes. Automated InSAR data processing and time series extraction
In this study, we develop an InSAR data processing system on Rapid and automated acquisition of landslide deformation infor-
the AI Earth Cloud platform to extract the deformation rates and mation is crucial for early warning and preliminary analysis, as it
displacement time series of the landslide in Zhaotong, Yunnan. helps to reduce the effects of disasters and facilitates the imple-
By integrating our InSAR data processing system with the CSDI mentation of effective emergency response measures. To achieve
method, we perform landslide early warning post-analysis after the this goal, this study has developed a cloud-based InSAR process-
landslide occurrence. This represents the first successful attempt ing system in collaboration with the AI Earth Cloud’s extensive
to combine the landslide motion statistical characteristics method computational and storage capabilities (https://​engine-​aiear ​th.​
with an InSAR deformation analysis tool. aliyun.​com/). It is constructed using cloud-native technologies
The structure of this paper is as follows: the “Study area and and offers an automatically managed elastic big data environ-
datasets” section briefly introduces the geological and climatic ment, which is built entirely on the AI Earth Cloud infrastructure,

Landslides
Fig. 1  Study area and data coverage. a The green box indicates Sentinel-1 data coverage by ascending track 128. b View before the occur-
rence of the landslide on January 22, 2024. c View after the occurrence of the landslide on January 22, 2024

with all subsystems and modules deployed 100% on the AI Earth infrastructure, the platform effectively merges extensive remote
Cloud, including Container Service for Kubernetes (ACK), Max- sensing data with computational resources. It empowers users to
Compute, Polar distributed databases, and Object Storage Service study algorithm models at any desired scale and validate them
(OSS). As a geospatial data computing platform built upon cloud through interactive programming (Xu et al. 2023).

Fig. 2  Spatial and temporal baseline network of the used interferograms for ascending track 128

Landslides
 Technical Note
The system utilizes an SBAS approach enhanced by Graph- automated processing, deformation velocity maps are generated,
ics Processing Unit (GPU) acceleration and an automated work- and researchers can select characteristic points representing the
flow, with the overall process outlined in Fig. 3. The system auto- study object for time series analysis.
matically synchronizes Sentinel-1 data from the European Space
Agency (ESA) and the Alaska Satellite Facility (ASF) website to
the AI Earth Cloud platform. Users can set an area of interest CSDI method
(AOI) and the SAR data acquisition dates, choosing ascending or
descending SAR data based on the landslide’s direction and slope. The core function of the CSDI is to transform landslide veloc-
To ensure the accuracy of the InSAR results, manual configura- ity into a dimensionless index that conforms to the standard
tion of data processing parameters is required, including temporal normal distribution through cumulative probability transfor-
baselines, high pass filtering coefficients, Gaussian filtering coef- mation (Zhang et al. 2024). This involves selecting the repre-
ficients in the time domain, and 2D spatial low pass filtering coef- sentative
ficients. Once these parameters are set, the system can begin auto- { landslide deformation}time series from the InSAR data:
D = (d 0 , t0 ), (d 1 , t1 ), ⋯ (d N , tN ) , N is the total number of displace-
mated processing using the traditional SBAS method. To enhance ments in SAR data. In this study, positive values of deformation in
processing efficiency, a rapid GPU-assisted InSAR module opti- the radar line of sight indicate movement towards the radar, while
mizes the most time-consuming steps in interferogram genera- negative values represent movement away from the radar. For
tion, such as geometric co-registration, resampling, and Enhanced convenience, we use the negation of the radar line-of-sight (LOS)
Spectral Diversity (ESD) correction, significantly accelerating the deformation. The landslide velocity can be expressed as follows:
calculations (Yu et al. 2019; Li et al. 2022b; Duan et al. 2023). After

Fig. 3  Technical scheme of InSAR processing system based on the AI Earth Cloud platform

Landslides
 Di − Di−1 Processing flowchart
vi = − (1)
ti − ti−1
To achieve early warning of landslides, the following processing
where ti is the time of ith displacement, and vi is the velocity of the steps are adhered to by integrating landslide motion characteristic
ith monitoring. statistics with InSAR technology for landslide early warning:
To facilitate comparison, the velocity is normalized by Mean
normalization to between 0 and 1. This normalized velocity can (1) InSAR Time Series Processing: On the AI Earth cloud plat-
be expressed as follows: form, select the AOI, and time range, and set parameters (tem-
vi − v poral and spatial baselines, filtering coefficients).
viM = (2) (2) Selection of Typical Time Feature Points: After obtaining the
vmax − vmin
velocity over time, select representative landslide deformation
where viM is the ith velocity after Mean normalization. vmax and vmin time series.
are the maximum and minimum velocities, respectively. v is the (3) Output Warning Levels: Output warning levels (safe, vigilance,
average velocity. and alerting) based on CSDI values (< 1.5 for safe, 1.5–2.5 for
Given the non-normality of the probability distribution vigilance, > 2.5 for alerting). Continue monitoring if the result
of the velocity and the restriction of the skewed distribution indicates safe; otherwise, proceed to the next step.
function’s domain, the velocity data are split into positive and (4) Expert Decision-Making: Experts conduct a comprehensive
non-positive groups and processed separately to better fit its assessment based on the warning level and multidisciplinary
probability distribution. For velocity values that are less than data, make final decisions, and provide different levels and types
or equal to zero, the probability is as follows: of response measures to government agencies and communities.
M
P= (3)
N
Results and analysis
where M is the number of velocity values less than or equal to Using the method described in the “Automated InSAR data process-
zero. For the rest of the velocity data, their probability distribution ing and time series extraction” section, we obtained a mean veloc-
can be approximated utilizing the probability distribution func- ity map in the LOS direction of the landslide, as shown in Fig. 5.
tion (PDF). In this study, the Gamma, Beta, Lognormal, Loglogis- Although the majority of the landslide body has few effective InSAR
tic, and Logistic distributions are adopted. Then, R2 is utilized to measurement points (IMPs) due to the influence of SAR imaging
evaluate the goodness-of-fit and identify the optimal PDF for the geometric distortion, there are still IMPs at the top and bottom of
given. Based on the fitted results, select the best probability dis- the landslide body that can represent the deformation trend of the
tribution function. Through equal probability transformation, the landslide. From Fig. 5, we can observe that the maximum deforma-
PDF is converted into standard normal distributions. Therefore, tion rate of the landslide exceeds 50 mm/a.
the corresponding cumulative probability density (CDF) can also To further analyze the deformation pattern of the landslide, this
be obtained and marked as F(v). Thus, the total CDF of landslide study selected points P1, P2, and P3 for time series analysis, with
velocity (v) can be calculated as follows: results shown in Fig. 6. Generally, these characteristic points with
significant deformation are chosen at the top, middle, and bottom of
G(v) = P + (1 − P)F(v) (4)
the landslide, respectively. However, due to the lack of effective meas-
M urement points in the middle section of the landslide, we selected two
­ SDI can be determined based on the
The threshold of the C
cumulative probability density (CDF) of random variables in points with significant deformation at the bottom. Figure 6 reveals
the standard normal distribution and is expressed as follows: a clear acceleration trend in deformation around February 2023 for
points P1, P2, and P3. Notably, in the final month before the landslide
CSDIM = R{G(v)} (5) occurred, the deformation acceleration at points P1 and P2 further
where R{} is the cumulative probability density (CDF) of random intensified. The standard deviation reflects the dispersion of the data.
variables. For our study, we should select the point with the largest deforma-
Based on the characteristics of the probability distribution tion standard deviation. Based on this criterion, we chose point P2
of the standard normal distribution and corresponding defor- over point P1 because the standard deviations for the last 10 dates for
mation stages (Zhang et al. 2024; Xu et al. 2020), ­CSDIM values points P1, P2, and P3 are 12.9754, 14.17, and 6.30, respectively. Based on
below 1.5 are classified as safe, indicating slow deformation. these results, this study chose the displacement time series of point
Values between 1.5 and 2.5 represent vigilance, due to various P2 as the input for the ­CSDIM method.
Through velocity normalization in the “CSDI method” section,
external factors. Values above 2.5 suggest alerting, correspond-
velocity over time is fitted using the PDF, as shown in Fig. 7a. The
­ SDIM
ing to a catastrophic landslide process. The process of the C
result indicates that the best-fitting PDF is the Beta distribution,
model is shown in Fig. 4.

Landslides
 Technical Note

­ SDIM model
Fig. 4  The procedure for using the C

with a fit (R2) of 0.90. By comparing Figs. 6 and 7b, it is observed processing, respectively. Furthermore, our computational method
that the ­CSDIM curve is relatively flat, sensitive only to the accel- can process a pair of interferometric measurements in approxi-
­ SDIM
eration of velocity. Figure 7b also shows that the calculated C mately 15 min, achieving a resolution of about 30 m. This processing
M
values range between 0.5 and 3.0. The C ­ SDI reaches 2.7 before time is half that required by the GAMMA software, a widely used
failure, and the alerting warning level is determined by the pre- commercial InSAR processing tool. This represents a significant
defined threshold. The issuance of a singular alerting warning by reduction in data processing time, offering promising prospects
­CSDIM just before the landslide event underscores its effectiveness for rapid InSAR processing.
in landslide early warning. Further testing was conducted on the computational efficiency of
cloud platforms. We evaluated our SAR dataset on an InSAR process-
Discussion ing platform equipped with an Intel(R) Xeon(R) Gold 6240 CPU with
144 cores and an NVIDIA Tesla T4 GPU card with 256 CUDA cores
AI Earth Cloud platform InSAR processing efficiency and 16 GB RAM, featuring a bandwidth of 320 GB/s, and the computer
Compared to traditional InSAR registration methods, our study itself had 256 GB of memory. The download and time-series process-
introduces a GPU-assisted interferometric processing approach ing times for Sentinel-1A data were recorded (see Table 1). A total of 57
that significantly improves the efficiency of handling large Sen- SAR images were downloaded, requiring approximately 6 h at a rate
tinel-1 TOPS InSAR datasets. According to Yu et al. (2019), this of 6 min per image. For our dataset, a total of 111 interferograms were
processing chain accelerates performance by factors of 37.27 and computed, with the entire time-series process taking approximately
164.79 when compared to multi-threaded and single-threaded CPU 6 h. Therefore, traditional data processing methods would require

Landslides
Fig. 5  LOS displacement rate obtained with ascending Sentinel-1 images from January 2022 to January 2024. The red dashed line denotes
the landslide

approximately 12 h. In contrast, processing the same task on a cloud displacement curves, deformation patterns can be classified into
platform takes only 3 h. This comparison highlights that the efficiency five categories: long-term creep, oscillating, convergence, expo-
of InSAR processing algorithms on the AI Earth Cloud platform is nential, and step-like patterns (Chen et al. 2021). The displacement
approximately four times that of traditional platforms. curve in this study corresponds to the exponential pattern, char-
acterized by a significant change in the acceleration of landslide
Analysis of early warning results using another different deformation from a stable initial state under the influence of heavy
normalization method rainfall, earthquakes, and other intense triggering factors.
Traditional landslide early warning methods rely on subjec-
The most evident external manifestation of landslide motion tively set static thresholds, complicating the achievement of effec-
is the variation in surface displacement. Based on the trends of tive warnings. This study pioneers the integration of the AI Earth

Fig. 6  InSAR displacement time series of points P1, P2, and P3

Landslides
 Technical Note

­ SDIM
Fig. 7  a Probability density fitting of velocity after Mean normalization. b The result of C

Cloud platform’s InSAR processing system with the CSDI method aforementioned five displacement patterns using the CSDI method.
for landslide warning. The uniqueness of the CSDI method lies in The results indicate the method’s effectiveness across all cases.
its threshold determination approach: using the probabilistic dis- This study utilized mean normalization, yet it is important to
tribution characteristics of the standard normal distribution to note that various normalization methods may result in differences
ensure that the warning proportion remains within 0.3% of the total in warning. Furthermore, the C ­ SDIM method, by not focusing on
outcomes, thereby overcoming the limitations of traditional meth- the localized characteristics of displacement curves, increases its
ods. Zhang et al. (2024) analyzed 24 landslide cases exhibiting the practical application in landslide early warning systems. In con-
trast, the Max–Min normalization method proves more effective in
representing the actual motion state and phases of landslides. This
Table 1  The runtime comparison between the AI Earth Cloud plat- normalized velocity can be expressed as follows:
form and traditional method vi −vmin
viM−M = vmax −vmin (6)
Method Download Data Total
(h) processing (h) where viM−M is the ith velocity after Min–Max normalization. vmax
(h) and vmin are the maximum and minimum velocities, respectively.
AI Earth Cloud platform 0 3 3 If employing the ­CSDIM−M method for warning, it necessitates
the re-establishment of warning thresholds. Referring to the Saito
Traditional method 6 12 18 model, the landslide movement is divided into three stages from

Landslides
 ­ SDIM. b The result of CSDI M−M
Fig. 8   a The result of C

the statistical perspective: the initial deformation stage with initia- Analysis of inducing factors for the landslide
tion phase and braking phase (Stage I), the uniform deformation
stage (Stage II), and the accelerated deformation stage (Stage III) The overall terrain of the landslide area is steep, with the slope mor-
­ SDIM and C
(Zhang et al. 2024). The results of C ­ SDIM−M are shown in phology characterized by a polyline shape. This is particularly true for
M−M
Fig. 8. ­CSDI values below 0 are considered safe, indicating slow the source area, where the slope angle is around 60°, creating a high and
deformation. Values between 0 and 1 indicate caution, correspond- steep free face that lays the foundation for collapse. The polyline slope
ing to the initiation phase. Values between 1 and 2 suggest vigilance, shape provides a pathway for the high-speed movement of debris flows
aligning with the braking phase. Values above 2 represent alerting, resulting from the collapse, which leads to scraping, impact, and com-
corresponding to the accelerated deformation stage (Stage III). pression, causing severe damage to the houses at the base of the slope.

Fig. 9  Geological profile reflecting the sliding mode before and after the landslide

Landslides
 Technical Note

Fig. 10  Correlation between rainfall and mean cumulative deformation of the landslide

The landslide material primarily consists of Feixianguan Forma- warning. Moreover, the paper analyzes the impact of different
tion sandstone, muddy sandstone interbedded with claystone, repre- velocity normalization methods on the warning results, demon-
senting strata of mixed soft and hard characteristics. Such lithology strating that Mean normalization is more suitable for landslide
is typically prone to sliding due to its susceptibility to weakening. early warning compared to the Min–Max normalization method.
The hard rock layers, characterized by well-developed fractures, often Additionally, the analysis of the Zhenxiong landslide’s trigger-
facilitate the infiltration and accumulation of rainwater, forming inter- ing factors reveals that its occurrence is closely associated with
layer aquifers (Fig. 9). Conversely, the claystone acts as a relatively topography, lithology, and rainfall. This insight further supports
impermeable layer but, being a soft rock, it can easily soften under the significance of considering various environmental factors
the influence of groundwater, leading to a rapid decrease in strength in landslide early warning. The study also notes that the CSDI
and often forming a weak zone. Therefore, this combination of soft method may be compromised when there is a lack of significant
and hard lithologies poses a threat to slope stability. displacement time series data within the landslide mass, sug-
Rainfall is often a critical trigger for sudden geological disasters gesting its use as a supplementary decision-making tool. This
such as landslides. In the southwestern mountainous regions of China, observation highlights the need for comprehensive data collec-
although the precipitation is not heavy, it persists for extended periods, tion and integration of additional factors. Future work should
sometimes even up to a month (Fig. 10). The precipitation data (https://​ explore the integration of physical and mechanical properties
www.​ncei.​noaa.​gov/) comes from the Bijie weather station (105.233° E, with the CSDI model to enhance its predictive capabilities and
27.300° N). The characteristic of this rainfall is minimal surface runoff, improve warning accuracy, thereby providing a more robust
with the majority slowly infiltrating into the slope, becoming a primary framework for landslide early warning systems.
factor in triggering landslides. Due to the highly developed fractures
in the rock mass of the landslide area, rainwater predominantly infil- Author contribution
trates through these pathways. This infiltration turns into groundwater, All authors contributed to the study’s conception and design. Mate-
which then softens the rock and soil, reducing their strength. Under rial preparation, data collection, and analysis were performed by
the gravity of the unstable rock mass, the soil and rock on the slip Bingquan Li, Yongsheng Li, Huizhi Duan, Tengfei Xue, and Ruiqing
surface undergo progressive failure, eventually leading to a shear break Niu. The first draft of the manuscript was written by Bingquan Li,
in the lock segment and continuity of the slip surface, resulting in a and all authors commented on previous versions. All authors read
high-position collapse. and approved the final manuscript.
The steep terrain provides the conditions for the occurrence of
landslides, with the lithology of the landslide being the primary
internal and objective factor. Sustained precipitation infiltrates and Funding
softens the weak planes within the rock, and additional rainfall This research is funded by the National Key Research and Devel-
before the landslide leads to groundwater accumulation. When the opment Program of China (Grant No. 2021YFC3001903), the
critical point is reached, these factors together trigger the landslide. National Natural Science Foundation of China under Grant
(Grant No. 42374039), and the Alibaba Innovation Research Pro-
Conclusions gram (24024696-CRAQ717HZ11230008). The Sentinel-1 datasets
This study represents the first effort to integrate the InSAR dis- were freely provided by Copernicus and ESA. One arc-second
placement time series processed on the AI Earth Cloud plat- Copernicus DEM was freely downloaded from https://​space​data.​
form with the CSDI method for landslide early warning. This coper​nicus.​eu/. The figures were prepared using GMT and Origin
approach enables rapid and effective landslide forecasting, offer- software.
ing a viable solution for early warning systems. The primary
findings of this research highlight the potential of combining Declarations
InSAR technology facilitated by AI Earth Cloud and the CSDI
method to enhance the efficiency and accuracy of landslide early Conflict of interest The authors declare no competing interests.

Landslides
References Intrieri E, Carlà T, Gigli G (2019) Forecasting the time of failure of land-
slides at slope-scale: a literature review. Earth-Sci Rev 193:333–349.
Berardino P, Fornaro G, Lanari R, Sansosti E (2002) A new algorithm for https://​doi.​org/​10.​1016/j.​earsc​irev.​2019.​03.​019
surface deformation monitoring based on small baseline differential Kang Y, Zhao C, Zhang Q, Lu Z, Li B (2017) Application of InSAR tech-
SAR interferograms. IEEE Trans Geosci Remote Sens 40(11):2375–2383. niques to an analysis of the Guanling landslide. Remote Sens
https://​doi.​org/​10.​1109/​TGRS.​2002.​803792 9(10):1046. https://​doi.​org/​10.​3390/​rs910​1046
Carlà T, Farina P, Intrieri E, Ketizmen H, Casagli N (2018) Integration of Lanari R, Mora O, Manunta M, Mallorquí J, Berardino P, Sansosti E (2004)
ground-based radar and satellite InSAR data for the analysis of an A small-baseline approach for investigating deformations on full-
unexpected slope failure in an open-pit mine. Eng Geol 235:39–52. resolution differential SAR interferograms. IEEE Trans Geosci Remote
https://​doi.​org/​10.​1016/j.​enggeo.​2018.​01.​021 Sens 42(7):1377–1386. https://​doi.​org/​10.​1109/​TGRS.​2004.​828196
Carlà T, Intrieri E, Gigli G (2017) On the monitoring and early-warning of Li Y, Jiang W, Li Y, Shen W, He Z, Li B, Li Q, Jiao Q, Tian Y (2022) Coseis-
brittle slope failures in hard rock masses: examples from an open-pit mic rupture model and tectonic implications of the January 7 2022,
mine. Eng Geol 228:71–81. https://​doi.​org/​10.​1016/j.​enggeo.​2017.​ Menyuan Mw 6.6 earthquake constraints from InSAR observations
08.​007 and field investigation. Remote Sens 14(9):2111. https://​doi.​org/​10.​
Casagli N, Intrieri E, Tofani V, Gigli G, Raspini F (2023) Landslide detection, 3390/​rs140​92111
monitoring and prediction with remote-sensing techniques. Nat Rev Li Y, Jiang W, Zhang J (2022) A time series processing chain for geologi-
Earth Environ 4(1):51–64. https://​doi.​org/​10.​1038/​s43017-​022-​00373-x cal disasters based on a GPU-assisted sentinel-1 InSAR processor. Nat
Cascini L, Scoppettuolo M, Babilio E (2022) Forecasting the landslide evo- Hazards 111(1):803–815. https://​doi.​org/​10.​1007/​s11069-​021-​05079-9
lution: from theory to practice. Landslides 19(12):2839–2851. https://​ Liu W, Zhang Y, Meng X, Wang A, Li Y, Su X, Ma K, Li H, Chen G (2024)
doi.​org/​10.​1007/​s10346-​022-​01934-3 Forecast volume of potential landslides in alpine-canyon terrain
Chen M, Huang D, Jiang Q (2021) Slope movement classification and new using time-series InSAR technology: a case study in the Bailong
insights into failure prediction based on landslide deformation evolu- River basin. China Landslides 21(1):35–51. https://​doi.​org/​10.​1007/​
tion. Int J Rock Mech Min Sci 141:104733. https://​doi.​org/​10.​1016/j.​ s10346-​023-​02135-2
ijrmms.​2021.​104733 Massonnet D, Feigl K (1998) Radar interferometry and its application to
Costantini M, Ferretti A, Minati F, Falco S, Trillo F, Colombo D, Novali F, changes in the Earth’s surface. Rev Geophys 36(4):441–500. https://​
Malvarosa F, Mammone C, Vecchioli F, Rucci A (2017) Analysis of sur- doi.​org/​10.​1029/​97RG0​3139
face deformations over the whole Italian territory by interferometric Mora O, Mallorquí J, Broquetas A (2003) Linear and nonlinear terrain
processing of ERS, Envisat and COSMO-SkyMed radar data. Remote deformation maps from a reduced set of interferometric SAR images.
Sens Environ 202:250–275. https://​doi.​org/​10.​1016/j.​rse.​2017.​07.​017 IEEE Trans Geosci Remote Sens 41(10):2243–2253. https://​doi.​org/​10.​
Ding C, Feng G, Liao M et al (2021) Displacement history and poten- 1109/​TGRS.​2003.​814657
tial triggering factors of Baige landslides, China revealed by optical Pecoraro G, Calvello M, Piciullo L (2019) Monitoring strategies for local
imagery time series. Remote Sens Environ 254:112253. https://​doi.​ landslide early warning systems. Landslides 16:213–231. https://​doi.​
org/​10.​1016/j.​rse.​2020.​112253 org/​10.​1007/​s10346-​018-​1068-z
Duan H, Li Y, Jiang H, Li Q, Jiang W, Tian Y, Zhang J (2023) Retrospective Peres D, Cancelliere A (2021) Comparing methods for determining
monitoring of slope failure event of tailings dam using InSAR time- landslide early warning thresholds: potential use of non-triggering
series observations. Nat Hazards 117(3):2375–2391. https://​doi.​org/​ rainfall for locations with scarce landslide data availability. Landslides
10.​1007/​s11069-​023-​05946-7 18(9):3135–3147. https://​doi.​org/​10.​1007/​s10346-​021-​01704-7
Fang Z, Tanyas H, Gorum T, Dahal A, Wang Y, Lombardo L (2023) Speech- Saito M (1965) Forecasting the time of occurrence of a slope failure.
recognition in landslide predictive modelling: a case for a next gener- In: Proceedings of 6th International Conference on Soil Mechan-
ation early warning system. Environ Model Softw 170:105833. https://​ ics and Foundation Engineering, Montreal, Canada. pp 537–541.
doi.​org/​10.​1016/j.​envso​ft.​2023.​105833 1572543024011185664
Ferretti A, Prati C, Rocca F (2001) Permanent scatterers in SAR interfer- Salee R, Chinkulkijniwat A, Yubonchit S, Horpibulsuk S, Wangfaoklang
ometry. IEEE Trans Geosci Remote Sens 39(1):8–20. https://​doi.​org/​ C, Soisompong S (2022) New threshold for landslide warning in the
10.​1109/​36.​898661 southern part of Thailand integrates cumulative rainfall with event
Gariano S, Brunetti M, Iovine G, Melillo M, Peruccacci S, Terrana O, Vernari rainfall depth-duration. Nat Hazards 113(1):125–141. https://​doi.​org/​
C, Guzzetti F (2015) Calibration and validation of rainfall thresholds 10.​1007/​s11069-​022-​05292-0
for shallow landslide forecasting in Sicily, southern Italy. Geomorphol- Segalini A, Valletta A, Carri A (2018) Landslide time-of-failure forecast
ogy 228:653–665. https://​doi.​org/​10.​1016/j.​geomo​rph.​2014.​10.​019 and alert threshold assessment: a generalized criterion. Eng Geol
He Y, Wang W, Zhang L, Chen Y, Chen Y, Chen B, He X, Zhao Z (2023) 245:72–80. https://​doi.​org/​10.​1016/j.​enggeo.​2018.​08.​003
An identification method of potential landslide zones using InSAR Valletta A, Carri A, Segalini A (2023) Alert threshold assessment based on
data and landslide susceptibility. Geomatics Nat Hazards Risk equivalent displacements for the identification of potentially critical
14(1):2185120. https://​doi.​org/​10.​1080/​19475​705.​2023.​21851​20 landslide events. Nat Hazards 115(2):1549–1570. https://​doi.​org/​10.​
He Y, Zhao Z, Yang W, Yan H, Wang W, Yao S, Zhang L, Liu T (2021) A 1007/​s11069-​022-​05606-2
unified network of information considering superimposed landslide Wang L, Xie M, Chai X (2014) Research on method of displacement speed
factors sequence and pixel spatial neighbourhood for landslide sus- ratio for spatial evaluation of landslide deformation. Rock Soil Mech
ceptibility mapping. Int J Appl Earth Obs Geoinf 104:102508. https://​ 35(2):519–528
doi.​org/​10.​1016/j.​jag.​2021.​102508 Xu Q, Yuan Y, Zeng Y, Hack R (2011) Some new pre-warning criteria for
He K, Zhao M, Zhang Y, Zhang J (2017) Unload-load displacement creep slope failure. Sci China Technol Sci 54:210–220. https://​doi.​org/​
response ratio parameter and its application in prediction of debris 10.​1007/​s11431-​011-​4640-5
landslide induced by rainfall. Environ Earth Sci 76:1–16. https://​doi.​ Xu Q, Peng D, Zhang S, Zhu X, He C, Qi X, Zhao K, Xiu D, Ju N (2020) Suc-
org/​10.​1007/​s12665-​016-​6372-0 cessful implementations of a real-time and intelligent early warning
He Y, Zhao Z, Zhu Q, Liu T, Zhang Q, Yang W, Zhang L, Wang Q (2024) An system for loess landslides on the Heifangtai terrace, China. Eng Geol
integrated neural network method for landslide susceptibility assess- 278:105817. https://​doi.​org/​10.​1016/j.​enggeo.​2020.​105817
ment based on time-series InSAR deformation dynamic features. Int Xu H, Man Y, Yang M, Wu J, Zhang Q, Wang J (2023) Analytical insight
J Digit Earth 17:1. https://​doi.​org/​10.​1080/​17538​947.​2023.​22954​08 of earth: a cloud-platform of intelligent computing for geospatial big
Hussain M, Shuai Z, Moawwez M, Umar T, Iqbal M, Kamran M, Muneer M data. arXiv preprint arXiv:2312.16385 . https://​doi.​org/​10.​48550/​arXiv.​
(2023) A review of spatial variations of multiple natural hazards and 2312.​16385
risk management strategies in Pakistan. Water 15(3):407. https://​doi.​ Yagüe-Martínez N, Prats-Iraola P, Gonzalez F, Brcic R, Shau R, Geudtner
org/​10.​3390/​w1503​0407 D, Eineder M, Bamler R (2016) Interferometric processing of Sentinel-1

Landslides
 Technical Note
TOPS data. IEEE Trans Geosci Remote Sens 54(4):2220–2234. https://​ Yongsheng Li (*)
doi.​org/​10.​1109/​TGRS.​2015.​24979​02 National Institute of Natural Hazards, Ministry of Emergency
Yu Y, Balz T, Luo H, Liao M, Zhang L (2019) GPU accelerated interfero- Management of China, Beijing 100085, China
metric SAR processing for Sentinel-1 TOPS data. Comput Geosci
129:12–25. https://​doi.​org/​10.​1016/j.​cageo.​2019.​04.​010 Yongsheng Li
Zhang X, Chen L, Zhou C (2023) Deformation monitoring and trend Email: [email protected]
analysis of reservoir bank landslides by combining time-series InSAR
and Hurst Index. Remote Sens 15(3):619. https://​d oi.​o rg/​1 0.​3 390/​ Ruiqing Niu
rs150​30619 School of Geophysics and Geomatics, China University
Zhang J, Tang H, Li C, Gong W, Zhou B, Zhang Y (2024) Deformation of Geosciences, Wuhan 430074, China
stage division and early warning of landslides based on the statisti-
cal characteristics of landslide kinematic features. Landslides: 1–19. Tengfei Xue
https://​doi.​org/​10.​1007/​s10346-​023-​02192-7 DAMO Academy, Alibaba group, Beijing 100102, China
Zhou C, Yin K, Cao Y, Intrieri E, Ahmed B, Catani F (2018) Displacement
prediction of step-like landslide by applying a novel kernel extreme Tengfei Xue
learning machine method. Landslides 15:2211–2225. https://​doi.​org/​ Hupan Lab, Hangzhou 310023, China
10.​1007/​s10346-​018-​1022-0
Huizhi Duan
School of Earth and Space Sciences, Peking University,
Beijing 100871, China
Bingquan Li
School of Automation, China University of Geosciences,
Wuhan 430074, China

Landslides

View publication stats