Content area
With the continuous promotion of railway construction in China, railway lines are increasingly extended to areas with complex geological environment, and such areas are prone to landslides and other geological disasters, which seriously threaten the safety of railway operation. The current landslide susceptibility assessment along the railway line relies on static factors such as topography and geology, and fails to take into account the significant time-varying and sudden nature of landslide disasters in complex geological environments, This poses a challenge in terms of satisfying the actual demand for dynamic perception of landslide hazards, and to reflect the deformation characteristics of potential landslides. For this reason, this paper utilizes to introduce the Interferometric Synthetic Aperture Radar (InSAR) technique to dynamically extract the surface deformation characteristics, as an effective supplement to the existing static factors, to enhance the promptness and precision of landslide susceptibility evaluation. Firstly, INSAR was used to obtain surface deformation in the study area and combined with optical remote sensing to identify landslides. Secondly, the deformation rate was taken as a dynamic factor, and 12 static factors, such as elevation and rainfall, were combined to construct a Mean Particle Swarm Optimisation -Random Forest (MPSO-RF) model, and the dynamic factors were introduced into the model through joint training and weighted superposition and performed. accuracy comparison and landslide susceptibility evaluation. Finally, the causes of landslides were analysed by combining the results of INSAR identification and model evaluation. The results show that: (1) the Small Baseline Subset Interferometric Synthetic Aperture Radar (SBAS-InSAR) technique can effectively identify potential landslide areas in slow deformation; (2) the accuracy of the joint training and weighted superposition models is improved by 6.54% and 3%, respectively, compared with that of the static model subsequent to the introduction of the INSAR deformation data; (3) the joint evaluation of the SBAS-InSAR and the MPSO-RF model can effectively supplement the traditional static evaluation with the lack of dynamic information. evaluation with the lack of dynamic information. The results of the study can provide theoretical basis and methodological support for the construction of line safety environment platform in railway disaster prevention and monitoring system.
Introduction
In recent years, landslides along railway lines have occurred frequently, posing a serious threat to normal railway operations. For example, on 19 June 2022, a landslide occurred in Longhu Village, Dajiangkou Town, Xiupo County, Huaihua City, Hunan Province, causing the northbound line of the Shanghai-Kunming Railway to suspend operations; on 15 September 2022, a mountain landslide occurred in Chengbei District, Xining City, Qinghai Province, with part of the landslide mass penetrating the bridge structure, causing the bridge piers of the Lanzhou-Xinjiang High-Speed Railway to shift, resulting in train suspensions and significant economic losses. These incidents highlight the severe situation of landslide disasters along railway lines. To ensure railway operational safety, China is accelerating the construction of an active protection system for railway operational environments, primarily comprising the railway comprehensive video surveillance system and the natural disaster and foreign object intrusion monitoring system [1]. Although some progress has been made, the railway safety environment management platform within the natural disaster monitoring system remains in the experimental or preliminary application phase, and existing monitoring methods still require optimisation in terms of timeliness and accuracy [2] Therefore, conducting landslide identification and susceptibility assessment along railway lines is of great significance for enhancing the capabilities of natural disaster monitoring systems and ensuring railway operational safety.
Current evaluations of landslide susceptibility along railway lines primarily rely on static factors such as topography, geology, and hydrology, without considering the dynamic evolution characteristics of landslides, leading to discrepancies between evaluation results and actual conditions [3]. In recent years, Interferometric Synthetic Aperture Radar (InSAR) technology, which enables large-scale, high-precision monitoring of ground deformation, has been gradually incorporated into landslide susceptibility evaluation systems. The InSAR technology system encompasses Differential InSAR (D-InSAR), Multi-Temporal InSAR (MTInSAR), Permanent Scatterer InSAR (PS-InSAR), and Small Baseline Subset InSAR (SBAS), enabling high-precision, long-term monitoring of large-scale surface deformation [4]. Compared to traditional landslide susceptibility assessment methods based on static factors, InSAR technology enables real-time monitoring of the slow deformation phase preceding landslides. This allows for the timely detection of early warning signs, thereby enhancing the accuracy and reliability of dynamic monitoring. It addresses the shortcomings of traditional methods, which cannot effectively track the evolution of landslides. For instance, Jia et al. [5] updated landslide susceptibility mapping using SBAS-INSAR technology, identifying additional potential landslide areas with an evaluation accuracy of 0.95. Ghaderpour et al. [6]combined PS-INSAR data with precipitation data to investigate the relationship between surface deformation and precipitation patterns. The results indicate that precipitation has a triggering effect on surface deformation.Tao et al. [7] proposed a spatio-temporal processing method for PS-INSAR to filter noise, achieving favorable results in settlement monitoring along railway lines. These studies validate the feasibility and effectiveness of INSAR technology for landslide monitoring.
Given the complex and variable geological environment and the prominent non-linear characteristics of landslide mechanisms, selecting an appropriate evaluation model is key to achieving high-precision landslide susceptibility assessment [8]. Machine learning models, which do not require extensive prior knowledge, possess strong data processing capabilities, and offer high predictive accuracy, have been extensively utilized in landslide susceptibility studies. Commonly used models include logistic regression, neural networks, and random forests [9]. However, a single machine learning model has limitations when dealing with complex terrain and multiple factors. Introducing optimisation algorithms can effectively improve model performance by adjusting parameters or structures to more efficiently process complex multidimensional data [10]. Yang Can [11] et al. utilised Bayesian optimisation algorithms to optimise the hyperparameters of machine learning models, which make landslide susceptibility estimates much more accurate. Razavi-Termeh [12] et al. employed three physics-based meta-heuristic algorithms to optimise random forests and extreme gradient trees, demonstrating improved accuracy in landslide susceptibility mapping. Although various optimization methods have been applied to landslide susceptibility assessment, model optimization remains a key factor in enhancing evaluation accuracy and adapting to complex environments.
In summary, most existing studies rely solely on single InSAR techniques or isolated machine learning methods for landslide susceptibility assessment, failing to effectively monitor landslide evolution processes and resulting in certain biases in evaluation outcomes. Furthermore, few studies integrate machine learning with InSAR technology to establish a comprehensive evaluation system incorporating both static and dynamic factors.To this end, this paper proposes a method for evaluating landslide susceptibility along railway lines by integrating SBAS-InSAR (Small Baseline Subset Interferometric Synthetic Aperture Radar, SBAS-INSAR) technology with a two-layer optimised machine learning model. The main contributions of this paper are as follows: (1) By integrating SBAS-INSAR technology with optical remote sensing data, an effective method for identifying potential landslides is proposed, enhancing monitoring accuracy and coverage. (2) SBAS-INSAR technology is innovatively combined with a random forest model optimized by the mean particle swarm algorithm, overcoming the limitations of traditional methods in dynamic monitoring and accuracy improvement. This not only enhances the model’s predictive capability for landslide susceptibility but also improves monitoring timeliness and accuracy. (3) Incorporating surface deformation as a dynamic factor alongside static factors into the landslide susceptibility evaluation system through joint training and weighted superposition methods. This effectively enhances the rationality and comprehensiveness of landslide susceptibility assessments, yielding more precise and reliable final evaluation results. The structure of this paper is as follows: First, based on the actual conditions of the study area, SBAS-INSAR technology is employed to get information about surface movement, and spatio-temporal evolution characteristics are analysed to identify potential landslides. Second, the mean particle swarm optimisation algorithm is used to tune the parameters of the random forest model, constructing a two-layer optimised MPSO-RF model, and dynamic factors are introduced through two methods: joint training and weighted superposition. Then, the accuracy of the model using only static factors was compared with that of the model incorporating dynamic factors, and the performance differences in landslide susceptibility zoning were analysed. Finally, combining the evaluation results of SBAS-INSA and MPSO-RF, a systematic analysis of the causes of landslides along the railway line in the study area was conducted, aiming to provide theoretical basis and methodological references for enhancing the geological disaster monitoring capabilities and scientific prevention and control levels of railway lines.
Materials and methods
Study region
Ya’an City(Fig. 1) is situated in the middle region of Sichuan Province, covering an area of approximately 15,300 square kilometres. It is situated at the junction of the western Sichuan mountainous region and the Sichuan Basin, with a terrain that slopes from west to east, reaching a maximum elevation of 4,884 m. The region is traversed by rivers such as the Qingyi River and the Zhou Gong River, and it has a subtropical humid monsoon climate. Annual average precipitation ranges from 1,200 to 1,800 millimetres, with rainfall concentrated between June and September. The railway line experiences heavy rainfall, which is a significant factor in triggering geological disasters. The region is situated at the intersection of the Longmen Mountain Fault and the Wolong-Luzhou Fault Zone, characterised by frequent tectonic activity, intense terrain erosion, and complex geological conditions, resulting in frequent landslides, mudslides, and other geological hazards.
[See PDF for image]
Fig. 1
Overview of the study area
The existing railway in Ya’an City is the Ya’an section of the Chengdu-Ya’an Railway, with a total length of 41.5 km. It runs east-west, traversing mountainous gorges and narrow river valleys, with high geological hazard risks along the route and significant landslide hazards. The railway under construction within the city is the Ya’an to Linzhi section of the Sichuan-Tibet Railway, which spans Tianquan County, Yucheng District, and Lushun County. The route crosses multiple active fault zones, including the Anning River-Zemu River, Longmen Mountain, and Jinsha River fault zones, with the terrain primarily consisting of high mountain gorges. Historical landslide disaster sites are densely distributed, and the geological environment is extremely complex. The construction of the railway line and disaster prevention efforts face significant challenges, necessitating enhanced monitoring and identification of landslides and other geological hazards.
Datasets
Based on existing research [13]- [14] and in conjunction with the current situation of the study area, static representative factors were selected from four major categories: topography and landforms, natural factors, geological factors, and human engineering activities. Elevation, slope gradient, slope aspect, planar curvature (Plcur), and Profile curvature (Procur) were selected as topographical factors; natural factors included precipitation, distance from rivers (DRI), and Normalized Difference Vegetation Index (NDVI); geological factors encompassed Geological Lithology (Geol), peak ground acceleration from earthquakes, and distance to faults (Falt); and land cover types (LC) were used to characterise the impact of human engineering activities. The specific sources of information for the static factors are shown in Table 1.
Table 1. Data sources
Impact factor | Data source |
|---|---|
Slope, Aspect, Elevation, Plcur, Procur | Geospatial Data Cloud DEM(https://www.gscloud.cn/) |
Geol, Falt | China Geological Survey (https://geocloud.cgs.gov.cn/) |
Rainfall, DRI | National Earth System Science Data Processing Centre (https://www.geodata.cn/) |
LC | Resource Environment Science and Data Processing Centre(https://www.resdc.cn/) |
NDVI | Geospatial Data Cloud(https://www.gscloud.cn/) |
PGA | Institute of Geophysics, China Earthquake Administration (http://www.gb18306.net/) |
To obtain the rate of deformation(Def), 46 scenes of Sentinel-1 A satellite ascending orbit images were selected, all of which were in interferometric wide-swath mode, with VV polarisation, data type Single Look Complex (SLC), time span from July 2022 to May 2024, with a temporal resolution of 12 days. The SAR data were sourced from ASF satellite data (https://search.asf.alaska.edu/). During processing, precise ephemeris data from the same period were used to correct orbital errors, and a digital elevation model was employed for terrain phase compensation and geocoding.
Methods
The overall framework of this study is illustrated in Fig. 2: (1) Data Preparation: SBAS-InSAR technology was employed to obtain surface deformation rates along railway corridors. Combined with high-resolution optical imagery for identifying potential landslides, an evaluation index system for landslide susceptibility was constructed, incorporating both dynamic and static InSAR factors. After resampling and collinearity testing of each factor, the dataset was divided into training and testing sets at a 7:3 ratio. (2) Model Construction: An MPSO-RF model is established, incorporating deformation rates as dynamic factors through both joint training and weighted superposition methods. Accuracy comparisons and landslide susceptibility zoning along the railway are conducted. (3) Experimental Analysis: Based on the integrated results of SBAS-InSAR and MPSO-RF, a comprehensive analysis of the triggering factors and causal characteristics of landslides along the railway in the study area is performed.
[See PDF for image]
Fig. 2
The research framework of the study
SBAS-INSAR
SBAS-InSAR technology represents an advancement in time-series InSAR analysis, evolving from conventional InSAR methodologies, primarily used for high-precision ground deformation monitoring, effectively reducing interference caused by differences in coherence [15]. Considering the study area’s rugged terrain, dense vegetation cover, and frequent rainfall, which can lead to reduced coherence and noise interference, SBAS-InSAR technology is an ideal method for ground deformation extraction and landslide identification in this study. During the interferometric processing stage, multiple Range Looks and Azimuth Looks are set at a ratio of 4:1. Pixel averaging in the direction and distance components effectively suppresses speckle noise. Each interferometric pair within a subset must undergo differential processing sequentially, following the specific workflow outlined below:
Interference map generation: The main image and sub-image processed by multi-view processing are convolved to generate an interference map. Subsequently, precise orbit data and DEM are used to perform differential processing on the phase to eliminate interference caused by flat ground phase and terrain.
Phase unwrapping and optimisation: The Minimum Cost Flow unwrapping method is used to address phase ambiguity issues, combined with the Goldstein filtering method to suppress residual noise. Interferograms with discontinuous stripes or low coherence are manually adjusted and removed through interactive processing;
Terrain phase removal: Terrain phase contributions are eliminated using an external DEM or reference interferogram;
Differential Model Construction: Separate deformation phase from residual error to generate a time-series interferometric phase dataset. The phase unwrapping expression is:
1
In the equation: represent the phase differences caused by elevation residuals, ground deformation, baseline errors, atmospheric delay, and noise at pixel (x, y) for the kth interferometric pair. After phase unwrapping, an SBAS deformation function model is constructed using linear or nonlinear deformation assumptions, and singular value decomposition is employed to solve the underdetermined system of equations and invert the time series deformation rate, as shown in Eq. 2. Finally, spatiotemporal filtering (combining high-pass temporal filtering and low-pass spatial filtering) is used to separate atmospheric phase and noise, ultimately yielding continuous deformation rates over the monitoring period. The results are then converted into radar line-of-sight (LOS) information in the ascending track.
2
In the formula: is the LOS deformation information of the point to be solved; λ is the radar wavelength.
MPSO-RF model
Random forest models are not only insensitive to collinearity between variables but also capable of flexibly fitting complex data relationships, making them widely used in landslide susceptibility assessment [16]. However, traditional random forest models have certain limitations in terms of learning efficiency and generalisation ability. To improve model performance, this paper introduces a two-layer optimisation strategy to enhance the random forest model. First, the standard particle swarm is optimised to further expand the particle search range. Kusum Deep [17] et al. proposed comparing the current particle position with a linear combination of pbest and gbest. By plotting a vector diagram to compare the particle search range, as shown in Fig. 3, the particle search range is expanded after linearly combining the individual best and group best values. The optimal linear combination obtained is to use and replace the pbestid and gbestid in the standard particle swarm velocity update formula to obtain a new particle velocity update formula as shown in Eq. 3, thereby obtaining the mean particle swarm. Then, the mean particle swarm is used to tune the hyperparameters of the random forest model, and after multiple training iterations, the mean particle swarm-optimised random forest model is obtained.
3
[See PDF for image]
Fig. 3
Comparison of particles in standard PSO and mean PSO motions
In the formula: represents the velocity of the particle in the kth iteration; represents the position of the particle in the kth iteration;µ is the inertial weight;c1 and c2 are learning factors representing the ability of the particle to learn pbest and gbest;r1and r2 are random numbers distributed in the interval [0,1] to increase the randomness of the search.
The MPSO-RF model is constructed by mapping key hyperparameters of the random forest to individual dimensions of the particle position vector. These primarily include the number of trees (n-trees), maximum tree depth (max_depth), minimum samples per split (min_samples_split), and maximum feature subset size (max_features). During particle swarm initialization, each particle corresponds to a set of hyperparameter values. The optimal parameter combination is determined by calculating fitness values using the objective function, where higher fitness values indicate superior particles [18]. The update formula for the i-th particle at generation (K + 1) is:
4
In the formula:andThe fitness values for particle i in generations k and k + 1.
To mitigate the risk of model overfitting, 5-fold cross-validation is implemented in each iteration. The dataset is randomly partitioned into five subsets, with one subset serving as the validation set in rotation while the remaining four are used for training. The fitness values of particles are calculated based on the model performance obtained from each training pass. As iterations progress, the particle swarm gradually converges, yielding initial hyperparameters through multiple training runs. Specific parameters are as follows: population size is 30, maximum iterations are 100, learning factors c1 and c2 are both set to 2, inertia weight w is 1, tree depth ranges from [0, 50], and all other parameters use default values.
Joint evaluation model
Prevalent techniques for integrating INSAR deformation into susceptibility assessment models encompass joint training, constructing evaluation matrices, and weighted superposition [19].
This study primarily employs joint training and weighted superposition methods to construct susceptibility evaluation models.
SBAS-INSAR and MPSO-RF joint training (IJMPSO-RF): After preprocessing INSAR deformation data, it is directly input into the model as a factor alongside static factors for training.
SBAS-INSAR weighted overlay with MPSO-RF (IWMPSO-RF): Based on the research of Chen [20], Meghanadh [21], and others, INSAR data is normalised considering both geological aspects and the spatio-temporal evolution characteristics of landslides. The weighted overlay method is then used to combine it with the landslide susceptibility map constructed using the MPSO-RF model based on static factors. The selection of coefficients is based on a comprehensive analysis of extensive experiments and relevant literature. By comparing the model’s performance under different weighting ratios and achieving high consistency with results from existing studies, the model’s validity and regional applicability are ensured.The formula for the weighted overlay is:
5
In the formula: LDSM refers to the landslide dynamic susceptibility map; LSM refers to the landslide susceptibility map; GDM refers to the ground deformation map.
Results
SBAS-INSAR landslide identification
Deformation rate analysis
Archival images of Sentinel-1 A in ascending orbit from July 2022 to May 2024 were selected to obtain the annual average LOS deformation rate. The black areas in Fig. 3 are incoherence zones, primarily caused by severe signal attenuation due to dense vegetation. Since vegetation roots have strong tensile and shear strength and can effectively stabilise the soil, the risk of landslides is low, so this area was not included in the analysis [15]. In Fig. 3, red (positive values) indicates surface deformation near the SAR satellite’s line of sight, while blue (negative values) indicates deformation away from the satellite. Drawing on the side sampling method proposed by Zhang Yu [22] et al., rich features were extracted from units along the railway line and within a 1-kilometre radius of their surroundings to reveal similar geological patterns. As shown in Fig. 4, the study area is generally in a stable state with no large-scale surface deformation. However, there are some discrete surface deformations around the under-construction and existing railways, with the maximum uplift deformation rate reaching 62.5 mm/a and the maximum subsidence rate reaching 96.8 mm/a. Specifically, Zones B and E of the existing railway and Zones 5 and 6 of the under-construction railway exhibit certain deformation trends.
[See PDF for image]
Fig. 4
Map of deformation rates in the study area
Identification of potential landslides
To more effectively identify potential landslide areas with higher risk, this study limits the monitoring scope to areas along railway lines. The deformation rate maps along the railway lines were converted into point grids and overlaid onto high-resolution imagery from Google Earth. The following landslide identification criteria were applied for screening: (1) absolute deformation rate exceeding 40 mm/year; (2) slope angle at deformation points greater than 5°; (3) spatial clustering of deformation points. The thresholds mentioned above are not empirically determined but rather established through comprehensive reference to prior research. Existing studies indicate that deformation rates exceeding 40 mm/a typically reflect significantly active deformation bodies. Areas with slopes less than approximately 5° rarely experience landslides or noticeable ground deformation, hence a slope > 5° is adopted as the preliminary lower limit for terrain screening. Additionally, genuine ground deformation often exhibits spatially clustered distribution characteristics. Therefore, this study integrates velocity, topography, and spatial clustering as three key indicators to ensure the rationality and reliability of potential landslide identification [23, 24–25].A total of 38 potential landslides were identified in this study. After spatial overlay analysis with existing landslide disaster points, it was found that 9 existing railway areas overlapped with existing landslide points, and 18 under-construction railway areas overlapped, with 11 new potential landslide points identified. Figure 5 shows some typical potential landslides located in the existing railway e area and the under-construction railway 5 area, which will be used for subsequent detailed analysis. Among these, the absolute value of the deformation rate in the existing railway e-zone is approximately 50 mm/a, with deformation points distributed in an elliptical pattern and located near the main railway line. The surrounding area has dense buildings and high human activity intensity, posing a high risk of ground subsidence. The under-construction railway 5-zone largely meets the criteria for landslide identification. The area is located in a steep slope region with a high deformation rate, and the points exhibit typical fan-shaped clustering, indicating a potential risk of landslides.
[See PDF for image]
Fig. 5
Typical regional deformation rates and optical images
Taking the two typical potential landslide areas shown in Fig. 4 as examples, we further analysed their spatio-temporal evolution characteristics and verified the identification results of the two areas using spatial evolution and time series curves.
Spatial evolution
Using 9 July 2022 as the deformation baseline, the study investigates the spatial evolution process of two typical regions. As shown in Fig. 6(a), most areas in the existing railway region e remained stable during the monitoring period, while relatively sustained and rapid subsidence deformation was detected in the existing railway region e. The deformation spreads outward from the deformation centre, beginning in February 2023. After May of the same year, the deformation rate increases, and after January 2024, it gradually stabilises, with the diffusion rate slowing down and the cumulative deformation gradually rising to its maximum value. As shown in Fig. 6(b), compared to the existing railway, the deformation of the under-construction railway evolved from early scattered small-scale distributions to concentrated clusters, stabilising within approximately seven months, with a relatively faster deformation development rate. The results indicate that, from the perspective of spatial evolution characteristics, both typical regions experienced continuous surface deformation, posing a potential risk of landslides. The evolutionary process corresponds with the traits of landslide development, confirming the validity of the potential landslide identification outcomes.
[See PDF for image]
Fig. 6
The process of spatial evolution. (a) Existing railway area e, (b)Under-construction railway area 5
Time-series curves
To conduct an in-depth analysis of the long-term time-series deformation characteristics of typical potential landslide areas, representative monitoring points A and B with deformation rates close to the regional average were selected from the existing railway e and the under-construction railway 5, respectively, to conduct time-series deformation analysis. It should be noted that points A and B are used solely to reflect the overall deformation trend of the area. The fact that their deformation values are slightly below the threshold does not affect the overall assessment of the potential landslide zone, as the maximum local deformation rate within the area has already exceeded the 40 mm/a threshold.The results are shown in Fig. 7. Point A exhibits overall continuous subsidence with distinct seasonal characteristics, with subsidence rates significantly accelerating during July–September. Previous studies have indicated that landslides in the study area are primarily triggered by heavy rainfall, daily rainfall reaching 35–50 mm may trigger accelerated deformation [26]. Combining concurrent daily rainfall data, it was found that when daily rainfall exceeds 40 mm, the rate of deformation at the monitoring points has changed.Point B is situated in the central to upper region of the landslide slope and shows an overall trend of continuous uplift. During July to September, under the influence of heavy rainfall, the uplift rate accelerates, with localised jump-like changes. Heavy rainfall caused an elevation in pore water pressure, weakening the shear strength of the slip zone and triggering shear movement of the landslide body, manifested as slope elevation. This deformation characteristic aligns with the typical landslide pattern of ‘forecast slope subsidence followed by slope body elevation.’ It should be noted that during short-term heavy rainfall events, pore water pressure in the shallow soil layers of slopes increases significantly. According to the Mohr–Coulomb criterion (τ = c + σ′ tan φ), effective stress σ′ decreases as pore pressure u increases, leading to a reduction in shear strength τ. Soils in the study area exhibit liquid limits around 35%–40%. When daily rainfall exceeds 40 mm, moisture content approaches liquid limit conditions, causing soil structure softening. Shear parameters c and φ decrease by approximately 10%–15% on average, readily inducing shear slip of the sliding body.Analysis results indicate that the surface deformation characteristics at the two typical points are consistent with the criteria for identifying potential landslides, thereby validating the rationality of the potential landslide identification results.
[See PDF for image]
Fig. 7
Relationship curves between daily rainfall and cumulative form variables. (a) Point A deformation sequence curve, (b)Point B deformation sequence curve
Landslide susceptibility model evaluation
Evaluation of co-linearity and importance of influencing factors
Uniform sampling of all factors at a resolution of 30 × 30 m, the Variance Inflation Factor (VIF) method in SPSS software was used to test for multicollinearity among the 13 selected static factors and deformation rate factors. A VIF value of ≥ 5 was used as the criterion for identifying multicollinearity issues. The analysis results are as follows: PGA (1.14), Def (1.16), NDVI (1.33), DRI (1.38), Procur (2.45), plcur (2.14), Aspect (1.29), Elevation (1.41), Slope (2.5), LC (1.46), Geol (2.07), Falt(1.44), and Rainfall (1.18). Among these, the VIF value for slope is the highest at 2.5, while the VIF values for the remaining factors are all between 1 and 2, significantly below the collinearity threshold. This indicates that there is no obvious collinearity issue among the selected factors, and they can all be used for evaluating landslide susceptibility along railway lines.
Under two conditions, with and without deformation rate, the MPSO-RF model was applied to obtain the importance ranking results. The results in Fig. 8 show that when only relevant static factors are included, the factors that have a greater impact on landslides mainly include rainfall, PGA, Elevation, and Slope. After introducing the Def, the importance of the original main controlling factors relatively decreased, and the importance of the deformation rate reached 0.165, indicating that the deformation rate is one of the important factors affecting landslides along the railway line.
[See PDF for image]
Fig. 8
Importance ranking results
Model accuracy
The Receiver Operating Characteristic (ROC) curve assesses the prediction efficacy of a binary classification model, illustrating the relationship between specificity and sensitivity at different decision thresholds. The Area Under the Curve (AUC) is computed to measure prediction accuracy, with AUC values approaching 1 signifying more model accuracy [27]. The training and testing datasets were divided in a 7:3 ratio, and the prediction results are shown in Fig. 9. The AUC value of the MPSO-RF model without considering the deformation rate factor is 84.5%; after introducing the deformation rate factor, the AUC values of IWMPSO-RF and IJMPSO-RF are 87.5% and 91.04%, respectively. Compared to the static factor alone, the proposed model improves accuracy by 3% and 6.54%, respectively, indicating that the deformation rate factor can enhance predictive accuracy to some extent.
[See PDF for image]
Fig. 9
ROC curves
Model volatility partitioning
Using the MPSO-RF model and the natural breakpoint method, a risk assessment zoning for landslide susceptibility along railway lines was conducted, with the results classified into five categories: low, moderately low, moderate, moderately high, and high susceptibility zones. The spatial distribution of these zones is shown in Fig. 10. As shown in Fig. 9, the disaster points of the three models are predominantly concentrated in the high-risk zones, which to some extent validates the rationality of the model’s predictive results [28]. The high-risk zones are primarily distributed in Yucheng District (existing railways f and j) and Tianquan County (under-construction railway 2), all located within rainfall-dominated landslide high-risk zones. When compared with actual landslide disaster points, the MPSO-RF model can effectively identify large-scale high-risk areas such as regions f and j, but it is less sensitive to small landslide points, posing a certain risk of misclassification. However, as shown in Fig. 10(b) and (c), the introduction of dynamic factors significantly enhances the model’s ability to identify potential landslide areas and improves the spatial resolution of key sections such as regions f and j. Further statistical analysis of the area distribution along the railway line revealed that the proportion of high-risk and moderately high-risk zones was 84.5% without the deformation rate factor, which increased to 87% after its introduction. This indicates that the inclusion of the deformation rate factor demonstrates good performance in identifying potential landslides and improving evaluation accuracy.
[See PDF for image]
Fig. 10
Three model susceptibility zoning maps. (a) MPSO-RF, (b) IJMPSO-RF, (c) IWMPSO-RF
Joint evaluation of the causes of landslide disasters
Based on the landslide identification results from SBAS-INSAR and the evaluation results from the MPSO-RF model, combined with various controlling factors, a systematic analysis of the causative mechanisms of landslides along railway lines was conducted. First, from the perspective of rainfall factors, heavy rainfall significantly increases soil moisture content and pore water pressure, weakens slope shear strength, and causes significant changes in deformation acceleration, thereby triggering landslides. Second, topographical and geological conditions indicate that the strata in the study area are primarily composed of mudstone, carbonate rock, and conglomerate, with loose structures and well-developed joints and fractures. Under the influence of heavy rainfall, these conditions are prone to softening and instability, exacerbating landslide risks. Finally, from a regional tectonic perspective, the study area is situated at the tectonic intersection of the Tianquan-Xingjing Fault and the Longmen Mountain Fault. Earthquake motions are prone to reflection and superposition within fault structures, enhancing local stress effects. This leads to stress accumulation and increased deformation rates in landslide bodies. When these exceed critical thresholds, slope instability is highly likely to occur.
In summary, landslides along the railway line in the study area are driven by multiple factors, including rainfall, geomorphology, rock properties, and tectonic stress. Incorporating potential landslides identified by SBAS-INSAR technology and deformation rate factors into the susceptibility evaluation system and conducting joint evaluations can achieve a more comprehensive and accurate identification and analysis of landslide risks.
Discussion
This study combines SBAS-InSAR technology with machine learning models to assess landslide susceptibility along railway corridors. While SBAS-InSAR provides high-precision ground deformation monitoring, deformation signal extraction is constrained in densely vegetated areas due to poor coherence. Although multi-view processing reduces noise, deformation data quality may deteriorate in complex terrain and thick vegetation zones, compromising landslide detection accuracy. Furthermore, deformation signal analysis is constrained by satellite image resolution and temporal data coverage, where temporal intervals and spatial resolution may limit detection of short-term or rapid landslide evolution.
Comparison with existing research: Ghaderpour et al. [6] employed PS-InSAR technology to monitor the reactivation timing and deformation rates of slow-sliding bodies. Their findings revealed a significant correlation between precipitation variations and surface deformation trends, validating the crucial role of deformation monitoring in identifying landslide activity. Although their study primarily focused on slow-sliding body monitoring rather than landslide susceptibility assessment along railway lines, their conclusions provided reference for selecting deformation factors in this research. Gao et al. [29] incorporated SBAS-InSAR deformation factors into a Bayesian network model for landslide susceptibility assessment, demonstrating that the inclusion of deformation information significantly enhanced model prediction accuracy. In contrast, this study employs a PSO-RF model combined with a weighted overlay strategy, organically integrating deformation factors with static geological environment factors. This approach further expands the application of deformation monitoring data in landslide susceptibility evaluation along railway corridors. In summary, although existing studies and this research differ in technical approaches, study areas, and evaluation objectives, they all demonstrate that methods integrating deformation monitoring with machine learning exhibit high feasibility and application value in landslide identification and susceptibility assessment.
This study employs a mean particle swarm optimization (MPSO) random forest model. While this approach effectively enhances the model’s predictive accuracy, uncertainty persists in hyperparameter selection and optimization. Parameters such as MPSO initialization, iteration count, and learning rate significantly influence model performance. Although cross-validation and multiple experiments were conducted to mitigate these factors, performance fluctuations under different parameter combinations may still lead to variations in final results. Due to limitations in data acquisition channels, this study could not account for other dynamic factors potentially influencing landslide occurrence, such as soil moisture and groundwater level changes. This omission may introduce bias into the landslide susceptibility assessment results.
Future research will integrate multi-source data for landslide susceptibility evaluation, combining LiDAR, high-resolution UAV imagery, and SBAS-InSAR data to enhance identification accuracy in complex terrain and vegetated areas. Data fusion can overcome the limitations of single data sources, enhancing model adaptability and accuracy. Simultaneously, exploring dynamic factors like soil moisture and groundwater levels will help improve the timeliness and precision of landslide early warnings. Furthermore, focusing on model universality, particularly in cross-regional and cross-scale applications, will provide more precise technical support for landslide disaster prediction and prevention.
Conclusion
This paper combines SBAS-INSAR technology and optical remote sensing to identify landslides along railway lines. It introduces deformation factors into the MPSO-RF model through joint training and weighted superposition to assess the landslide susceptibility of the study region along railway lines, drawing the following conclusions:
SBAS-InSAR technology was employed to observe deformation in the study area and identify landslides in combination with optical remote sensing. Typical potential landslide areas along the railway line were selected for spatio-temporal evolution characteristic analysis. The identified areas were verified to meet the landslide criteria through spatial evolution and time series curves, proving that SBAS-InSAR technology can be used for landslide identification along railway lines.
After obtaining the deformation rate factor, the joint training and weighted superposition methods were used to introduce the factor into the MPSO-RF model. In comparison to the the static model, the precision of the joint training and weighted superposition models improved by 6.54% and 3%, respectively. Among them, the joint training method mainly improved the model accuracy, while the weighted superposition method contributed less to the improvement in accuracy but helped to expand the identification range of high-risk areas in the susceptibility partition.
The joint evaluation based on the SBAS-INSAR and MPSO-RF models can effectively improve the accuracy of landslide identification along railway lines, providing a strong theoretical basis for the construction of a railway line safety environment management platform.
Acknowledgements
The authors wish to express their heartfelt gratitude to all experts and collaborators who supported this research. We thank the teams involved in SBAS-InSAR data processing, evaluation factor construction, and experimental validation for their dedicated efforts. We also acknowledge the institutions that provided optical imagery, DEM data, geological information, and meteorological datasets essential to this study. The insightful comments and suggestions provided by the reviewers played a crucial role in enhancing the quality of the manuscript. The successful completion of this work would not have been possible without the assistance and support from all contributors, to whom the authors extend their sincere appreciation.
Author’ contributions
ZSH collected and provided some of the data, analysed the experimental data, and wrote the first draft of the paper; GRC provided guidance on the paper topic and research approach, participated in the overall structure design of the paper, and provided some references and algorithm optimisation suggestions; both authors participated in the language refinement and paper revision work. All authors read and approved the final manuscript.
Funding
This paper was funded by the Natural Science Foundation of Gansu Province (No. 21JR1RA254).
Data availability
The data used and analysed in this study are available on request from the corresponding author.
Declarations
Competing interests
The authors declare that they have no conflict of interest.
Abbreviations
Interferometric Synthetic Aperture Radar
Mean Particle Swarm Optimisation -Random Forest
Small Baseline Subset- Interferometric Synthetic Aperture Radar
Planar curuature
Profile curuature
Distance from rivers
Normalized Difference Vegetation Index
Geological Lithology
Distance to faults
Land cover types
Deformation
Single Look Complex
Line of Sight
Receiver Operating Characteristic
Area Under the Curve
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
1. Guo, ZW; Shi, WF; Li, JB et al. Design and realisation of line safety environment management system. China Railway; 2023; 09, pp. 34-39. [DOI: https://dx.doi.org/10.19549/j.issn.1001-683x.2023.07.17.004]
2. Liu QH, Guan ZB, Zhao Y et al (2023) A review of environmental safety monitoring system for high-speed railway operation. China Railway 0440–47. https://doi.org/10.19549/j.issn.1001-683x.2022.11.18.003
3. Jiang, N; Li, YG; Han, Z et al. A side-sampling based Linformer model for landslide susceptibility assessment: a case study of the railways in China. Geomat Nat Hazards Risk; 2024; 15, [DOI: https://dx.doi.org/10.1080/19475705.2024.2354507] 2354507.
4. Aswathi, J; Kumar, RBB; Oommen, T et al. InSAR as a tool for monitoring hydropower projects: a review. Energy Geosci; 2022; 3,
5. Jia, Y; Wu, CY; Wang, LJ et al. Landslide susceptibility assessment integrating InSAR and machine learning. J Geodesy Geodyn; 2025; 45,
6. Ghaderpour, E; Mazzanti, P; Bozzano, F; Scarascia Mugnozza, G et al. Ground deformation monitoring via PS-InSAR time series: an industrial zone in Sacco River Valley, central Italy. Remote Sensing Applications: Society & Environment; 2024; 34, [DOI: https://dx.doi.org/10.1016/j.rsase.2024.101191] 101191.
7. Rui, TAO; Lau, A; Mossefin, ME; Kong, G; Nordal, S; Pan, Yu-tao et al. Monitoring of ground displacement-induced railway anomalies using PS-InSAR techniques. Measurement; 2025; 248, 116863. [DOI: https://dx.doi.org/10.1016/j.measurement.2025.116863]
8. Guo, JQ; Xi, WF; Yang, ZQ et al. Landslide hazard susceptibility evaluation based on SBAS-InSAR technology and SSA-BP neural network algorithm: a case study of Baihetan Reservoir area. J Mt Sci; 2024; 21,
9. Hu, YX; Xu, ZJ; Yan, W et al. Prediction of deformation intervals of earthy landslide slopes based on bootstrap and GA-Elman. J Railway Sci Eng; 2023; 20,
10. Wu, YC; Zhou, HX; Che, AL et al. A rough set-neural network based study of seismic landslide susceptibility in IBURI. J Rock Mech Eng; 2021; 40,
11. Yang, C; Liu, LL; Zhang, YL et al. Landslide susceptibility assessment based on bayesian optimisation of machine learning hyperparameters. Geoscience Technol Bull; 2022; 41,
12. Razavi Termeh SV, Sadeghi Niaraki et al (2025) Investigating the efficacy of physics-based metaheuristic algorithms in combination with explainable ensemble machine-learning models for landslide susceptibility mapping. Stoch Env Res Risk Assess 1–33. https://doi.org/10.1007/s00477-025-02911-7
13. Mou, JQ; Zhuang, JQ; Wang, SB et al. Evaluation of landslide susceptibility in ya’an based on deep neural network model. Chin J Geol Hazards Prev; 2023; 34,
14. Jiang, XC; Liu, T. Research on risk assessment of geological disasters based on hierarchical analysis-informative model – taking Ya’an City, Sichuan Province as an example. Geol Hazards Environ Prot; 2024; 35,
15. Zeng, TR; Wu, LY; Jin, BJ et al. Landslide dynamic susceptibility mapping based on stacking integration strategy and SBAS-InSAR. J Rock Mech Eng; 2023; 42,
16. Liu, S; Wang, T; Cao, JW et al. Evaluation of rainfall-induced landslide susceptibility based on an optimised random forest model – a case study of extreme rainfall events in the West Qinling mountains. Geol Bull; 2024; 43,
17. Deep, K; Bansal, JC. Mean particle swarm optimisation for function optimisation. Int J Comput Intell Stud; 2009; 1,
18. Wang, WD; Liu, P; Gong, L. Landslide disaster susceptibility zoning in Sichuan Province based on support vector machine model. J Railway Sci Eng; 2019; 16,
19. Qin JS, Li WL, Shan YF, Bulletin et al (2025) Evaluation of landslide susceptibility in Asbestos County by integrating time-series InSAR deformation. Geoscience and Technology Bulletin, 1–14. [Online/OL, 2025-04-05]. https://doi.org/10.19509/j.cnki.dzkq.tb20240342
20. Cao, C; Zhu, K; Xu, P et al. Refined landslide susceptibility analysis based on InSAR technology and UAV multi-source data. J Clean Prod; 2022; 368, [DOI: https://dx.doi.org/10.1016/j.jclepro.2022.133146] 133146.
21. Devar, M; Tiwari, A et al. Landslide susceptibility mapping using MT-InSAR and AHP enabled GIS-based multi-criteria decision analysis. Nat Hazards Risk; 2021; 12,
22. Zhang, Y; Li, L; Hou, YL et al. Study on the identification and risk of landslides after railway construction in mountainous areas. J Railway Eng; 2024; 41,
23. Yan, L; Xiong, Q; Li, D et al. InSAR-Driven dynamic landslide hazard mapping in highly vegetated areas. Remote Sens; 2024; 16, 3229. [DOI: https://dx.doi.org/10.3390/rs16173229]
24. Chen, Z; Song, D; Du, Y et al. Investigation on the spatial distribution of landslides in Sichuan Province, southwest China. Geomat Nat Hazards Risk; 2023; 14,
25. Dai, H; Zhang, H; Dai, H et al. Landslide identification and gradation method based on statistical analysis and spatial cluster analysis. Remote Sens; 2022; 14, [DOI: https://dx.doi.org/10.3390/rs14184504] 4504.
26. Wang, WD; Fu, QX; Tang, R. Time-space sensitivity analysis of landslide geological hazards in the line area of Sichuan-Tibet railway (Chengdu–Ya’an section). J Railway Sci Eng; 2018; 15,
27. Gao, BH; He, Y; Zhang, LF et al. Dynamic assessment of CNN landslide susceptibility considering InSAR deformation: a case study of Liu Jiaxia reservoir area. J Rock Mech Eng; 2023; 42,
28. Akinci, H Ayse Yavuz Oalp. Investigating the effects of different data classification methods on landslide susceptibility mapping. Adv Space Res; 2025; 75,
29. Gao, X; Wang, B; Dai, W et al. A landslide susceptibility assessment method using SBAS-InSAR to optimize Bayesian network. Front Environ Sci; 2025; 13, [DOI: https://dx.doi.org/10.3389/fenvs.2025.1522949] 1522949.
© The Author(s) 2025. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.