Landslide Hazard Analysis Combining BGA-Net-Based Landslide Susceptibility Perception and Small Baseline Subset Interferometric Synthetic Aperture Radar in the Baige Section in the Upper Reaches of Jinsha River

Abstract

The geological and topographic conditions in the upper reaches of the Jinsha River are intricate, with frequent occurrences of landslides. Landslide Susceptibility Prediction (LSP) in this area is a crucial aspect of geological disaster risk management. This study constructs an LSP model using a Convolutional Neural Network (CNN) based on a Bilateral Aggregation Guidance (BAG) strategy, termed BGA-Net. A comprehensive landslide hazard analysis, integrating static landslide susceptibility zonation with dynamic surface deformation monitoring, was therefore conducted. The study area selected was the upper reaches of the Jinsha River, particularly the site of the Baige landslide. The BGA-Net model was first proposed for LSP generation, achieving an accuracy exceeding 85%, while the Small Baseline Subset Interferometric Synthetic Aperture Radar (SBAS-InSAR) technology was jointly applied to comprehensively analyze the dynamic geological hazard risk at a regional scale. The final results were presented in a lookup table format and mapped to delineate and grade each risk level. The results show the method is practical, with high feasibility. Compared with traditional machine learning methods, the BGA-strategy-oriented CNN model more effectively differentiated the extremely low- and extremely high-susceptibility areas, enhancing decision-making processes.

1. Introduction

The upper reaches of the Jinsha River area have complex geological and topographical conditions, with the typical development of landslide disasters. The upper regions of the Jinsha River are situated within a plateau mountainous terrain characterized by elevated peaks and deep valleys. Within this area, prevalent rock formations consist predominantly of schist, shale, and quartzite, exhibiting pronounced foliation and abundant joints and fractures [1]. These geological features render the terrain susceptible to deformation and erosion, particularly under the influence of geological tectonic forces and water erosion. Noteworthy faults within the vicinity include the Jinsha River Fault, Maisu Fault, Zengke–Shuoqu Fault, Litang Fault, and Yushu–Ganzi Fault, indicating a history of frequent geological tectonic movements. In October 2018, a large landslide occurred in the Baige district, resulting in the blockage of the Jinsha River and the formation of a barrier lake, which posed great threats to multiple hydropower stations and millions of residents. The incident also led to the destruction of numerous residential buildings and cattle sheds, necessitating the urgent evacuation and resettlement of over 21,000 individuals for their safety. In November of the same year, another collapse at the same location caused the formation of an additional barrier lake, impacting approximately 34,200 people and causing damage to houses, roads, bridges, and agricultural land downstream. Additionally, the landslides significantly altered the hydrological conditions, impacting downstream ecosystems and disrupting aquatic habitats. Moreover, field station monitoring results indicate that surface deformation in the area is ongoing, indicating serious geological and environmental safety hazards. Therefore, conducting landslide susceptibility and hazard analysis in the Baige area of the upper Jinsha River is of great significance.

Landslide susceptibility and hazard assessment are essential for understanding the spatiotemporal distribution of geological disasters and are crucial for the effective prevention and mitigation of such events [2,3]. Currently, commonly used LSP models can be categorized into three types: qualitative analysis based on empirical reasoning, deterministic models based on physical mechanisms, and semi-quantitative methods using mathematical and statistical models [4]. The qualitative analysis based on empirical reasoning includes methods such as the engineering geological analogy method and slide causative mechanisms process simulation. These methods can identify the potential hazards of geological disasters from a macroscopic perspective, yet are limited by the experts’ experience and subjectivity, and are unable to provide quantitative results. Moreover, the methods struggle with nonlinear relationships and integrating dynamic data processions. Deterministic models based on physical mechanisms can more accurately simulate slope stability. However, these methods typically require a large amount of data and complex calculations and are only suitable for small-scale studies as the formulas have limitations at a large scale and regarding dynamic surface deformation data integration. Semi-quantitative methods based on statistical models are relatively effective for large-scale landslide susceptibility mapping and prediction, but have poor compatibility with dynamic changes and nonlinear relationships in the data.

Overall, researchers have proposed numerus methodologies for landslide susceptibility assessments. The widely used models include statistical models such as the Analytic Hierarchy Process (AHP) [5] and Information Value (IV) model [6], and machine learning models such as Logistic Regression (LR) [7,8], the Artificial Neural Network (ANN) model [9], Random Forest (RF) [10], and Support Vector Machine (SVM) [11,12]. However, these models cannot fully utilize the abundant information of high-resolution graphical layers and have limitations in terms of model depth and semantic structure extractions [13,14,15]. The landslide formation mechanisms are complex and the triggering factors are typical, with differences in their non-linear features. In recent years, with the rise of deep learning technology, the Convolutional Neural Network (CNN) model has developed rapidly and has gradually been applied in the field to geological disaster risk assessments. Applications include debris flow detection [16], collapse susceptibility assessments [17], and landslide susceptibility assessments [18,19]. Deep learning models belong to the machine learning models category, which is built on artificial neural network structures. The core principle is to learn data representations through the multiple non-linear transformed layers forming the “neural network”, thereby modeling complex patterns and relationships. The CNN addresses the limitations of traditional models in automatically learning the complex feature patterns, feature representations, hierarchical feature learning, parameter sharing, and sparse connections. By stacking multiple convolutional and pooling layers with activation functions, CNNs can progressively extract complex semantic features from a hierarchical, non-linear data structure, which is of great help when attempting to improve the accuracy of LSP results. Therefore, it is crucial to construct a CNN model suitable for susceptibility assessments.

To meet this application demand and considering the fusion of features at different scales, this paper constructs a Bilateral Aggregation Guidance strategy CNN, termed BGA-Net, for LSP. A joint landslide hazard risk analysis was then performed, integrating the LSP results with the temporal Interferometric Synthetic Aperture Radar (InSAR) derived surface deformation monitoring. The aim was to further provide technical support for elaborate disaster prevention and mitigation around Baige in the upper reaches of the Jinsha River and to accurately capture the semantic surface geographic features that reflect the landslide characteristics. BGA-Net introduced a Context Embedding Block (CEB) at the end of the convolutional procession to enhance the interaction between deep and shallow features. The BGA-Net and LSP combination and the focus on deformation is a novel approach that leverages both static and dynamic monitoring. This enables a comprehensive analysis of landslide disasters.

2. Study Area and Data

2.1. Overview of the Study Area

The upper reaches of the Jinsha River are located in a plateau mountainous area, characterized by high mountain and valley landforms. The study area is situated within a plateau mountainous terrain, distinguished by elevated peaks and deep valleys. The predominant rock formations in this area include schist, shale, and quartzite, which display pronounced foliation as well as numerous joints and fractures [1]. These geological characteristics make the terrain highly susceptible to deformation and erosion, particularly under the influence of tectonic forces and water erosion. Significant faults in the vicinity include the Jinsha River Fault, Maisu Fault, Zengke–Shuoqu Fault, Litang Fault, and Yushu–Ganzi Fault, indicating a history of frequent tectonic movements. The geological conditions of the upper reaches of the Jinsha River are complex, characterized by frequent tectonic activities that lead to the recurrent development of geological disasters. This region is a typical, large-scale, landslide-prone area in the southwest region of China. According to statistics, there have been more than 1300 landslides in the upper reaches of the Jinsha River, including documented major landslides such as the Temi ancient landslide [20], the Laojuntan landslide (1965), and the Baige landslide (2018). The specific typical landslide cases are shown in

Table 1. The detailed content of landslide cases.

Slide Name	Time	Triggering Factor	Impact
Temi ancient landslide	about 1.4 ka BP	Earthquake-triggered	Barrier lakes formed; the terrain changed
Laojuntan landslide	1965	Rainfall-triggered	Significant loss of life and property
Baige landslide	2018	Gravity-triggered	Direct losses of about CNY 153 million; serious consequences, resulting in multiple deaths and injuries

.

Landslides in the upper reaches of the Jinsha River are influenced by both geological structures and human disturbances. The slope foot undergoes riverine scouring while being jointly eroded by precipitation and other hydrology factors. The foot of the slope is eroded by river water, rainwater, and other factors. The roads in this area are often built along the mountainside. In this way, vegetation is destroyed, infrastructure is built, and industry and agriculture develop, increasing the overall instability of the slope, leading to the frequent occurrence of large landslides, and posing a serious threat to people’s lives and the safety of property. This paper takes the upper reaches of the Jinsha River as the study area to perform a comprehensive landslide hazard analysis. Detailed information of the study area is shown in Figure 1.

2.2. Research Data

For the landslide susceptibility assessment, multi-resource, raw data were collected (

Table 2. Data source.

Index	Data Source	Data Type	Data Usage
1	Sentinel-1A satellite data	Radar data	Large-scale surface deformation monitoring
2	Field geological survey data	Geological survey textual and image records	Analyze landslide and slope development and disaster mechanisms
3	High-Resolution Gaofen-2 data	Optical remote sensing image data	Analyze surface characteristics and extract disaster-bearing bodies in hazardous areas
4	SRTM DEM	DEM	Assist with Sentinel-1A satellite data preprocessing; extract hazard factors: elevation, slope, aspect, curvature, TWI, and TRI
5	Road vector	Vector data	Extract typical disaster-bearing bodies; extract hazard factor: distance to road
6	River vector	Vector data	Extract hazard factor: distance to river
7	Fault vector	Vector data	Extract hazard factors: distance to fault and fault distribution density
8	CHIRSP Pentad data	Meteorological satellite inversion data	Extract hazard factor: annual average precipitation
9	Global 30 m resolution land use type data	Land use type classification data	Extract hazard factor: land use type
10	Landsat 8 OTL satellite data	NDVI	Extract hazard factor: NDVI

) for the construction of predisposing factors, as shown in Figure 2. In this study, the environmental predisposing factors contained lithology, vegetation, hydrology, topography, geomorphology, and human activities [21]. The geological significance of these features was quantified and transformed into attribute information that can be processed by model. The relevant operations mainly include satellite image processing, geological mapping digitization, and Digital Elevation Model (DEM) operations. The software platforms used include ArcMap (version 10.2) and Google Earth Engine. After processing the data, all feature layers were uniformly processed into raster data (.tiff format) with a spatial resolution of 30 m. In terms of spatial reference systems, the Universal Transverse Mercator (UTM) projection under the WGS84 coordinate system was uniformly selected, and the 47N zone was selected based on the latitude and longitude of the study area.

The landslide point data were taken from the GEOVIS Earth Open Platform (https://www.geovisearth.com/ accessed on 6 March 2022). A total of 338 historical landslide points were distributed within the study area. Based on the collection and statistical analysis of historical landslide data investigation materials, this study created a 1 km buffer zone around landslide points, which is considered a stable area that is less prone to landslides. Outside the buffer zone, an equal number of non-landslide points were randomly selected as negative samples for training [22]. Landslides and non-landslide samples were collected at a ratio of 1:2, therefore, a total of 676 negative samples were selected, as shown in Figure 3.

3. Landslide Susceptibility Prediction Based on BGA-Net

3.1. Basic Framework of Landslide Susceptibility Prediction (LSP) Model Based on CNN

A Convolutional Neural Network (CNN) is often applied in image tasks such as image classification, image segmentation, and object detection. The use of a CNN can allow for the weight sharing of convolution kernels in the hidden layers and establish local connections that greatly reduce the complexity of the model, preventing the overfitting of the processing process and improving the processing speed. During model training, CNN can automatically update model parameters by finding the direction of weight updates that reduces the loss function through gradient descent and the backpropagation of gradients, ultimately minimizing the loss function to determine the optimal model parameters. The feature learning ability of CNN is achieved through convolutional calculations.

This study constructed BGA-Net, a model that extracts and enhances feature information through convolution operations and a dual-branch approach (the Texture Extraction Branch (TEB) and Semantic Extraction Branch (SEB)). To avoid the loss of semantic information, this study introduces the Context Embedding Block (CEB). Finally, the landslide classification results were output through the fully connected layer. The convolution kernel slid over the input image or intermediate layer features and performed convolution operations with the corresponding regions to extract image features [23], as shown in the following principles:

$${\mathrm{X}}_{\mathrm{i}}=\mathrm{f}({\mathrm{W}}_{\mathrm{i}}\otimes {\mathrm{X}}_{\mathrm{i}-1}+{\mathrm{B}}_{\mathrm{i}}),$$

(1)

Here, ${\mathrm{X}}_{\mathrm{i}}$ represents the output of the $\mathrm{i}$ th layer in the CNN model, $\mathrm{f}(·)$ is the activation function, ${\mathrm{W}}_{\mathrm{i}}$ is the weight matrix of the $\mathrm{i}$ th layer, ${\mathrm{B}}_{\mathrm{i}}$ is the bias matrix of the i th layer, and ⊗ represents matrix multiplication.

The CNN model is mainly composed of three parts: the input layer, the hidden layers, and the output layer. The hidden layers primarily include the following types: convolutional layers, pooling layers, and fully connected layers. The detailed principles of each layer are as follows.

Convolutional Layer: This layer is mainly used to extract local features from the input data. The principle is to capture different features of the input data, such as edges, textures, and shapes, by applying multiple convolution kernels, as shown in Equation (2).

$$\left(I\otimes K\right)\left(i,j\right)=\sum _{m=0}^{M-1}\sum _{n=0}^{N-1}I\left(i+m,j+n\right)\otimes K\left(m,n\right)$$

(2)

where $I$ represents the input matrix, $K$ represents the convolution kernel matrix, M and N are the height and width of the convolution kernel, respectively, and m and n are the indices. $\left(i,j\right)$ represents the coordinates of the output feature map.

Fully Connected Layer: This layer flattens the input features into a one-dimensional vector, then performs a linear transformation through the weight matrix and bias, and finally applies an activation function to obtain the output. This layer maps the high-dimensional feature vector to the final classification space. It can be expressed by Equation (4), where y represents the output vector, x represents the input vector, $W$ represents the weight matrix, b represents the bias vector, and $f$ represents the activation function (such as the Rectified Linear Unit (ReLU), Sigmoid, etc.).

$$y=f\left(W\otimes x+b\right)$$

(3)

Thirteen factors, including NDVI, TRI, TWI, elevation, distance to roads, distance to rivers, distance to faults, annual average precipitation, plan curvature, slope, aspect, profile curvature, and lithology, were used as static predisposing factors for LSP. A CNN model for LSP was constructed, as shown in Figure 4.

3.1.1. BGA-Net Network Structure

This study adopted a dual-branch strategy for CNN named Bilateral Guided Aggregation (BGA) (as shown in Figure 4), which is based on the characteristics of landslides and integrates convolution technology. The model aims to improve the accuracy and efficiency of risk assessment by accurately capturing the surface geographical features that reflect the spatial distribution, scale, and boundaries of landslides through integrating local spatial features with global semantic understanding.

The BGA module can output feature representations as dual branches at different levels. Simple fusion methods ignore the diversity of the two types of information. Therefore, the specific steps are as follows: First, a unified normalization of the data was carried out. In the TEB branch, the convolutions were performed by depthwise separable convolution (DWConv), followed by standard convolutions. The result was then multiplied with the outputs from the SEB branch, where standard convolution and up-sampling had already been performed. The revised process was applied equally to the TEB branch. Finally, the results of the two matrix multiplications were combined using the Stacking operation. This module uses a rich feature combination method to integrate features at different scales, allowing for the effective correlation and fusion of information from the two branches.

3.1.2. Context Embedding Block (CEB)

This paper integrated the CEB at the end of the SEB branch to enhance the extraction of the comprehensive semantic information learning [18]. As mentioned earlier, the Semantic Branch needs a large receptive field to capture high-level semantic information. This module uses global average pooling to concentrate and integrate feature hierarchy information to obtain global semantic information. The CEB employs a residual structure to prevent the gradient vanishing problem caused by the increased depth of the network, ensuring that large-scale semantic information is not lost during input.

3.1.3. Bilateral Guided Aggregation (BGA)

In the network structure, there are generally simple information fusion methods such as stacking or concatenation (Concatenate). However, the outputs of the two branches in this model have feature representations at different levels. Simple fusion methods ignore the diversity of the two types of information. Therefore, this experiment designed a Bilateral Guided Aggregation (BGA) module. The specific steps are as follows: First, a unified normalization of the data was carried out. In the DB branch, a portion of the data were selected for depthwise separable convolution (DWConv), followed by standard convolution (Conv). The result was then multiplied with the output from the SB branch, where standard convolution (Conv) and upsampling had already been performed. The same process was applied to the SB branch. Finally, the results of the two matrix multiplications were combined using the Add operation. This module uses a rich feature combination method to integrate features at different scales, allowing for the effective correlation and fusion of information from the two branches.

3.2. Validation

This study used the following quantitative metrics to validate the performance of BGA-Net: Overall Accuracy (OA), Precision, Recall, F1-Score, and Area Under the Curve (AUC). The formulas for calculating these metrics are listed as Equations (4) to (8).

$$\mathrm{O}\mathrm{A}=\frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P}+\mathrm{F}\mathrm{N}}$$

(4)

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}$$

(5)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$

(6)

$$\mathrm{F}1-\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=\frac{2\times \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}$$

(7)

$$\mathrm{A}\mathrm{U}\mathrm{C}=\frac{\sum \mathrm{T}\mathrm{P}+\sum \mathrm{T}\mathrm{N}}{\mathrm{P}+\mathrm{N}}$$

(8)

Additionally, the numbers of True Positive (TP), False Positive (FP), True Negative (TN), and False Negative (FN) samples were counted. The ROC curve was plotted with the False Positive Rate (FPR) on the X-axis and the True Positive Rate (TPR) on the Y-axis. The formulas for TPR and FPR are listed as Equations (9) and (10), respectively.

$$\mathrm{T}\mathrm{P}\mathrm{R}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}$$

(9)

$$\mathrm{F}\mathrm{P}\mathrm{R}=\frac{\mathrm{F}\mathrm{P}}{\mathrm{F}\mathrm{P}+\mathrm{T}\mathrm{N}}$$

(10)

3.3. Deformation Monitoring Based on SBAS-InSAR

InSAR technology provides spatially continuous surface deformation information over short periods. To avoid long intervals, spatial incoherence, and the atmospheric effects of traditional interferometry measurement methods, and considering the practical requirements of monitoring slow linear and nonlinear deformations over long sequences, this study used the Small Baseline Subset (SBAS) InSAR technology for surface deformation monitoring in the study area [24,25,26].

The Small Baseline Subset Interferometry (SBAS-InSAR) is a new time series analysis method proposed by Berardino, Lanari, and others for studying low-resolution, large-scale deformations [27]. This method generates a time series of interferograms from multiple master images according to the principle of short temporal baselines and uses the Singular Value Decomposition (SVD) of matrices to invert and obtain the deformation sequence and average deformation rate for the observation period in the study area [28,29,30,31]. The key steps are as follows.

Firstly, $N+1$ Synthetic Aperture Radar (SAR) images covering the same study area were obtained, with acquisition times $t_0$, $t_1$, …, $t_n$. Appropriate temporal and spatial baseline thresholds we re set to co-register the other SAR images to the master image, resulting in $M$ interferometric pairs. $M$ satisfied the following condition:

$$\frac{N+1}{2}\le M\le \frac{N\left(N+1\right)}{2}$$

(11)

Next, multi-master time series interferograms were generated from the M differential interferometric pairs. The interferometric phase consists of a reference ellipsoid phase, topographic phase, deformation phase (along the radar line of sight), atmospheric phase, and noise phase. An external reference DEM was used to remove the topographic phase and effectively remove noise through multi-looking and filtering. The Minimum Cost Flow (MCF) method was applied for phase unwrapping to obtain the cumulative displacement along the radar line of sight.

A regression analysis was conducted of the deformation data set to estimate and remove elevation residuals. Noise and atmospheric delay residual phases were separated based on filtering combinations.

The short baseline set time series deformation model was used to reconstruct the deformation time series.

If the baseline set contains multiple subsets, the Singular Value Decomposition (SVD) method was used to obtain the unique solution, ultimately deriving the deformation time series for the pixel.

4. Landslide Hazard Analysis Using Joint SBAS-InSAR Technology

4.1. Model Building and Training

4.1.1. Model Building

Building on the theoretical foundation described above, this study implemented the BGA-Net model in code. The model includes two convolutional layers with a kernel size of 5 × 5, two pooling layers with a size of 3 × 3, one Flatten layer, one Dropout layer, and two fully connected layers. ReLU activation functions were used between each hidden layer, and Softmax was used for the classification of results. After multiple rounds of parameter tuning, the following optimal parameters were determined: Epochs set to 100, Learning Rate set to 0.001, Optimizer set to Adam, and Loss Function set to Categorical Crossentropy.

The Categorical Crossentropy function measures the difference between the true probability distribution $y$ and the predicted probability distribution $\widehat{y}$, and is commonly used in classification problems. The formula is:

$$L=-\sum _{i=1}^n{y}_i\log {\widehat{y}}_i$$

(12)

where $y_i$ is the true probability distribution and ${\widehat{y}}_i$ is the predicted probability distribution. In binary classification problems, the formula can be simplified to the following:

$$L=-\left[y\log \widehat{y}+\left(1-y\right)\log \left(1-\widehat{y}\right)\right]$$

(13)

Cross-entropy loss amplifies the impact of low-probability events through the logarithm function, making the model more accurate in predicting the correct class. During training, optimization algorithms minimize the cross-entropy loss to improve the model parameters.

4.1.2. Model Training

In the study area, there were a total of 338 known historical landslide points. A buffer zone of 1 km was established for each of these points. Corresponding negative sample points were randomly selected outside the buffer zones, resulting in a total of 676 sample points. A 54 × 54 area was cropped, centered on the sample points that were to be used for model training, and the corresponding labels (Landslide and Non-Landslide) were assigned. The data were then divided into training and validation sets in a 7:3 ratio, and finally, they were fed into the model for training.

4.2. Satellite-Borne SBAS-InSAR Surface Deformation Monitoring

In this study, we used the sun-synchronous orbit data of the Sentinel-1A satellite equipped with SAR sensor. Sixty-five ascending orbit images were generated for the monitoring period from 11 April 2017 to 24 June 2019, with a time interval of approximately 12 days. The deformation monitoring annual deformation rate map (m/year) for the Baige area is shown in Figure 5. The deformation monitoring annual deformation rate map (m/year) for the Baige area is presented, with the monitoring extent of ascending orbit shown in Figure 5 and the monitoring extent of descending orbit shown in Figure 6.

Based on the results of the ascending and descending orbit monitoring, a visual interpretation was conducted using Google Earth. Based on the level of deformation, the coordinates and deformation range of geological hazard points were determined. Ultimately, 98 geological hazard points were interpreted, and the file format was *.shp (a format used to store the location, shape, and attributes of geographic features). Among the identified hazard points, the ascending orbit deformation monitoring results indicate that the maximum annual deformation rate away from the line of sight (LOS) is 0.15 m/year, with a maximum cumulative deformation of −0.33 m; the maximum annual deformation rate toward the LOS is 0.07 m/year, with a maximum cumulative deformation of 0.15 m. The locations of the maximum deformation rate points are shown in Figure 5(a-1,a-2). Special attention was paid to the area near the large landslide in Baige, Tibet, and the deformation rate results are shown in Figure 7.

Based on the results of ascending and descending orbit monitoring, combined with optical images, a hazard investigation was conducted. The coordinates and deformation range of geological hazard points were determined based on the level of deformation; ultimately, 98 geological hazard points were interpreted, as shown in Figure 5a.

4.3. Results of Landslide Susceptibility Prediction (LSP)

In an LSP based on BGA-Net, the predisposing factors of landslide hazards contain a large amount of geographical surface information, such as geological, hydrological, and human factors. In this study, the model training parameters were set with 30 epochs, an Adam learning rate of 0.0001, and the categorical cross-entropy loss function. Based on the experience from relevant articles, the landslide disaster occurrence point dataset was split into a 7:3 ratio for model training and validation, respectively [32]. After repeated validations, it was found that this splitting ratio resulted in the best model fitting and the highest accuracy. A slope unit susceptibility assessment was conducted for the surrounding counties using the natural breakpoint method, which ensured that the data within each category were as similar as possible while the differences between different categories were maximized, facilitating analysis and visualization. The assessment results were categorized into five levels—extremely high, high, moderate, low, and extremely low—as shown in Figure 8. By separately counting the number of correctly and incorrectly predicted landslide and non-landslide samples in the validation set and using a confusion matrix, we can conclude that the overall accuracy and precision of LSP are 85.5% and 87% respectively. It can also be seen that high-risk areas are mainly concentrated around faults and rivers, showing a banded distribution. Compared to other areas, the southwestern side has a denser distribution. Low-risk areas are primarily distributed in the eastern part. Overall, the high-risk areas occupy a large proportion of the total, making landslide disasters more likely to occur.

4.4. Joint Risk Assessment Based on Landslide Susceptibility and Surface Deformation

The ascending orbit deformation rate (abbreviated as sg) obtained by SBAS-InSAR is classified into different deformation risk levels according to certain rules to represent the degree of deformation risk. The specific classification is as follows: |sg| < 0.001 m/year is classified as extremely low, 0.001 < |sg| < 0.005 m/year is classified as low, 0.005 < |sg| < 0.01 m/year is classified as moderate, 0.01 < |sg| < 0.02 m/year is classified as high, and |sg| > 0.02 m/year is classified as extremely high. Landslide risk is classified into extremely low-risk areas, low-risk areas, moderate-risk areas, high-risk areas, and extremely high-risk areas using the natural breakpoint method, which can distinguish between significant deformation and unrestricted deformation and lift and fall in the track line-of-sight direction. After consulting the relevant materials and experts, the following risk joint lookup table (

Table 3. SBAS-InSAR joint risk lookup table.

	Surface Deformation Hazard	Extremely Low	Low	Moderate	High	Extremely High
LSP
Extremely Low		Extremely Low	Extremely Low	Low	Moderate	High
Low		Extremely Low	Low	Moderate	Moderate	High
Moderate		Low	Low	Moderate	High	Extremely High
High		Moderate	Moderate	High	High	Extremely High
Extremely High		High	High	High	Extremely High	Extremely High

) was established. In theory, as long as the surface deformation of the region can be obtained in a timely manner, the risk zoning of the region can be obtained based on Table 3 within a short period of time, thereby achieving a dynamic early warning of landslides in the region [33].

The deformation rate map and LSP map of this area can be separated into five classes: extremely low, low, moderate, high, and extremely high. Then, ArcMap’s raster calculator can be used to add the two maps together. Finally, according to Table 3, the sum can be classified according to the standards needed to obtain the risk zoning map, as shown in Figure 9. In addition to the Baige landslide, some other high-risk areas are still visible, indicating the need for focused inspection and prevention measures in the remaining high-risk areas.

5. Discussion

5.1. Accuracy and Superiority of BGA-Net for LSP

To compare the prediction accuracy and performance of BGA-Net in landslide prediction in the study area, this study introduces Convolutional Neural Network (CNN) and Residual Network (Res-Net) models for comparative research on the fitting effect of landslide hazard in the surrounding areas of Baige. The CNN model and Res-Net model are used to predict the landslide probability at various locations in the study area, and a fitting effect comparison of the three models is shown in Figure 10.

The number of historical landslide points that fall within areas classified as High and Extremely High was statistically analyzed, and

Table 4. Statistics of historical landslide points in high- and extremely high-risk areas.

Model	High	Extremely High	Ratio (%)
BGA-Net	10	312	95.27
CNN	2	305	90.83
ResNet	17	253	79.88

was generated. From this table, it can be clearly seen that, compared to the other two models, the BGA-net model’s prediction results have the highest proportion of historical landslide points falling within high- and extremely high-risk areas, indicating the best fit.

The numbers of True Positives (TP), False Positives (FP), True Negatives (TN), and False Negatives (FN) were counted separately for the training and testing sets of the three models, and the confusion matrix was plotted as shown in Figure 11. From the figure, it can be seen that BGA-Net performed the best, with the least amount of FP and FN. There is not much difference between the training and validation sets, indicating the absence of issues such as overfitting or underfitting for BGA-Net. The other two models performed less satisfactorily.

As shown in

Table 5. Accuracy evaluation of LSP around Baige.

Model	Dataset Type	OA	Precision	Recall	F1-Score	AUC
BGA-Net	Training set	0.946	0.971	0.92	0.945	0.923
	Validation set	0.855	0.870	0.835	0.852	0.856
CNN	Training set	0.897	0.877	0.852	0.897	0.898
	Validation set	0.749	0.755	0.717	0.736	0.748
Res-Net	Training set	0.888	0.912	0.862	0.886	0.888
	Validation set	0.744	0.753	0.707	0.729	0.743

, the susceptibility to landslides in the surrounding area of Baige was evaluated for accuracy using five statistical evaluation criteria: Overall Accuracy (OA), Precision, Recall, F1-Score, and Area Under the Curve (AUC).

Five accuracy evaluation metrics were chosen for the validation set and a line chart was created (Figure 12) to more intuitively compare the differences between the models.

To more intuitively compare the accuracy of each model, the ROC curves for the three models were plotted, as shown in Figure 13.

From the overall sample ROC curve in Figure 13, it can be seen that the AUC value of the BGA-Net is also higher than that of the other two models, indicating that this model performs well in the overall sample.

According to Figure 10 and Table 5, it can be concluded that compared to the CNN and ResNet models, the BGA-Net model has the highest accuracy and stability in its predictions. It can accurately identify high-risk and low-risk areas, with good local prediction details and high accuracy in predicting high-risk areas. In terms of the accuracy evaluation, the BGA-Net training set scores for all five indicators are above 0.9, and for the validation set, they are all around 0.85, indicating high accuracy. In contrast, the CNN and Res-Net models perform less well, with all indicators for the validation set being below 0.8. Therefore, the comparative experiments prove the effectiveness and superiority of BGA-Net for LSP in practical use. From the overall sample ROC curve in Figure 13, it can be seen that the AUC value of the BGA-Net is also higher than that of the other two models, indicating that this model performs well in the overall sample.

5.2. Eeffectiveness of Landslide Risk Assessment Jointly Based on LSP and SBAS-InSAR

To validate the effectiveness of the landslide risk assessment method proposed in this paper, the disaster overlap index was used to verify the results of the risk assessment. Using optical images after the occurrence of the large landslide in Baige in 2023, nine landslides were extracted through visual interpretation (Figure 14 shows the optical image interpretation results of some landslide areas, where the yellow boundary areas indicate the interpreted landslide areas). All these landslides fell within the high-risk areas delineated in the joint risk mapping shown in Figure 9, thus demonstrating the effectiveness of the proposed landslide joint risk assessment method based on landslide susceptibility and surface deformation. The joint assessment method fully utilizes the advantages of both technologies, integrating geological hazard susceptibility and surface deformation data to enhance the comprehensive evaluation of landslide risk. The method demonstrates strong practicality, as the required data and technologies are accessible and operable at the current technological level. However, the method imposes high demands on the quality of the input data, where poor data quality may affect the accuracy of the results. Moreover, the method is based on certain model assumptions that require further validation. Subjectivity exists in geological hazard interpretation and risk assessment, especially in the interpretation of optical images, which may affect the objectivity of the results. In conclusion, while the joint method shows promising results in this study, further exploration and practice are needed when applying it to other regions, considering their specific circumstances.

5.3. Deformation and High-Risk Zonation-Forming Mechanism Inference

Based on the LSP results of the BGA-Net model, three areas were selected for detailed zooming (as shown in Figure 15) to study the potential causes and mechanisms of surface deformation and to correlate the observed deformation patterns with known landslide locations and geological structures.

By comparing Figure 9 and Figure 15, it can be seen that significant deformation areas and high/high-sigh-susceptibility zones have a high degree of overlap. Through a detailed analysis of disaster-causing factors, it is evident that areas with notable surface deformation and potential lanslide hazards in the Baige region are mostly located on both sides of the upper Jinsha River, at medium elevations (around 3000 m above sea level), and in areas with relatively steep slopes and large elevation differences. These topographic conditions are prone to erosion and traction-type deformation. Long-term erosion by river water at the slope base causes instability and rainwater infiltration under gravitational force leads to overall slope sliding. Furthermore, due to the generally gentle terrain along both sides of the river, some residential areas and roads are built along the mountains, and human activities such as slope cutting exacerbate the instability of the slopes. Additionally, most significant deformation areas are situated in the central part of the study area, where geological faults are densely distributed. Geological activities have exposed the rock mass, which is predominantly composed of metamorphic rocks (MA3/MB1). These rocks are prone to alteration due to rainwater erosion, making the rock mass loose and weak. When facing lubrication from rainwater and gravitational traction, they are prone to instability. In summary, landslide development in the Baige region exhibits typical characteristics of rock landslides in alpine valley areas.

6. Conclusions

This study constructed a BGA-Net network based on bilateral guided aggregation for LSP. Using multi-source disaster-causing factor data and landslide disaster point data for the surrounding area of the Baige section in the upper reaches of the Jinsha River, a large-scale LSP was conducted, with an accuracy of over 85%. Based on this, combined with the InSAR surface deformation monitoring results and the joint risk lookup table, a zoning map of large-scale landslide hazard risks was constructed, realizing the monitoring and risk assessment of large-scale landslides. The experiments show that the LSP model based on the BGA-Net network proposed in this study, as well as the landslide disaster risk joint zoning method based on surface deformation characteristics, have strong feasibility and practicality. This achievement can be applied to the practical disaster prevention and mitigation needs in the surrounding area of the Bai Ge section of the upper Jinsha River. The method accurately identifies areas at high risk of landslides, providing a reference for land use planning to avoid the construction of important facilities in landslide-prone areas and reduce future economic losses and casualties caused by natural disasters. Compared with traditional machine learning methods, the CNN-based landslide risk assessment has a higher differentiation ability for extremely low- and extremely high-susceptibility areas, which is more conducive for decision-making. Secondly, this study combined the landslide susceptibility with the time-series InSAR surface deformation monitoring results to achieve a comprehensive analysis of geological disaster risk at a regional scale. The innovation of this method lies in its enhancement of the scientific value and efficiency of evaluation, fully leveraging the advantages of both approaches. It is of great significance for promoting the modernization of landslide disaster assessment and risk management practices. However, this study is limited by the data, as it did not obtain ground monitoring station data. Combining remote sensing wide-area disaster monitoring with ground station monitoring technology to achieve the “point-to-area” combined monitoring of large-scale landslide risks is a direction that needs to be further improved in subsequent work. Furthermore, in subsequent work, we hope to further improve the prediction accuracy and generalization ability of the model by integrating bedrock monitoring data and exploring ensemble or hybrid model methods. We will also verify its universality and promote its application to other landslide-prone areas.