Environmental Controls and Transition of the Baige Landslide Deformation Revealed by Time-Series Remote Sensing Observations

Highlights

What are the main findings?

• A three-stage deformation pattern of the Baige landslide (slow pre-failure deformation, post-failure residual creep, long-term sustained acceleration) was quantitatively revealed via 2017–2024 SBAS-InSAR monitoring, accompanied by systematic restructuring of the deformation field and intensified spatial heterogeneity.
• The deformation driving mechanism underwent a fundamental transition: from multi-factor (temperature, precipitation, freeze–thaw cycles) short-lag (0–5 days) synergy before failure to a temperature-dominated long-lag (12 days) environmental control regime after failure, while gravity-driven creep and internal structural adjustment remained important background controls.

What are the implications of the main findings?

• This study clarifies the stage-dependent spatiotemporal evolution and driving mechanism transition of high-altitude fractured rock landslides, enriching the theoretical understanding of post-failure sustained activation.
• It provides quantitative methodological references and scientific basis for long-term monitoring, early warning, and risk prevention of similar landslides in the upper Jinsha River and alpine canyon regions.

Abstract

High-altitude rock slides frequently occur in the high-mountain canyon regions of the eastern Tibetan Plateau, posing significant disaster risks. The Baige landslide catastrophically failed in October 2018, blocking the Jinsha River and forming a major landslide-dammed lake. However, quantitative understanding of the spatiotemporal evolution and environmental control mechanisms remains insufficient, particularly regarding stage-dependent driving mechanisms. This study investigates the Baige landslide using mall Baseline Subset Interferometric Synthetic Aperture Radar (SBAS-InSAR), Seasonal-Trend decomposition based on Loess (STL) time-series decomposition, Principal Component Analysis–Independent Component Analysis (PCA-ICA) signal analysis, and slope-unit spatial statistics. Results indicate that: (1) deformation exhibited three stages separated by October 2018: slow pre-slide deformation, post-slide residual creep, and long-term sustained acceleration; (2) instability caused systematic restructuring of the deformation field, with valid pixels decreasing from 2766 to 560, deformation changing from slight positive line-of-sight (LOS) displacement to pronounced negative LOS displacement, and global standard deviation increasing from 21.40 mm to 40.55 mm, with stronger disturbances in the steep front zone; and (3) the driving mechanism shifted from short-term multi-factor control to a temperature-dominated long-term environmental control regime after failure, while gravity-driven creep and post-failure structural adjustment remained important background controls. Slope fragmentation and structural reorganization likely contributed to this transition.

1. Introduction

Landslides are among the most destructive geological hazards in mountainous regions. The eastern margin of the Tibetan Plateau, particularly the upper reaches of the Jinsha River, is characterized by intense tectonic activity and dramatic topographic relief, where high-elevation landslides frequently block rivers. Under the combined effects of complex geological structures and external triggers, this region is highly prone to “landslide–debris flow–dammed lake–flood” disaster chains, posing long-term threats to riverside communities, major infrastructure, and the ecological environment [1]. The Baige landslide, the most severe river-blocking event on the main stem of the Jinsha River in nearly a century, developed within a tectonically mixed rock belt [2]. It exhibits typical characteristics of large scale, high elevation, high velocity, staged sliding, significant disaster-chain effects, and persistent hazards. Its spatiotemporal deformation evolution and multi-factor driving mechanisms remain key scientific issues in geological hazard research in tectonically active regions [3].

The Baige landslide is located on the right bank of the Jinsha River in Baige Village, Boluo Township, Jiangda County, Qamdo City, Tibet Autonomous Region. During 10–11 October 2018, a massive high-altitude landslide occurred, with initial failure recorded at approximately 22:06 on 10 October 2018. The sliding mass rushed into the Jinsha River at extremely high speed, instantly blocking the river channel and forming a large landslide-dammed lake, with backwater extending up to 55 km upstream. In the early hours of 13 October, natural overflow of the debris dam began. Only 20 days later, fragmented rock masses in the rear traction zone gradually disintegrated, triggering a second large-scale landslide with a volume of approximately 3 × 10⁶ m³. This event directly blocked the previously formed natural spillway and created a larger debris dam with a higher breaching risk [4]. The cascading disasters caused severe destruction: upstream villages and farmland were submerged, while the outburst flood propagated downstream along the Jinsha River for more than 500 km, severely damaging villages, farmland, roads, bridges, and power facilities. In total, 102,000 people across three provinces were affected, and 86,000 were urgently evacuated. Direct economic losses in Yunnan Province alone reached 7.43 billion yuan. In addition, the residual landslide mass remains enormous, with an estimated volume of approximately 3470 × 10⁴ m³, and is still undergoing continuous creep deformation. Under extreme rainfall or seismic conditions, the residual mass retains a high risk of reactivation and river blockage, posing a long-term threat to the safe operation of major hydropower, water conservancy, and transportation projects in the upper Jinsha River [5].

Following the Baige landslide, extensive investigations were rapidly conducted by domestic scholars. Regarding geological characteristics, formation mechanisms, and disaster impacts, Deng et al. divided the landslide into a blocking zone, a main sliding zone, and a traction zone based on elevation through detailed field investigations [6]. They identified the progressive destruction of the blocking zone by the gravity-driven wedge-shaped mass in the main sliding zone as the key instability mechanism and reconstructed the entire movement process from rapid initiation to opposite-bank impact and jet formation. From a regional tectonic perspective, Feng et al. demonstrated that the landslide developed within the Jinsha River tectonically mixed rock belt, where the slope consists of tectonic lens-shaped blocks and highly fractured, mylonitized strata [7]. They classified it as a “translational–tensional hybrid” landslide formed by the coupling of tectonic fragmentation, gravitational unloading, and fluvial erosion. Liu et al. subdivided the landslide deposits into five zones and used numerical simulations to identify instability risks in the forezone under heavy rainfall and seismic loading [5]. Cao et al. further delineated the controlling fault zones and potential slip surfaces using geophysical exploration, providing key evidence for mechanistic studies [2].

In terms of deformation monitoring and evolution analysis, multi-source remote sensing has become a core approach. Through interpretation of historical remote sensing imagery, Xu et al. reconstructed a deformation history spanning more than 50 years [4]. Liu et al. integrated Keyhole, Gaofen satellite, Unmanned Aerial Vehicle (UAV), and airborne Light Detection and Ranging (LiDAR) data and, using SBAS-InSAR technology, obtained a maximum deformation rate of 130–135 mm/a in the middle and upper slope sections from January 2018 to May 2024, enabling quantitative long-term monitoring [5]. Yang C. S. et al. and Yang F. et al. employed time-series Interferometric Synthetic Aperture Radar (InSAR) and pixel offset tracking techniques, respectively, to reconstruct the spatiotemporal deformation characteristics before and after failure [8,9]. Building on these studies, Li et al. integrated multi-platform Synthetic Aperture Radar (SAR) datasets from Sentinel-1 and ALOS-1/2 and applied multi-geometry pixel offset tracking (MG-POT) to reconstruct a high-precision 3D deformation time series from 2016 to 2018 [3]. By introducing Green–Lagrange finite strain tensor analysis, they characterized surface kinematic patterns from both displacement and strain perspectives, revealing a longitudinal kinematic zoning pattern of “tension-dominated rear section—nearly rigid translation in the middle—compression-dominated front.” This provided a new analytical framework for understanding the complex deformation behavior of large high-altitude rock landslides.

However, existing research has achieved a fairly systematic understanding of the geological genesis and disaster effects of the Baige landslide, yet shortcomings remain. Although previous studies have conducted deformation monitoring and stage division based on remote sensing data, quantitative correlations have yet to be established between long-term deformation time series and environmental driving factors such as precipitation, air temperature, and freeze–thaw processes. The contribution, coupling effects, and time-lag response of environmental factors to landslide deformation have not been quantified by stage, nor has the intrinsic transition from multi-factor synergistic regulation to single-factor dominance been revealed against the backdrop of slope structural failure.

This study addresses the lack of quantitative correlations between deformation and environmental factors and the insufficient understanding of environmental control mechanisms and their transition in current Baige landslide research. Taking the Baige landslide as a typical case, this study integrates SBAS-InSAR long-term deformation monitoring, STL time-series decomposition, slope unit spatial statistics, and cross-correlation analysis. It aims to establish quantitative correlations between long-term deformation time series and environmental factors such as precipitation, air temperature, and freeze–thaw processes, to quantify by stage the contribution, coupling effects, and time-lag response of environmental factors to landslide deformation, and to reveal the intrinsic transition of the driving mechanism from multi-factor synergistic regulation to single-factor dominance before and after failure. This will lead to a systematic understanding of the environmental control mechanisms governing the long-term deformation of high-elevation rock landslides and provide a scientific basis for the risk prevention and control of similar hazards in the upper Jinsha River.

The main contributions of this study are as follows:

(1) Using SBAS-InSAR technology, long-term deformation datasets from 2017 to 2024 were obtained to systematically analyze the overall distribution characteristics of surface deformation across the Baige landslide area. The fundamental development pattern of post-failure continuous deformation in high-altitude fragmented rock landslides was clarified, providing a data foundation for subsequent analyses of spatiotemporal evolution and driving mechanisms.

(2) By accurately delineating the evolutionary stages of the landslide, deformation rates, trends, and activity intensities at each stage were quantitatively characterized. This revealed the complete temporal evolution pathway of the Baige landslide, characterized by “pre-failure slow deformation, post-failure residual creep, and long-term sustained acceleration,” and clarified the intrinsic temporal mechanisms underlying the continuous reactivation and progressive deterioration of the residual landslide mass.

(3) Combined with slope-unit-based spatial quantification methods, the dynamic differences in deformation fields before and after failure were comparatively analyzed. By quantifying indicators such as effective pixels, mean deformation, and dispersion, the spatial differentiation patterns jointly controlled by slope gradient and spatial location were identified, and the systematic reshaping effect of large-scale instability on the regional deformation field was clarified.

(4) By integrating time-series decomposition and correlation statistical analysis, quantitative relationships between deformation time series and environmental factors were established. The differences in driving mechanisms before and after landslide failure, were systematically compared, revealing a transition from multi-factor synergistic regulation before failure to a more concentrated temperature-dominated environmental control regime after failure. In addition, the staged evolution of response lag effects and the transformation mechanisms of environmental driving factors induced by slope structural failure were clarified.

This paper is organized into five sections. Section 1 introduces the geological background, research status, existing gaps, and objectives of this study. Section 2 describes the study area, data sources, and methodologies, including SBAS-InSAR processing, STL decomposition, slope unit spatial statistics, and environmental factor correlation analysis. Section 3 presents the spatiotemporal deformation characteristics, evolutionary stages, spatial differentiation patterns, and driving mechanism transformations of the Baige landslide. Section 4 discusses the main findings, limitations, and implications of the study. Section 5 summarizes the major conclusions and provides references for the monitoring and risk management of similar high-altitude rock landslides.

2. Materials and Methods

2.1. Study Area and Data Sources

2.1.1. Study Area

The Baige landslide is located within the Jinsha River suture zone along the eastern margin of the Tibetan Plateau, straddling the boundary between Jiangda County (Tibet Autonomous Region) and Baiyu County (Sichuan Province). It is situated on the right bank of a deeply incised V-shaped valley in the upper reaches of the Jinsha River (98°42′7″E, 31°4′57″N), at an average elevation of approximately 3600 m, with a relative relief exceeding 1000 m. The area represents a typical alpine canyon geomorphic system (Figure 1), characterized by steep slopes with gradients generally exceeding 30°, which are highly conducive to landslide development. The study area is located near the Jiangda–Boluo–Jinsha River fault zone, with the rear edge of the landslide controlled by the Boluo–Muxie reverse fault, indicating strong tectonic activity and a complex stress regime. The exposed strata belong to the Permian–Lower Triassic Gangtuo Formation and are mainly composed of tectonically mixed rocks. Under the combined influence of repeated tectonic deformation and weathering-induced stress release, the rock mass has become highly fragmented, providing unfavorable material conditions for long-term slope stability and promoting creep deformation.

The study area has a plateau semi-humid climate with distinct dry and wet seasons and an extremely uneven spatiotemporal distribution of precipitation. The long-term average annual precipitation is approximately 650 mm, with the rainy season (June–September) accounting for 87.2% of the annual total. The coupled effects of localized heavy rainfall and high-altitude freeze–thaw cycles are the core external triggers driving the continuous deterioration of slope stability. In October–November 2018, the Baige landslide experienced two successive episodes of high-altitude collapse and river block-age, marking two critical turning points in its deformation evolution. The rapid blockage of the river by the landslide mass formed a massive landslide-dammed lake, subsequently causing significant disaster losses. Following the landslide’s destabilization, the residual mass did not stabilize but instead continued to exhibit significant residual creep deformation, providing an excellent natural experimental platform for systematic research into the spatiotemporal evolution characteristics and driving mechanisms of the Baige landslide.

2.1.2. Data Sources

In this study, Sentinel-1 SAR data acquired between 2017 and 2024 were used to monitor pre- and post-failure deformation characteristics and to capture long-term deformation trends associated with landslides. Sentinel-1 is a dual-satellite Earth observation mission developed by the European Space Agency (ESA), equipped with C-band SAR that enables all-weather, day-and-night imaging. Detailed data information is provided in

Table 1. Key Parameters of the Sentinel-1 Satellite.

Data Type	Date	Number of Scenes	Incidence Angle /°	Azimuth Angle /°	Data Source
Sentinel-1 (Ascending)	18 March 2017 – 10 July 2024	166	33.9	−10.1	https://search.asf.alaska.edu, 13 February 2026
Sentinel-1 (Descending)	11 February 2017 – 25 December 2024	230	31.3	192.8	https://search.asf.alaska.edu, 13 February 2026
DEM	2014	4	/	/	https://earthexplorer.usgs.gov, 13 February 2026

.

To reduce phase errors induced by orbital inaccuracies, precise orbit ephemerides (POD) provided by ESA were applied for orbit correction. In addition, the Shuttle Radar Topography Mission (SRTM) digital elevation model (DEM) with a spatial resolution of 30 m, jointly developed by the National Aeronautics and Space Administration (NASA) and the National Imagery and Mapping Agency (NIMA), was used to mitigate topographic phase effects.

Meteorological data used for environmental driving-factor analysis were obtained from the National Oceanic and Atmospheric Administration (NOAA) Global Historical Climatology Network Daily (GHCN-Daily) dataset for the Dege meteorological station (station ID: 56144), located at 31°44′N, 98°34′E with an elevation of 3201.2 m. Daily precipitation and air temperature data (including daily maximum and minimum temperature) were extracted from this station-based dataset. As a long-term continuous observational record, it provides ground-based meteorological information for the study area and represents the closest available climatic observations to the Baige landslide region.

To evaluate the representativeness and reliability of the station-based observations, ECMWF Reanalysis v5-Land (ERA5-Land) reanalysis data was used as an independent reference dataset. The comparison results show strong consistency between ERA5-Land and station observations in air temperature variability, with high correlation at both daily (R = 0.94) and monthly scales (R = 0.96). In addition, freeze–thaw indicators derived from ERA5-Land show good agreement with station-based calculations, with an overall cycle accuracy of 87.97% and strong correlation in thawing degree days (TDD, R = 0.88). These results confirm that the Dege Meteorological Station provides a reliable representation of regional climatic variability for the study area.

Prior to the environmental-factor analysis, all meteorological datasets were temporally aligned with the deformation monitoring period to ensure consistency among multi-source observations. Missing records were filled using time-based linear interpolation to construct continuous daily time series. Daily mean temperature values were subsequently used to derive freeze–thaw indicators. Furthermore, because deformation measurements and environmental variables have different units and magnitudes, all variables were standardized using the Z-score method before PCA, ICA, and cross-correlation analyses to eliminate scale effects and improve comparability among heterogeneous datasets.

2.2. Methods

2.2.1. Overview

This study utilizes Sentinel-1 time-series SAR imagery and applies the SBAS-InSAR technique to retrieve long-term surface deformation. By integrating STL time-series decomposition, multivariate statistical analysis, and multiscale spatial partitioning, the spatiotemporal evolution characteristics and driving mechanisms of the deformation field are systematically analyzed. An integrated analytical framework is established for time-series feature extraction, spatial pattern quantification, and driving mechanism analysis, providing methodological support for investigating the long-term deformation evolution and disturbance-source characteristics of large-scale post-disaster landslides (Figure 2).

Based on long-term deformation results derived from SBAS-InSAR inversion, the Baige landslide is selected as the core study object. Segmented linear regression is employed to delineate deformation evolution stages, identify rate inflection points, and characterize activity intensity. Combined with slope-unit-based and global–subarea spatial statistics, deformation mean values, extrema, standard deviations, and coefficients of variation are calculated to quantitatively characterize spatiotemporal pattern changes and spatial heterogeneity before and after landslide failure. STL time-series decomposition is further applied to separate deformation trends from periodic signals, while PCA, ICA, and cross-correlation analyses are used to identify dominant driving factors, quantify driving intensities and response lags, and reveal the transformation characteristics of deformation driving mechanisms before and after failure. Ultimately, this study provides a systematic quantitative analysis of the spatiotemporal evolution and intrinsic dynamic driving mechanisms of the Baige landslide.

2.2.2. Data Processing

SBAS-InSAR technology can effectively mitigate decorrelation caused by changes in surface scattering properties by selecting interferometric pairs through temporal and spatial baseline thresholds [10,11]. In this study, Sentinel-1A data acquired between 2017 and 2024 were used, including 230 descending-track images and 166 ascending-track images (Level-1 Single Look Complex data). SBAS-InSAR processing was conducted to retrieve surface deformation in the study area.

The same temporal and spatial baseline thresholds were consistently applied to both ascending and descending datasets (temporal baseline ≤ 180 days and spatial baseline ≤ 45%), ensuring methodological consistency between the two orbit geometries. Based on standard SBAS processing practice, interferometric pairs were generated independently within each processing subset.

Due to differences in orbital viewing geometry, spatial coverage, and interferometric network connectivity, the number of feasible interferometric combinations may vary even when identical baseline thresholds are applied. Therefore, the ascending and descending datasets inherently produce different interferometric pair densities under the same processing strategy. In this study, a total of 1235 ascending-track interferometric pairs and 2873 descending-track interferometric pairs were generated for subsequent analysis. This difference reflects the intrinsic combinatorial structure of SBAS networks rather than differences in data quality or methodological inconsistency.

To ensure temporal continuity, annual deformation time series were first generated separately for each processing period and then concatenated to construct a continuous long-term deformation record covering 2017–2024. All annual deformation products were referenced to a unified stable reference frame defined by twenty bedrock reference points, ensuring consistency across multi-temporal SBAS processing results.

For interferogram processing, an adaptive Goldstein filter was applied to suppress noise and enhance phase continuity, followed by phase unwrapping using the Minimum Cost Flow (MCF) algorithm, with a coherence threshold of 0.3. This threshold has been widely adopted in SBAS-InSAR studies conducted in mountainous and vegetated environments and represents a practical compromise between phase reliability and spatial coverage. Considering the strong topographic relief, vegetation cover, and temporal decorrelation characteristics of the Baige landslide area, a threshold of 0.3 was selected to retain sufficient valid observations while minimizing the propagation of phase-unwrapping errors. Low-coherence areas were masked to minimize error propagation and improve unwrapping accuracy [12].

During error correction, precise orbit ephemerides from the ESA were used to remove orbital errors. The 30 m resolution SRTM DEM was then incorporated to eliminate topographic phase contributions, isolating the deformation phase. In addition, GACOS data were used to further correct atmospheric phase errors, reducing the influence of tropospheric delays on deformation measurements. Initial deformation rates were estimated using a linear model, and the interferometric phase was subsequently flattened.

Because SBAS-InSAR provides relative deformation measurements, stable reference points were selected before time-series inversion to establish a reliable deformation reference frame. Reference points were manually selected according to four criteria: (1) absence of residual topographic fringes after topographic phase removal; (2) absence of obvious deformation fringes and location outside identified deformation zones; (3) absence of phase discontinuities, phase jumps, or phase-unwrapping artifacts; and (4) consistently high coherence throughout the interferometric network. Considering that coherence conditions vary among interferometric pairs and that a single reference point may not remain suitable for all interferograms, twenty reference points were distributed across stable bedrock areas surrounding the landslide and used to define the deformation reference frame during SBAS processing. All reference points were located outside the identified deformation zones and were selected from geomorphologically stable bedrock sectors surrounding the Baige landslide. These points exhibited consistently high coherence and stable phase behavior throughout the observation period, thereby minimizing the influence of landslide deformation on the reference frame. These points were visually inspected to ensure temporal phase stability and to minimize the influence of residual phase ramps and local deformation signals. The adoption of multiple stable reference points reduces potential biases associated with individual reference-point instability and improves the robustness of the retrieved deformation field.

Finally, LOS deformation time series were inverted from the corrected phase using Singular Value Decomposition (SVD) [13]. A multi-looking ratio of 1:3 (range: azimuth) was applied, yielding final deformation products with a spatial resolution of approximately 15 × 20 m.

Both ascending- and descending-orbit datasets were processed independently. Owing to the steep canyon topography of the Baige landslide area, the ascending-orbit observations provide more continuous spatial coverage and are less affected by ter-rain-induced geometric distortions such as layover and shadow. In addition, the ascending dataset contains a larger number of valid deformation observations within the landslide body. Therefore, the ascending-orbit dataset was selected as the primary dataset for detailed deformation analysis, whereas the descending-orbit results were used as Supplementary Materials for evaluating the overall deformation evolution trend.

2.2.3. Spatial-Temporal Evolution Analysis of the Baige Landslide Based on SBAS-InSAR and Slope Units

In this study, a unified LOS deformation sign rule is adopted: positive values correspond to ground displacement toward the satellite (uplift), and negative values represent ground displacement away from the satellite (subsidence).

This section is based on long-term deformation inversion results derived from SBAS-InSAR. By integrating time-series extraction at typical monitoring points, segmented linear regression, and slope-unit-based spatial statistics, the deformation evolution patterns before and after the Baige landslide failure are quantitatively analyzed from both temporal and spatial perspectives, thereby characterizing the spatiotemporal patterns and differentiation features of surface deformation.

Temporal deformation evolution was analyzed using representative monitoring points extracted from the SBAS-InSAR deformation field. Three monitoring points located within the main deformation zone were selected to characterize the long-term deformation behavior of the Baige landslide. Cumulative deformation time series from 2017 to 2024 were extracted to construct deformation–time evolution curves. Taking the October 2018 failure event as the temporal boundary, the deformation process was divided into pre-failure and post-failure stages. Segmented linear regression was then applied to quantify deformation trends and evaluate changes in deformation rates between different evolutionary stages. The coefficient of determination (R²) and statistical significance level (p < 0.05) were used to assess the reliability of the regression results [14].

To distinguish long-term deformation evolution from seasonal environmental influences, STL was applied to the deformation time series. The original deformation signal was decomposed into trend, seasonal, and residual components, allowing the separation of long-term deformation behavior from periodic variations and short-term disturbances. The decomposed components were subsequently used for deformation evolution analysis and driving-factor identification [15].

The identification of macroscopic deformation patterns was based on high-precision time-series deformation data derived from the SBAS-InSAR technique. All core processing procedures, including interferogram generation, phase filtering, phase unwrap-ping, and time-series deformation inversion, follow the algorithm proposed by Berardino et al. [13]. After correcting for atmospheric delays, topographic effects, and orbital errors, a 3 × 3 Gaussian filter was applied to the deformation-rate maps to suppress high-frequency noise and random interference. Taking the October 2018 failure event as the temporal boundary, cumulative deformation fields and annual average deformation-rate fields were generated for both the pre-failure and post-failure stages [16]. On the basis of these multi-temporal deformation products, a multi-scale statistical framework combining global-scale and subregional-scale statistics was established for spatial deformation evolution analysis, to quantitatively characterize the spatial heterogeneity and dynamic evolution of the deformation field [17].

To quantify spatial variations in deformation before and after failure, comparative analyses were conducted at both the global and subregional scales. This established a two-level analytical framework consisting of global and subregional analyses [15]. The spatial extent, deformation intensity, and heterogeneity of the deformation field were evaluated by comparing deformation characteristics between the pre-failure and post-failure stages. This framework provides a basis for identifying spatial response patterns associated with large-scale landslide failure and subsequent deformation evolution.

To quantitatively evaluate spatial deformation characteristics and their changes before and after failure, the following statistical indicators were calculated for each analysis unit:

(1) descriptive statistics (mean, standard deviation, and extrema of deformation);

(2) deformation change, defined as the difference between post-slide and pre-slide mean deformation ($\Delta \overline{D}$);

(3) uncertainty estimation of $\Delta \overline{D}$ using standard error propagation, followed by the error formula as follows:

$${\mathsf{\sigma }}_{\Delta \overline{\mathrm{D}}}=\sqrt{{\left({\displaystyle \frac{{\mathsf{\sigma }}_{pre}}{\sqrt{n_{pre}}}}\right)}^2+{\left({\displaystyle \frac{{\mathsf{\sigma }}_{post}}{\sqrt{n_{post}}}}\right)}^2}$$

(1)

Based on these metrics, spatial thematic maps of deformation changes are generated, with uncertainty incorporated as error bars. Subregions lacking pre-slide data are separately annotated to maintain analytical completeness.

Macroscale global statistics can obscure local heterogeneous responses; therefore, slope units are introduced as a vehicle for microscale analysis [18]. Based on a 30 m SRTM DEM, 128 slope units were generated using the r.slopeunits tool in GRASS GIS with a circular variance threshold of 0.3 and an area threshold of 0.03–0.1 km² [19]. Sensitivity tests were conducted to evaluate the robustness of the delineation parameters.

For each slope unit, mean deformation, extrema, standard deviation, and coefficient of variation were calculated to characterize deformation intensity and spatial variability. Frequency-distribution and stability analyses were subsequently performed to quantify changes in deformation patterns and spatial heterogeneity following landslide failure.

2.2.4. Analysis of Deformation Driving Mechanisms for the Baige Landslide Based on STL-PCA-ICA and Cross-Correlation Analysis

This section investigates the driving mechanisms of landslide deformation based on long-term deformation time-series data and STL decomposition results. An integrated analytical framework combining STL decomposition, PCA, ICA, and cross-correlation analysis was established to identify dominant environmental controls, quantify their driving intensities, and determine deformation response lags. By comparing the pre-failure and post-failure stages, the transformation of deformation-driving patterns and the intrinsic mechanisms controlling landslide evolution were systematically analyzed.

All deformation and environmental time series were first reconstructed onto a regular daily temporal grid using time-aware interpolation to address the irregular acquisition intervals of Sentinel-1 observations. STL decomposition was then applied to the temporally regularized time series to separate trend, seasonal, and residual components. The extracted signals were then combined with meteorological and freeze–thaw indicators for multivariate statistical analysis. PCA was employed to reduce data dimensionality and identify dominant groups of controlling factors, whereas ICA was used to separate statistically independent driving signals and improve the physical interpretability of the identified deformation controls. Finally, cross-correlation analysis was conducted to quantify the response lag and driving intensity between environmental factors and landslide deformation.

To achieve these objectives, an integrated analytical framework incorporating signal decomposition, factor correlation, and mechanism analysis was established. Based on the obtained daily temperature data, three freeze–thaw indicators were defined:

(1) Thaw degree-days (TDD), defined as the cumulative sum of daily mean temperatures above 0 °C:

$$TDD=\sum max\left(T_{mean},0\right)$$

(2)

(2) Freezing degree-days (FDD), defined as the cumulative sum of the absolute values of daily mean temperatures below 0 °C:

$$FDD=\sum \left|min\left(T_{mean},0\right)\right|$$

(3)

(3) Freeze–thaw cycle count, defined as the number of days when the daily mean maximum temperature > 0 °C while the daily minimum temperature < 0 °C.

The base temperature threshold was set to 0 °C, following standard practice. All in-dices were calculated at a daily resolution and then aggregated to monthly values for cross-correlation analysis. Secondly, since cumulative deformation, meteorological factors, and freeze–thaw indices are heterogeneous variables with different units and magnitudes, we applied Z-score standardization to eliminate scale effects and numerical biases among variables, thereby ensuring the comparability of subsequent multi-factor analyses [20].

Subsequently, PCA was applied to the standardized multivariate dataset for dimensionality reduction [21]. Key principle components were selected and extracted based on their contribution to variance, and the contribution of each environmental variable was analyzed using the factor loading matrix to rapidly identify the dominant clusters of environmental factors controlling landslide deformation. Considering that PCA based on orthogonal transformations tends to confound mixed signals from co-variant physical fields, making it difficult to separate independent driving components, ICA was further introduced to address this technical limitation [22].

In this study, the full dataset was divided into two independent subsets separated by the landslide collapse timing (pre-failure and post-failure), and ICA blind source de-composition was performed individually for each time-series subset; three statistically independent source signals denoted as IC1, IC2 and IC3 were extracted from each subset. ICA was used to blindly separate statistically independent physical source signal components from multi-source mixed observation signals, thereby enhancing the accuracy and physical plausibility of driving factor identification.

Finally, cross-correlation analysis was employed to quantitatively characterize the time-delay relationships between the driving factors and deformation responses [23]. To capture potential seasonal to monthly lag effects associated with freeze–thaw processes and cumulative precipitation, a relatively wide time lag window of −30 to +30 days was adopted. Within a wide time lag window of −30 to 30 days, the Pearson correlation coefficients (ρ) were calculated between the original deformation time series, the seasonal components derived from STL decomposition, and the independent components separated by ICA, and each environmental factor.

A unified criterion was established: ρ > 0.5 indicates a strong correlation, 0.3–0.5 indicates a moderate correlation, and ρ < 0.3 as weak correlation to judge the correlation magnitude. Based on this predefined evaluation system, independent components with remarkable indicative significance for driving effects were objectively screened for comparative analysis across pre- and post-failure stages. Combined the peak positions of the correlation coefficients to determine the optimal response lag and driving intensity of deformation to each driving factor. Ultimately, this systematically reveals the synergistic regulatory mechanism of meteorological elements and the freeze–thaw cycle on the deformation of the Baige landslide, as well as the reconstruction patterns of driving modes before and after instability.

To ensure consistency between the methodological design and graphical presentation, the full ±30-day cross-correlation results were explicitly computed and examined. This analysis confirms that all statistically significant correlation peaks (|ρ| ≥ 0.3) are concentrated within the 0–12 day lag window, whereas no secondary peaks are observed in the negative lag interval (−30 to 0 days) or the long positive lag interval (12–30 days). Accordingly, although a wide lag window was adopted to ensure comprehensive analysis, the effective deformation response is dominated by short-term lag behavior. The results are fully consistent with the complete ±30-day analytical framework.

2.2.5. Integrated Quantitative Analysis of the Spatiotemporal Evolution and Driving Mechanisms of the Baige Landslide

Based on SBAS-InSAR time-series inversion, slope-unit spatial statistics, STL–PCA–ICA, and cross-correlation analysis, an integrated quantitative framework of “time-series stage identification–spatial quantitative characterization–driving mechanism determination” was established to systematically investigate the spatiotemporal evolution and driving mechanisms of Baige landslide deformation.

For the quantification of spatiotemporal deformation evolution, the October 2018 failure event was used as the temporal boundary. Segmented linear regression was first applied to deformation sequences at typical monitoring points. Regression slopes were used to quantify deformation rates at different stages, while relative changes in the absolute slope values between pre-slide and post-slide stages were used to determine acceleration or deceleration trends. The coefficient of determination (R²) was used to evaluate the significance of deformation trends and identify deformation-rate inflection points before and after failure. Based on these results, a three-stage temporal evolution pathway of “slow deformation–residual creep–long-term acceleration” was established to characterize the evolutionary stages and activity intensity of landslide deformation.

Spatial quantitative analysis was further conducted using global–subregional polygon comparisons and slope-unit statistics. At the global scale, the severity and heterogeneity of deformation were quantified using changes in the number of valid pixels, mean deformation, and standard deviation. At the subregional scale, differences in deformation magnitude among subregions were compared to quantify the attenuation effects associated with topographic gradients. At the slope-unit scale, units were classified into three stability categories according to standard deviation: stable (<5 mm), moderately stable (5–10 mm), and unstable (>10 mm). Changes in the proportion of each category were then used to quantify destabilization intensity and spatial heterogeneity. By integrating temporal evolutionary stages with spatial differentiation characteristics, the spatiotemporal evolution mechanisms of the Baige landslide were systematically quantified.

For the analysis of deformation driving mechanisms, STL time-series decomposition was first used to separate trend and seasonal components from the deformation series. The deformation data and environmental variables, including meteorological and freeze–thaw factors, were standardized prior to analysis. PCA was then employed to identify dominant driver groups based on variance contributions and factor loadings, while ICA was used to separate independent physical driving signals and reduce interference from mixed signals. Cross-correlation analysis was subsequently applied to quantify the relationships between deformation and environmental factors using Pearson correlation coefficients, and optimal response lags were determined according to the peak lag positions. By comparing variations in correlation strength and lag characteristics before and after failure, changes in the contributions of different driving factors were identified. Ultimately, the transformation of the deformation driving pattern from “multi-factor synergistic short-lag driving” to “a temperature-dominated environmental control regimen with a longer response lag” was revealed, thereby clarifying the intrinsic dynamic mechanisms governing landslide deformation.

Through the integration of spatiotemporal quantification and multi-method driving-factor analysis, this section establishes a complete analytical framework from deformation characterization to mechanism interpretation and driving-pattern identification, providing methodological support for the evolution analysis and risk assessment of the Baige landslide.

3. Results

3.1. Surface Deformation Characteristics of the Baige Landslide Based on SBAS-InSAR

This study focuses on the Baige landslide as the core research subject. Leveraging SBAS-InSAR long-term interferometric measurement technology, we inverted and interpreted the cumulative surface deformation from 2017 to 2024, based on areal-averaged deformation values of all valid InSAR coherent pixels covering the entire landslide area to characterize the overall average deformation trend of the whole slope, with the results shown in Figure 3. Based on the time-series deformation results, and using the catastrophic overall instability of the landslide in October 2018 as a spatiotemporal boundary, the evolution of surface deformation was divided into two core phases: pre-failure steady-state incubation and precursor manifestation, and post-failure multi-stage dynamic adjustment. This comprehensively illustrates the overall evolution patterns of surface deformation before and after landslide instability and clarifies the macro-deformation characteristics of the landslide’s entire life cycle. To ensure consistency across different analytical scales, the stage classifications used in this study are organized within a unified hierarchical framework. The classification presented in this section represents the area-averaged deformation evolution of the entire landslide derived from all valid InSAR coherent pixels, whereas the monitoring-point classification presented in Section 3.2.1 describes local deformation behavior within the core sliding zone. These classifications are complementary representations of the same deformation process at different spatial scales rather than independent classification systems.

Prior to the failure (18 March 2017–3 October 2018), the landslide deformation followed a progressive pattern of steady-state oscillation followed by the sudden emergence of precursory signs. In 2017, the landslide was in a state of natural mechanical equilibrium, with cumulative deformation fluctuating within the range of −5 to +9.64 mm in an undirected, high-frequency pattern. Deformation rates alternated sharply between positive and negative values, and deformation signals were dominated by seasonal elastic disturbances and random noise, with no signs of sustained instability; In early 2018, the landslide entered the critical pre-instability phase, with deformation characteristics undergoing a fundamental shift. Following an anomaly reset, the deformation transitioned to a unidirectional, sustained negative LOS deformation trend. By October 2018, the cumulative deformation reached −17.99 mm, during this period, significant negative strain rate peaks emerged repeatedly, and deformation shifted completely from random fluctuations to irreversible accumulation, clearly reflecting the precursory characteristics of instability, such as internal stress imbalance and progressive structural failure within the slope.

Post-failure (2019–2024), the landslide departed from its original steady state and entered a phase of continuous post-disaster disturbance and adjustment. At the area-averaged scale, the post-failure evolution can be further subdivided into five successive stages, reflecting changes in the overall deformation intensity and spatial organization of the entire landslide. The evolution of deformation exhibited a clear five-stage progressive rhythm: 2019 marked the initial slow adjustment phase following the disaster, characterized by sustained negative LOS deformation, during which the slope completed the initial redistribution of post-failure stresses through gradual subsidence; In 2020, the deformation field reached its maximum spatial extent and intensity following the 2018 failure. However, the deformation pattern exhibited pronounced spatial heterogeneity. Negative deformation remained concentrated within the Baige landslide body, reflecting continued post-failure settlement, creep, and compaction of the displaced material. In contrast, positive deformation was observed in parts of the surrounding slope sectors outside the main landslide boundary. The large-scale collapse in October 2018 substantially altered the original stress field of the slope. As the displaced landslide mass underwent continued compaction and post-failure adjustment, stress redistribution within the surrounding rock mass may have induced localized compensatory deformation, resulting in positive LOS deformation signals in some peripheral slope sectors. In addition, deformation accommodation associated with the compaction of post-failure deposits may have contributed to spatially heterogeneous displacement patterns characterized by the coexistence of localized uplift and subsidence. However, no in situ monitoring data are currently available to directly verify the underlying mechanism, and therefore this interpretation should be regarded as a plausible inference based on the observed SBAS-InSAR deformation pattern; In 2021, the overall deformation intensity declined relative to the 2020 peak, with areal-averaged deformation fluctuating gently and the pattern reverting to negative LOS deformation, reflecting weakened large-scale post-failure structural adjustment at the whole-landslide scale. Nevertheless, deformation remained concentrated in the main sliding zone with persistent local cumulative deformation, indicating ongoing localized adjustment processes. In 2022, the areal-averaged deformation of the entire landslide showed relatively small fluctuations, suggesting a reduction in large-scale deformation activity. However, localized deformation persisted within active sectors of the landslide, indicating that post-failure adjustment processes were still ongoing; From 2023 to 2024, localized deformation activity continued to be observed in several sectors of the landslide. Although no evidence of large-scale renewed failure was detected, the slope remained dynamically active and underwent continued post-failure adjustment.

Overall, at the area-averaged scale, surface deformation of the Baige landslide from 2017 to 2024 followed a five-stage evolutionary sequence of “stable oscillation–precursory initiation–post-disaster adjustment–deformation peak–deformation attenuation–continued post-failure evolution”. The October 2018 instability event served as the pivotal turning point in the restructuring of the deformation pattern, rather than a long-term maintenance of a dynamically active state. Although the overall deformation intensity gradually decreased after 2020, localized deformation within the main sliding zone persisted throughout the monitoring period.

3.2. Spatial-Temporal Evolution Mechanisms of the Baige Landslide

Based on the deformation quantification results, this section adopts long-term time-series analysis of high-coherence monitoring points as the temporal foundation, and integrates refined slope-unit spatial statistics to systematically reveal the deformation evolution patterns and disturbance response mechanisms before and after the Baige landslide failure from both temporal and spatial dimensions.

3.2.1. Temporal Evolution Characteristics of Deformation in the Baige Landslide

To systematically analyze the temporal evolution patterns and phase transition mechanisms of the Baige landslide, three representative monitoring points (MP1–MP3) were selected in the core deformation zone based on high coherence, remarkable deformation magnitude and coverage of key slope movement areas. Monitoring points MP1–MP3 are located within the core fractured main sliding zone, and their time-series deformation reflects local slip-belt characteristics rather than the overall average deformation of the entire landslide. The deformation time-series of MP1 and MP3 are provided in the Supplementary Materials (Figures S1 and S2). Although differences exist in deformation magnitude and direction among the monitoring points, all three sites exhibit similar stage-transition characteristics before and after the 2018 failure event, indicating that the identified evolutionary pattern is spatially representative of the main deformation zone of the landslide. Therefore, MP2, which contains the most complete and continuous deformation record, was selected as the characteristic monitoring point for detailed temporal analysis.

Notably, the phase division at the point scale represents a simplified expression of the same deformation evolution process described in Section 3.1, derived from representative monitoring points within the core sliding zone. The three-stage framework presented here characterizes local deformation behavior within the core fractured sliding zone derived from representative monitoring points. This point-scale classification is hierarchically consistent with the five-stage area-averaged framework, with each stage reflecting the localized manifestation of the broader landslide evolution process across different spatial scales.

Using 2017–2024 time-series deformation data and taking the October 2018 catastrophic failure as the temporal boundary, the deformation curve of MP2 presents a three-stage evolutionary pattern of “slow pre-slide deformation–post-slide residual creep–long-term sustained acceleration” (Figure 4).

During 2017–2018, cumulative deformation fluctuated between 0–14 mm without a distinct directional trend, with an average monthly rate of −0.26 mm and a low R² of 0.20. The slope remained in weak initial creep near critical stability with no obvious acceleration precursors. From November 2018 to January 2021, deformation transformed into continuous unidirectional accumulation at an average monthly rate of +1.20 mm, dominated by stress redistribution, fragmented rock compaction and local plastic adjustment. According to the adopted LOS sign convention, positive deformation values indicate movement toward the satellite. Therefore, the positive rate observed during the post-slide residual creep stage reflects a temporary movement toward the satellite following the 2018 collapse.

After January 2021, the landslide entered a long-term acceleration stage, showing a stable and highly consistent linear trend in its time-series deformation. The average monthly deformation rate reached −2.12 mm, which was approximately 715% higher than that in the pre-slide stage, and the coefficient of determination (R²) was as high as 0.95, indicating a strong linear correlation between deformation and time. This significant acceleration and high fitting degree fully reflects the progressive deterioration of the internal structure of the slope rock mass and the continuous decline of its long-term stability. From a rock mechanics perspective, these features are diagnostic of tertiary creep: the STL-derived trend component accounts for 94% of the total variance, and the deformation rate jumped from +1.20 mm/month during the post-slide residual creep stage (movement toward the satellite) to −2.12 mm/month during the long-term acceleration stage. Under the adopted LOS sign convention, the negative deformation rate indicates renewed movement away from the satellite during the long-term acceleration stage. This evolution pattern is consistent with accelerated creep driven by stress corrosion and strain softening within fragmented rock masses under gravitational loading [24,25]. Although numerical modeling is not performed in this study, these statistical characteristics provide indirect but robust evidence for a gravity-dominated internal creep mechanism. In addition, the pronounced short-term fluctuations observed after 2020 indicate that seasonal deformation signals became increasingly evident during the post-failure adjustment process. Following the 2018 collapse, the fragmented landslide mass continued to undergo structural reorganization and residual creep, while seasonal environmental forcing, particularly temperature variations, exerted a stronger influence on slope deformation. As a result, the MP2 time series exhibits both a sustained acceleration trend and enhanced seasonal fluctuations.

The complete evolutionary sequence represents a conceptual synthesis of the multi-scale deformation characteristics identified in this study, rather than an independent classification system. It integrates the area-averaged five-stage evolution described in Section 3.1 and the point-scale three-stage behavior observed in Section 3.2.1, thereby demonstrating that the Baige landslide did not enter a stable dormant state after the major collapse but instead maintained persistent internal activity with continuous deformation development. These quantified temporal evolution characteristics provide reliable observational evidence and solid quantitative support for revealing the spatiotemporal coupling law of slope deformation and exploring its internal driving mechanism transition.

3.2.2. Spatial Evolution Characteristics of Deformation in the Baige Landslide

Based on 2017–2024 long-term deformation data from SBAS-InSAR inversion and spatial statistics of 128 slope units, this study quantitatively compares the spatial deformation differences in the Baige landslide before (2017–October 2018) and after failure (October 2018–2024), and reveals its spatial evolution features and control mechanisms. Landslide deformation is jointly dominated by topographic conditions, spatial location and pre-failure deformation state, showing obvious spatiotemporal heterogeneity. Relevant quantitative statistical results are listed in

Table 2. Statistical indicators of the Baige landslide.

Statistical Dimension	Indicator	Pre-Sliding (2017–October 2018.10)	Post-Sliding (November 2018–2024)	Change Analysis
Mean Deformation	Mean	−3.86	6.72	+117.1% (Reversal)
	Minimum	−18.84	−18.91	Stable
	Maximum	7.98	30.40	+281.0% (Elevated)
Extreme Deformation	Maximum	26.98	165.36	+512.9% (Intensified)
	Minimum	−47.18	−80.15	−69.9% (Enhanced)
	Extreme Difference	35.62	98.75	+177.2% (Diverged)
Dispersion	Standard Deviation	9.87	35.42	+258.9% (Expanded
	Coefficient of Variation	15.73	8.96	−43.0% (Concentrated)

Summarizes statistical indicators calculated from slope units, while pixel-based statistical results of the entire landslide and three subregions (A, B, C) are detailed in Tables S1–S4 of the Supplementary Materials.

.

To support the slope-unit-based spatial analysis, slope-unit delineation results are presented in Figure S3. Based on the 30 m SRTM DEM and the r.slopeunits algorithm, a total of 128 slope units were generated within the study area. The delineated units effectively capture local topographic boundaries and provide the basic spatial framework for subsequent deformation and stability analyses.

A sensitivity analysis was further conducted using circular variance thresholds (cvmin) of 0.25, 0.30, and 0.35 while keeping all other parameters unchanged. The results showed that all three parameter settings produced identical slope-unit configurations, yielding the same 128 slope units with unchanged spatial boundaries. Consequently, the derived deformation and stability statistics remained unaffected by the tested parameter variations, indicating that the selected delineation parameters are robust and suitable for spatial heterogeneity analysis in the Baige landslide area.

At the regional scale, 2766 effective deformation pixels were identified before sliding with an overall slight positive LOS deformation trend. After the landslide failure, the number of effective pixels dropped sharply to 560, and the deformation field evolved into pronounced negative LOS deformation, with the mean deformation value reversing from positive to negative and deformation amplitude greatly increased. The reduction in effective pixels is mainly attributed to the disintegration and fragmentation of frontal rock and soil masses, which increases surface roughness, weakens radar coherence and impairs SAR observation quality.

Subregional analysis adopting the whole landslide and three subregions (A, B, C) further reflects spatial deformation heterogeneity. Statistics based on all valid raster pixels within the whole-landslide polygon indicate that the overall standard deviation of deformation increased from 21.40 mm before failure to 40.55 mm after failure, accompanied by an expanded extreme value range, demonstrating aggravated spatial dispersion of post-failure deformation. These values are derived from pixel-based zonal statistics and therefore differ from the slope-unit-based statistics reported in Table 2. Detailed statistical results for the entire landslide and Subregions A, B and C are provided in Tables S1–S4 of the Supplementary Materials.

Slope-unit-based statistics further confirm the increase in spatial heterogeneity after failure. As shown in Table 2, the standard deviation of slope-unit mean deformation in-creased from 9.87 mm before failure to 35.42 mm after failure, accompanied by substantial increases in deformation extremes and dispersion. Although the absolute values differ from the pixel-based statistics because of the different statistical units adopted, both approaches consistently indicate a marked enhancement of post-failure spatial deformation heterogeneity.

Controlled by topography and location, the steep front Subregion A experienced early negative deformation and suffered the most intense destabilization with greatly amplified post-failure negative LOS deformation. In contrast, the gentle rear Subregion C maintained stable positive LOS deformation before failure and presented moderate deformation magnitude after failure, reflecting the buffering effect of gentle terrain on disturbance propagation. Localized positive LOS deformation in the front disintegration zone indicates that the downslope movement of the main sliding mass triggered regional stress release and compensatory adjustment.

To evaluate the robustness of the deformation interpretation, deformation statistics derived from ascending- and descending-orbit datasets were compared using the same slope-unit delineation framework. Although the absolute deformation magnitudes differ between the two viewing geometries, both datasets indicate that the 2018 catastrophic failure significantly altered the deformation characteristics of the Baige landslide. The observed differences are mainly attributed to variations in satellite line-of-sight geometry, terrain-induced distortion effects in the steep canyon environment, and differences in the spatial distribution of coherent pixels. Because the ascending-orbit dataset provides more continuous spatial coverage and a larger number of valid observations within the landslide area, it was selected as the primary dataset for detailed spatiotemporal analysis, while the descending-orbit results were used as an independent reference for evaluating the overall deformation evolution. Detailed statistical comparisons between the two datasets are provided in Table S5 of the Supplementary Materials.

In summary, the landslide failure induced systematic restructuring of the regional deformation field, including reduced valid pixels, transformation from slight positive LOS deformation to pronounced negative LOS deformation, and enhanced topographically controlled spatial heterogeneity. Supported by zonal statistics and slope-unit-based analyses, these spatial evolution characteristics deepen the understanding of landslide evolution mechanisms and provide reliable quantitative evidence for instability analysis and regional slope stability assessment.

3.3. Deformation Driving Mechanisms of the Baige Landslide

As the core research focus of this paper concerning environmental controls and deformation transition, this section aims to reveal the intrinsic deformation response characteristics and stage evolution law of driving mechanisms of the Baige landslide before and after catastrophic failure. Based on the time-series deformation data of typical monitoring point MP2, STL time-series decomposition was firstly adopted to separate multi-scale deformation components. Combined with PCA, ICA and cross-correlation analysis, this study further quantitatively explores the coupling relationship between landslide deformation and environmental factors, and systematically interprets the essential reason for the transition of deformation driving patterns.

The decomposed deformation signal can be divided into three independent components with clear physical implications: trend component, seasonal component and residual component (Figure 5). The trend component dominates the long-term evolutionary process of landslide deformation and reflects the continuous creep induced by internal rock mass structure and gravitational adjustment, with a goodness-of-fit as high as 94% against the original time series. Its slope presents obvious step-like changes, precisely marking the stability transition of the slope at different evolutionary stages. Under the adopted LOS sign convention, negative values indicate movement away from the satellite, whereas positive values indicate movement toward the satellite. The trend-component rate changed from −0.0076 mm/d in the pre-slide steady stage to +0.0322 mm/d in the post-slide residual creep stage, indicating a temporary reversal of LOS deformation direction following the 2018 collapse. Subsequently, the trend-component rate evolved to −0.0702 mm/d in the long-term accelerated deformation stage, reflecting renewed movement away from the satellite together with a substantial increase in deformation intensity. The seasonal component shows a stable annual periodic variation with an amplitude of approximately ±3 mm. Although the seasonal component reflects the influence of seasonal environmental variations, its periodicity is relatively weak and irregular. This is mainly because the deformation of the Baige landslide is dominated by long-term creep and post-failure structural adjustment, while seasonal climatic forcing contributes only a secondary modulation effect. The residual component accounts for 19% of the total signal, mainly reflecting short-term random disturbances caused by local micro-topography, sudden meteorological changes and observation noise without evident periodicity. The relatively weak periodicity of the seasonal component suggests that deformation at MP2 is not primarily controlled by regular seasonal forcing. Instead, long-term gravitational adjustment, structural degradation of the rock mass, and post-failure stress redistribution dominate the deformation process, thereby masking a clear seasonal signal.

Standardized deformation sequences and environmental factor datasets were further processed by PCA and ICA for dimensionality reduction and potential source signal separation (Figure 6). The results indicate that pre-failure deformation was governed by a multi-factor coupling mechanism. Temperature and TDD exhibit strong positive loadings on PC1, whereas freeze–thaw cycles show a negative loading, suggesting a coupled control pattern characterized by positive thermal forcing and negative freeze–thaw regulation. PC1 explains 64.93% of the total variance, indicating that the deformation response before failure was dominated by a relatively coherent environmental control system. After the large-scale collapse in 2018, the original deformation-environment response framework was fundamentally restructured. The contribution rate of PC1 decreased to 55.44%, while the influence of the second principal component PC2 increased significantly, indicating a remarkable change in the composition of dominant driving factors.

To further disentangle the physical sources of each principal component and isolate in-dependent driving signals, ICA was performed on the same standardized datasets, yielding three independent signals denoted as IC1, IC2 and IC3. Pre-IC1 represented short-term seasonal deformation dominated by precipitation and shallow freeze–thaw before landslide col-lapse, showing zero-lag strong correlation with displacement (ρ = 0.90); Pre-IC3 was governed by accumulated temperature and annual freeze–thaw cycles and reflected long-term thermal creep of deep bedrock; Pre-IC2 exhibited weak correlations with all considered environmental variables (|ρ| < 0.5) and did not show a clear physical association with the environmental factors analyzed in this study. After the catastrophic collapse in October 2018, physical properties of each component underwent obvious conversion: Post-IC1 inherited the long-period thermal features of Pre-IC3 and evolved into a lagged temperature-freeze–thaw controlled component; Post-IC3 took over the deformation characteristics of Pre-IC1 and denoted self-weight creep of post-failure deposits; Post-IC2 showed a transient response that was temporally associated with intensive rainfall events. Therefore, Pre-IC1, Pre-IC3 and their corresponding post-collapse counterparts were selected for subsequent cross-correlation analysis, whereas IC2 was not further considered in the environmental driving-mechanism interpretation because it lacked a clear association with the analyzed environmental variables.

It should be noted that the pre-failure dataset spans approximately 18 months, resulting in a relatively limited sample size for ICA decomposition. Therefore, the extracted independent components should be interpreted as exploratory representations of the dominant deformation signals rather than definitive physical processes. Moreover, the interpretation of deformation-driving mechanisms in this study is not based solely on ICA results but is jointly supported by STL decomposition, PCA, and subsequent cross-correlation analyses. The consistency among these independent analytical approaches provides qualitative support for the inferred deformation-driving mechanisms. Nevertheless, uncertainty remains regarding the statistical robustness of the extracted components, and the ICA-derived signals should therefore be interpreted with appropriate caution. Future studies incorporating longer observation records and independent validation approaches, such as bootstrap stability testing, may further improve the assessment of component robustness and reproducibility.

Cross-correlation analysis (Figure 7) further verifies that pre-slide deformation formed a highly coordinated short-lag response network with environmental factors such as air temperature, precipitation and freeze–thaw processes, with lag response time mainly concentrated within 0–5 days. Specifically, the seasonal deformation component exhibited strong synchronous correlation with precipitation (ρ = 0.57) and a prominent 5-day lag correlation with regional freeze–thaw activities (ρ = 0.58). By contrast, post-failure deformation gradually evolved into a temperature-dominated environmental control regime. The response lag between temperature and deformation extended to 12 days with a higher correlation coefficient (ρ = 0.64), whereas the correlation intensity of deformation with precipitation and freeze–thaw cycles weakened obviously and fell below the critical level (ρ < 0.5). It is worth noting that although the cross-correlation analysis was performed within a ±30-day window, all significant correlation peaks (|ρ| ≥ 0.3) occurred within 0–12 days, and no secondary peaks were observed in the ranges of −30 to −12 days or +12 to +30 days (|ρ| < 0.3). This confirms that the deformation response to environmental forcing is predominantly short-lag, and the choice of a wider window does not affect the reliability of the conclusions. To ensure methodological consistency with the analytical framework defined in Section 2.2.4, the full ±30-day cross-correlation computation was used in all analyses presented in Figure 7. This ensures that the graphical results are fully comparable with the methodological design, while the observed response behavior is interpreted based on statistically significant correlation peaks.

The observed transition of the deformation driving mechanism may be associated with internal structural reorganization of the slope following the catastrophic failure. The large-scale collapse likely resulted in substantial fragmentation and redistribution of the slope materials, which may have altered the original hydrogeological conditions and modified rainfall infiltration and subsurface runoff pathways. These changes may have reduced the sensitivity of slope deformation to precipitation and freeze–thaw processes observed before failure. Accordingly, the deformation-control pattern appears to have shifted from a pre-failure regime characterized by the coupled influence of multiple environmental factors to a post-failure regime with stronger thermal influence and weakened responses to precipitation and freeze–thaw activities.

Although the available deformation observations and statistical analyses do not provide direct evidence of internal structural reconstruction, the combined results of STL decomposition, PCA, ICA and cross-correlation analyses consistently indicate a substantial reorganization of deformation–environment interactions after the 2018 failure event. These findings provide quantitative evidence for understanding the stage-dependent evolution of environmental controls on high-altitude fractured-rock landslides in tectonically active mountain regions, and may contribute to long-term deformation assessment, hazard monitoring and risk management of similar landslides in the upper Jinsha River region.

The fundamental transition of the deformation driving mechanism is predominantly controlled by the drastic internal structural reconstruction of the slope after failure. The intact and integrated bedrock was completely disintegrated into loose, fragmented and disordered rock–soil masses, which greatly changed the original hydrogeological conditions, disrupted rainfall infiltration and underground runoff pathways, and substantially weakened the cumulative fatigue damage effect induced by long-term freeze–thaw cycles. Accordingly, the dominant control mechanism of landslide deformation has undergone a clear shift: from the synergistic regulation of multiple external environmental factors under intact rock mass structure before failure, to the internal self-adjustment dominated by slope structural reconstruction, stress redistribution and thermodynamic evolution after failure. This study quantitatively clarifies the stage-dependent transition law of environmental control and deformation driving mechanism for high-altitude fractured rock landslides in tectonically active mountainous areas. It also provides a reliable theoretical basis and quantitative reference for the long-term evolution assessment, instability early warning and disaster chain risk prevention of similar landslides in the upper reaches of the Jinsha River.

4. Discussion

4.1. Objective Limitations of InSAR Deformation Observations

Based on Sentinel-1 time-series imagery and SBAS-InSAR technology, this study obtained a long-term surface deformation dataset covering 2017–2024. This enabled continuous monitoring of the complete evolution of the Baige landslide before, during, and after failure, providing key data support for analyzing its full-cycle evolution patterns and driving mechanisms. However, due to the combined effects of complex high-mountain canyon topography, severe geological disturbance during failure, and limitations in satellite acquisition capability, the InSAR deformation observations still contain several objective limitations.

First, significant decorrelation occurred in the main sliding zone during the catastrophic failure stage. Liu et al. reported that the rapid and large-scale deformation of the Baige landslide on 11 October 2018 caused severe SAR decorrelation, making deformation retrieval during this period impossible [5]. Studies by Yang C. S. et al. and Yang F. et al. further confirmed this limitation [8,9]. The former adopted multiple remote sensing methods for complementary analysis, whereas the latter combined InSAR phase and amplitude information to compensate for the limitations of single-technique monitoring. Both studies demonstrated that InSAR alone cannot fully capture the complete deformation process during catastrophic failure.

Furthermore, the observational characteristics of Sentinel-1 data and the complex environmental conditions of the study area further increased data uncertainty. Sentinel-1A imagery has a spatial resolution of approximately 40 m, and SAR observations can only retrieve one-dimensional deformation along the radar line of sight. When ground movement is nearly perpendicular to the satellite viewing direction, deformation signals are significantly weakened. The Baige landslide area also exhibits an elevation difference exceeding 800 m, resulting in severe geometric distortions such as layover and shadowing. Combined with dense vegetation cover, seasonal changes introduce strong temporal decorrelation effects, while atmospheric delays in mountainous regions further complicate deformation signal extraction [26]. It is worth noting that all deformation results in this study are LOS measurements and have not been decomposed into a three-dimensional displacement field. This may lead to underestimation of true surface displacements. However, because the Baige landslide maintained a relatively stable sliding direction after failure, and our main conclusions rely on relative changes, trend consistency, and variations in correlation strength rather than absolute LOS values, the systematic LOS projection bias is consistent across stages and does not affect the identification of acceleration trends or driver transitions. Similar LOS-only analyses have been successfully applied in other high-altitude landslide studies [27,28]. In addition, the observed deformation concentration in the main sliding zone and the long-term post-failure de-formation trend are consistent with previously reported monitoring and remote-sensing studies of the Baige landslide [3,5,8,9], which further supports the reliability of the SBAS-InSAR results. Future work will combine multi-orbit InSAR and GNSS for full 3D displacement inversion.

Specifically, a coherence threshold of 0.3 was adopted during phase unwrapping to exclude low-coherence pixels and reduce error propagation. Although the selected threshold inevitably affects the number of retained pixels, it is widely regarded as a reasonable compromise between phase reliability and spatial coverage in mountainous SBAS-InSAR applications. The pronounced reduction in effective deformation pixels after the 2018 failure event is mainly attributed to severe decorrelation caused by slope frag-mentation, large displacement gradients, and substantial changes in surface scattering characteristics following catastrophic slope collapse. Therefore, the observed decrease in valid pixels primarily reflects the physical evolution of the landslide rather than being solely controlled by the selected coherence threshold.

In addition to the inherent limitations of InSAR technology and environmental noise interference, the continuity of the later-stage data in this study was also constrained by satellite image acquisition capabilities. From 2022 to 2024, only Sentinel-1A ascending-track images were available for the study area, with only 13 valid scenes in 2022, 8 in 2023, and 12 in 2024. The low image availability led to extended temporal sampling intervals, which became an important factor reducing the continuity of late-stage deformation observations.

These limitations impose two main constraints on the present analysis. First, the absence of deformation information during the catastrophic failure stage restricts quantitative characterization of the critical transition from accelerated deformation to complete failure. Therefore, interpretation of this stage mainly relies on optical remote sensing observations and deformation reconstructions reported in previous studies. Second, the limited image availability and extended temporal intervals between 2022 and 2024, combined with post-failure coherence loss in some areas, resulted in sparse observations during the slow deformation stage, affecting the continuous estimation of cumulative deformation and deformation rates. Nevertheless, these data gaps are mainly concentrated in short periods of intense activity and years with limited image coverage. Their influence on the overall deformation evolution pattern, including pre-slide slow deformation, early accelerated deformation, and post-failure continuous creep, remains limited. The deformation time series constructed in this study still effectively captures the phased evolution characteristics of the Baige landslide.

To address these limitations, this study primarily relies on valid InSAR observations and supplements incoherent or missing periods using optical remote sensing imagery, published studies, and regional multi-source monitoring data for cross-validation. This integrated approach minimizes the limitations of single-source InSAR observations and improves the completeness and reliability of the analysis of the full life-cycle deformation evolution of the Baige landslide.

4.2. Comparative Analysis of Spatiotemporal Deformation Evolution Characteristics with Previous Studies

Based on long-term SBAS-InSAR deformation data, this study identifies a three-stage evolution pattern of the Baige landslide, namely “pre-slide slow deformation–post-slide residual creep–long-term sustained acceleration,” together with spatial differentiation jointly controlled by topography and spatial location. Compared with previous studies, the results not only confirm the established evolutionary framework but also provide refinements in stage division, quantitative characterization, and mechanistic interpretation.

The overall deformation evolution characteristics are generally consistent with previous studies. Temporally, the framework of “pre-slide slow deformation–post-slide sustained deformation” agrees with the long-term evolution models proposed by Xu et al. and Liu et al. [4,5], confirming that the Baige landslide underwent prolonged pre-failure evolution and continuous post-failure activity rather than sudden catastrophic failure. Spatially, the blockage–main sliding–traction zoning proposed by Deng et al. [6] and the K1–K5 residual mass classification of Liu et al. [5] correspond closely to the front-strong and rear-weak deformation gradient pattern identified in this study, further confirming the dominant control of topography on landslide deformation.

The main difference from previous studies lies in the interpretation of post-slide deformation evolution. Liu et al. regarded the post-slide stage as a slow but continuous creep process (post-sliding slow deformation) without further sub-dividing it into distinct temporal phases [5]. In contrast, our study reveals that the post-slide deformation rate increased from +1.20 mm/month to −2.12 mm/month, indicating that post-slide deformation did not gradually decay but instead evolved into a sustained acceleration process. This finding refines the conventional understanding of post-failure deformation in the Baige landslide by revealing a stepwise rate increase hidden within what was previously described as a uniform slow creep.

Spatially, previous studies mainly relied on qualitative geological zoning, whereas this study quantitatively characterizes deformation heterogeneity using slope-unit statistics. Subregion A exhibited significantly stronger post-slide deformation fluctuations than the regional average, while Subregion C showed relatively stable deformation responses, quantitatively demonstrating the attenuation and regulatory effects of topography on deformation disturbances. In addition, the sharp decrease in effective observation pixels and the increase in deformation standard deviation after failure are consistent with the staged instability process proposed by Deng et al. [6]. Areas lacking effective pre-slide deformation signals but showing post-slide positive LOS deformation also correspond spatially with the discontinuous slip zones identified by Cao et al. [2], providing additional evidence for the multi-stage collapse and sliding mechanism.

Overall, this study confirms the general evolutionary framework proposed in previous research while refining the understanding of post-slide deformation through quantitative stage division and slope-unit-based spatial analysis. Importantly, the apparent difference between the decreasing area-averaged deformation trend reported in Section 3.1 and the sustained acceleration observed at MP2 in Section 3.2.1 does not represent a contradiction. The former reflects the averaged response of the entire landslide system, including stabilized and slowly deforming peripheral zones, whereas MP2 captures the localized behavior of the main active sliding belt where deformation remains concentrated and continuously evolving. This indicates that post-failure deformation is spatially heterogeneous, with overall attenuation at the landslide scale coexisting with persistent or accelerating deformation within key active zones. These findings improve the understanding of the full life-cycle evolution and deformation mechanisms of the Baige landslide.

4.3. Comparative Analysis of Deformation Driving Mechanisms with Previous Studies

This study integrates STL time-series decomposition, principal component analysis, cross-correlation analysis, and independent component analysis to systematically reveal the transformation of deformation-driving mechanisms before and after failure of the Baige landslide. Before failure, deformation was jointly controlled by multiple environmental factors, including air temperature, precipitation, and freeze–thaw cycles, exhibiting high environmental sensitivity and short response lags. After failure, the slope structure became highly fragmented, the driving system contracted significantly, and deformation evolved into a temperature-dominated environmental control regime with a longer response lag, while gravity-driven creep and post-failure structural adjustment remained important back-ground controls. Overall, these findings are generally consistent with previous studies at the macroscopic level but provide further quantitative insights into environmental driving factors, lag-response characteristics, and driving-mechanism transformation.

At the macroscopic level, the results agree well with previous research. Gravity remains the dominant endogenous driving force controlling landslide evolution. Deng et al. identified gravitational loading as the primary factor triggering instability [6], while Cao et al. emphasized the combined effects of tectonic fragmentation, gravitational unloading, and river erosion [2]. Consistent with these studies, the present analysis shows that the trend-component deformation rate increased from −0.0076 mm/d before failure to −0.0702 mm/d during the long-term acceleration stage, quantitatively confirming the dominant role of gravity-driven deformation. Previous studies also recognized staged evolution characteristics. For example, the creep–shear–collapse model proposed by Feng et al. [7] and the stability evolution patterns simulated by Liu et al. [5] both indirectly reflect changes in deformation responses and driving conditions during different stages. Building on these studies, the present work provides a quantitative interpretation of the staged evolution of environmental driving mechanisms.

The main differences from previous studies lie in the detailed quantitative analysis of driving mechanisms. First, this study clarifies the transition from multi-factor environmental regulation before failure to a more concentrated temperature-dominated environmental control regime after failure. Previous studies mainly focused on geological conditions such as rock fragmentation and river erosion or evaluated slope stability through numerical simulations, without quantitatively identifying environmental driving factors. In contrast, the PCA and correlation analyses in this study demonstrate that pre-slide deformation was jointly controlled by air temperature and freeze–thaw processes, with the first principal component explaining 64.93% of the variance. After failure, correlations between deformation and precipitation or freeze–thaw processes weakened significantly, while air temperature exhibited the strongest statistical association with deformation among the analyzed environmental variables. The transition from intact bedrock to loose debris likely altered infiltration pathways and freeze–thaw processes, which may have contributed to the observed simplification of environmental response patterns.

Second, this study quantitatively identifies stage-dependent lag responses between environmental factors and deformation, addressing a gap in previous research. Before failure, deformation responded rapidly to environmental forcing, with precipitation showing strong synchronous correlations and freeze–thaw processes exhibiting a strong correlation with a short lag of approximately 5 days. After failure, the fragmented debris structure reduced heat-transfer and stress-transmission efficiency, extending the deformation response lag to air temperature to approximately 12 days. Previous studies only qualitatively discussed seasonal influences and did not perform systematic cross-stage lag comparisons or establish quantitative lag relationships between deformation and multiple environmental factors [5]. Through long-term statistical analysis, this study quantifies both lag times and response intensities at different evolutionary stages, refining the understanding of time-dependent environmental controls on high-altitude landslides.

Third, this study explains the intrinsic mechanism responsible for the continued acceleration of post-slide deformation. Previous research mainly focused on post-failure stability evaluation and hazard assessment but did not explain why deformation continued to accelerate rather than gradually decay. In exploring its causes, we examined two potential external triggers: a lowered rainfall threshold and fluvial undercutting at the toe. Cross-correlation analysis (Figure 7) shows that the post-slide deformation–precipitation correlation dropped from a strong ρ = 0.57 (pre-slide) to a weak ρ < 0.3 (post-slide), contradicting the hypothesized threshold lowering and indicating a weakened role of rainfall. Meanwhile, Figure 1b confirms that the landslide debris blocked the Jinsha River and displaced the main channel away from the toe, so sustained fluvial undercutting would not occur. Thus, the contribution of external environmental factors has significantly decreased and cannot explain the post-2021 sustained acceleration. The time-series fitting results show that the landslide evolved from a residual creep stage into a long-term acceleration stage, with deformation trend rates increasing stepwise over time. The trend component achieved a goodness-of-fit of 94%, indicating that the slope evolved from short-term post-failure stress adjustment into a continuously active state primarily associated with continued gravitational loading, internal creep, and post-failure structural adjustment within fragmented debris masses. This finding provides quantitative support for the mechanism of persistent post-failure deformation and supplements previous risk assessments.

From an operational monitoring perspective, the identified 12-day lagged response between temperature and landslide deformation provides a potentially useful forecasting window. Although the present study does not establish quantitative warning thresholds, sustained temperature increases may serve as a precursor signal for enhanced landslide activity. Therefore, periods of rapid temperature rise could be used to trigger intensified deformation monitoring during the subsequent two weeks. Integrating temperature forecasts with near-real-time InSAR observations and ground-based monitoring may improve the timeliness of hazard identification and support the development of site-specific early-warning strategies for large high-altitude landslides. However, the establishment of quantitative warning thresholds require long-term multi-source monitoring data and remain an important direction for future research. It should be noted that the meteorological forcing data represent regional-scale climatic variability rather than site-specific microclimatic conditions, which may introduce scale-related uncertainties in the quantitative interpretation of environmental controls.

Overall, this study confirms the gravity-dominated and stage-dependent driving framework proposed in previous research while further quantifying the transition from multi-factor environmental regulation before failure to a temperature-dominated environmental control regime after failure, with gravity-driven creep and post-failure structural adjustment remaining important background controls. Importantly, the cross-correlation results consistently show short-lag responses within 0–12 days across all driving factors, ensuring the temporal coherence of the identified mechanism. It also reveals the reduced statistical influence of precipitation and freeze–thaw processes on post-failure deformation and clarifies the evolution of deformation-response lags. These findings advance the understanding of the Baige landslide from qualitative interpretation to quantitative mechanism analysis and provide references for the monitoring and early warning of similar high-altitude landslides.

5. Conclusions

This study takes the Baige landslide as its subject. Based on Sentinel-1 time-series imagery from 2017 to 2024, and utilizing SBAS-InSAR technology, time-series analysis of feature points, and spatial statistical methods for slope units, we systematically investigated the surface deformation characteristics, spatiotemporal evolution mechanisms, and deformation driving mechanisms of the landslide. We elucidated the deformation response patterns and underlying mechanisms throughout the landslide’s entire life cycle. The core conclusions are as follows:

(1) Based on SBAS-InSAR monitoring results, the fundamental deformation characteristics of the Baige landslide were systematically clarified. Using the catastrophic failure in October 2018 as the key temporal boundary, a multi-scale deformation interpretation framework was established. At the area-averaged scale, the deformation evolution is characterized by a five-stage sequence describing the overall landslide response. At the point scale, a three-stage behavioral pattern is identified from representative monitoring points within the core sliding zone. In addition, a seven-node descriptive sequence is introduced as a conceptual synthesis to summarize the full evolutionary process of landslide deformation, integrating observations from different spatial scales rather than representing an independent classification system. Deformation is concentrated in the main slip zone, and the residual mass continues to be activated. The high-mountain canyon topography and fractured rock mass provide inherent geological conditions for the landslide’s long-term deformation. Overall, the Baige landslide exhibits a clear scale-dependent and stage-structured deformation evolution pattern.

(2) The study revealed the temporal evolution of deformation in the Baige landslide. Taking the catastrophic failure in October 2018 as a key dividing point, the landslide exhibited three distinct phases: “slow pre-slide deformation–post-slide residual creep–long-term sustained acceleration.” Pre-slide (March 2017–October 2018): The monthly average deformation rate was only −0.26 mm (R² = 0.20), with significant deformation fluctuations and no evidence of directional accumulation; In the post-slide residual creep phase (November 2018–January 2021), the rate increased to +1.20 mm/month, marking the onset of stress redistribution and rock mass consolidation; After 2021, the system entered a long-term acceleration phase, with the rate surging to −2.12 mm/month, approximately 715% increase compared to the pre-slide period (R² = 0.95)—marking a transition from passive stress adjustment to sustained active activation.

(3) The spatial differentiation and evolution characteristics of deformation were clarified. The 2018 landslide event caused a systematic restructuring of the regional deformation field. The number of effective deformation pixels decreased sharply from 2766 before failure to 560 after failure, accompanied by obvious deterioration of radar coherence. The mean deformation value reversed from positive LOS deformation to pronounced negative LOS deformation, with the overall deformation amplitude increasing by 117.1%. The global standard deviation of deformation rose from 21.40 mm to 40.55 mm, indicating significantly intensified spatial heterogeneity. Controlled by topography and location, deformation disturbances were most severe in the steep sections at the front edge, while the response was mild in the gentle sections at the rear edge. Localized positive LOS deformation signals in the frontal disintegration zone suggest regional stress release and compensatory adjustment during post-failure deformation evolution.

(4) The transformation mechanism of deformation drivers before and after the Baige landslide’s instability was elucidated. Prior to the slide, the slope structure was intact, and deformation was jointly regulated by multiple environmental factors such as air temperature, precipitation, and freeze–thaw cycles, with environmental responses exhibiting a strong correlation with a short time lag of 0–5 days. After destabilization, the rock mass became fragmented and loose, and the driving system was significantly simplified, shifting to a temperature-dominated environmental control regime with a long time lag, extending the response lag to 12 days, while the driving associations with precipitation and freeze–thaw cycles were significantly weakened. The structural failure of the slope body is the core trigger for the transformation of the driving mechanism. Combined with the continuous action of gravity and the creep effects within the debris mass, these factors jointly govern the long-term accelerated deformation dynamics of the post-slide residual mass.