Rain-Induced Landslide Hazard Assessment Using Inception Model and Interpretability Method—A Case Study of Zayu County, Tibet

Abstract

Geological landslide disasters significantly threaten the safety of people’s lives and property. Landslides are a significant threat in Zayu County, Tibet, resulting in numerous geological disasters, including the 1950 earthquake that caused significant casualties and river blockages. More recent landslides have caused substantial economic losses and infrastructure damage, posing ongoing risks to the local population and their property. Landslide hazard assessment is a critical task in geological disaster prevention and mitigation. This study applied the Inception model to assess landslide hazard in the Zayu area. The Inception model excels at capturing multi-scale features efficiently through its architecture. Fifteen disaster-causing factors were selected as the primary indicators for landslide susceptibility assessment. On this basis, the Inception model was used for landslide susceptibility assessment. Combined with daily precipitation data in the Zayu area, the landslide hazard assessment of the “25 April 2010, heavy rainstorm in Zayu, Tibet” was completed. Back Propagation Neural Network (BPNN), Residual Neural Network (ResNet), Convolutional Neural Network (CNN), and Visual Geometry Group-16 (VGG-16) were introduced for comparison of the fitting effects, and SHapley Additive exPlanations (SHAP) was used for interpretability analysis. The comparative experimental results show that the Inception model performed best in landslide susceptibility assessment and is feasible in practical use. The results also show that the most critical factors in the model were topographic wetness index (TWI), normalized difference water index (NDWI), and road density. This study is significant for assessing landslide hazard in geological landslide disaster prevention and mitigation. It provides a reference for further research and response to similar disasters.

1. Introduction

Landslides are a prevalent geological hazard, posing a serious threat to people’s lives and property [1]. Landslide monitoring represents the most direct and effective method for studying the evolution of landslides. Notably, landslide risk assessment can provide a crucial decision-making basis for early warning systems, mitigating harm to lives and property and facilitating timely risk avoidance [2,3,4]. Therefore, reasonable and effective preventive research is paramount for the stable development of landslide-prone areas.

In recent years, deep learning (DL) has been widely employed in geographical research, with applications in image recognition [5,6], feature prediction [7], change detection [8,9], and more. In the context of landslide prediction, compared to the machine learning (ML) techniques extensively used in previous years [10,11,12], DL demonstrates superior capabilities in feature extraction and extensive data fitting, thereby enhancing the accuracy of remote sensing image processing [13,14,15]. Notably, BPNN and CNN have performed exceptionally well in landslide prediction [16,17,18].

There are 21 highly developed sections of landslides in Tibet, spanning a length of 1398 km and encompassing 434 landslides. The study area of Zayu County is situated in the high-altitude mountain canyon region between the Himalayas and the Hengduan Mountains, a high-incidence area for landslides [19]. This region is characterized by intense tectonic movement, frequent seismic activity, and concentrated rainfall, with an average annual precipitation of 801.1 mm, and predominantly consists of high mountains and canyons. These conditions have resulted in frequent geological disasters, particularly along the main road sections of the Zayu area, where high and steep slopes often cause damage, posing a threat to the lives and property of the local population. Therefore, conducting landslide disaster susceptibility and risk assessments in the Zayu area is significant.

Zayu County has long suffered from numerous casualties and property losses due to landslide disasters as a typical high-risk landslide area in the southwest region. According to statistics, 15 geological hazards developed in and around the Zhuhuagan Township of Zayu County, threatening 374 households and 2607 people, with threatened assets amounting to approximately CNY 408.63 million [20]. The Murdock-Chashum earthquake that occurred in 1950 induced a series of secondary geological disasters, including landslides, resulting in a cumulative death toll of more than 4500 and causing the blockage of the Yarlung Zangbo River and river breakage at multiple locations along its lower reaches [21]. On 25 April 2010, persistent heavy rainfall in Tsashum County of Tibet triggered geological hazards in some townships, leading to communication interruptions and road blockages. On 29 July 2016, a large landslide occurred on the road near the Tsatsum Highway in Tsatsum County, causing an economic loss of approximately USD 50,000.

The Inception model, initially proposed by Google’s research team in 2014, is a deep learning network architecture primarily used for computer vision tasks, such as image classification. Landslide influencing factors contain a large amount of geographic information, which is small in scale and large in quantity. The advantages of the Inception model, such as “multi-scale feature extraction” and “high computational efficiency,” make it well-suited for landslide hazard prediction.

Deep learning models inherently involve multi-layered complex nonlinear operations, often resulting in poor interpretability, commonly referred to as a “black box” approach [22]. However, geographic research necessitates spatial analysis [23], and understanding the causes and connections of disasters is crucial for disaster risk research and prevention. In recent years, Explainable Artificial Intelligence (XAI) models have emerged to provide explanations for input sets, aiding in comprehending deep learning models and identifying critical factors in landslide susceptibility assessment. Standard model explanation methods include SHAP, LIME (Local Interpretable Model-agnostic Explanations), and PDP (Partial Dependence Plot). SHAP is a prominent method inspired by Shapley’s values from game theory. It provides a systematic framework for quantifying and explaining the contribution of each feature in the model to specific predictions. SHAP not only provides an overview of global feature importance but also excels in providing local explanations, showing in detail how each feature positively or negatively influences the final prediction for a single prediction instance.

Therefore, this paper proposes a method to assess landslide risk by combining rainfall statistical data as dynamic triggering factors based on deep-learning landslide susceptibility. It constructs a joint risk lookup table to achieve landslide risk grading assessment. It also quantitatively explains the contributions of different samples and factors to the model-driven based on interpretable models. This paper takes Zayu County as the research area, conducts interpretable landslide risk assessment and analysis based on deep learning models, and proves the feasibility and effectiveness of the proposed method through experiments.

2. Study Area and Data Processing

2.1. Study Area

Zayu County is located in the southeastern part of the Tibet Autonomous Region, on the southeast edge of the Qinghai–Tibet Plateau, at approximately 97°27′ E and 28°24′ N. The county is characterized by steep terrain, overlapping mountains, and deep valleys. The average elevation is 2300 m, with a general tilting trend from northwest to southeast, resulting in a higher elevation in the northwest and a lower elevation in the southeast, with a relative height difference of 3600 m. The significant vertical height difference forms a typical high mountain canyon and mountain–valley landform. The valley’s southern edge is only 1400 m above sea level, while there are more than ten peaks above 5000 m, with the highest peak being the 6740 m Meili Snow Mountain. The area is traversed by dozens of major and medium rivers, including the Zayu River and the Nu River, featuring large drops, both tributaries of the Yarlung Tsangpo River. Due to differences in terrain elevation and precipitation triggers during the rainy season, severe geological disasters, such as landslides and debris flows, occur frequently, making it one of the critical areas for research and development of geological disaster prevention. The schematic diagram of the study area is shown in Figure 1.

2.2. Construction of Evaluation Index System

The evaluation index system of landslides is crucial in assessing landslide hazards and evaluating susceptibility. In this study, the landslide index system is divided into two major parts: intrinsic factors and extrinsic factors. Based on the geographical conditions of Zayu County and scientific selection principles, 15 evaluation factors, such as elevation, aspect, and slope, were selected [24]. The constructed evaluation index system is illustrated in Figure 2.

To ensure spatial consistency, the factor layers in this study were resampled to a uniform spatial resolution of 30 m. Furthermore, all factors were normalized to prevent the results from being influenced by the magnitude of the numerical values. The extraction and processing of factors were primarily conducted using the ArcGIS 10.2 and ENVI 5.3 platforms. The processed thematic maps of the factors are presented in Figure 3.

• Elevation
  Elevation significantly influences the regional climate, particularly in high-altitude mountainous areas where landslides are closely dependent on elevation [25].
• Slope
  Steeper slopes generate greater shear forces, increasing landslide likelihood.
• Aspect
  Variations in aspect lead to differences in regional climate conditions, indirectly impacting slope stability [26].
• Plane curvature and profile curvature
  Slope curvature affects the internal stress distribution, influencing landslide development.
• Terrain ruggedness index
  The Terrain Ruggedness Index (TRI) is crucial for describing the earth’s surface topography.
• Normalized vegetation index
  The vegetation index reflects the degree of vegetation coverage [27]. Vegetation coverage can mitigate erosion and the scouring of rocks by rainwater and reduce rainwater infiltration. Conversely, vegetation roots can intrude into rocks, compromising overall stability [28]. Therefore, vegetation can have either a positive or negative impact on landslides. The calculation formula is as follows:

  $$NDVI=\frac{NIR-R}{NIR+R}$$

  (1)

  where NIR and R represent the spectral reflectance in the near-infrared and red bands, respectively.
• Stream density
  River erosion and infiltration can reduce slopes’ shear resistance, thereby affecting their overall stability [29].
• Topographic wetness index
  The Topographic Wetness Index (TWI) evaluates soil moisture and quantitatively assesses runoff trends. The calculation formula is as follows:

  $$TWI=\ln \frac{\alpha }{\tan \beta }$$

  (2)

  where $\alpha$ is the length of the contour line per unit area, and $\beta$ is the slope.
• Normalized difference water index
  The Normalized Difference Water Index (NDWI) visually represents the distribution of rivers in the area. The calculation formula is as follows:

  $$NDWI=\frac{Green-NIR}{Green+NIR}$$

  (3)

  where NIR and Green represent the spectral reflectance in the near-infrared and green bands, respectively.
• Land cover type
  Different land use types affect slope stability differently [30].
• Fault density and distance to faults
  Faults disrupt the integrity of rock masses, thereby increasing the impact of other factors on slope stability.
• Road density
  A higher road density indicates more frequent human activities, leading to increased surface disturbances and potentially causing landslide disasters.
• Rainfall
  Rainfall increases groundwater levels, alters the mass distribution within the soil, and changes the slope stress, thereby affecting slope stability. Intense rainfall can cause substantial soil surface erosion, directly leading to landslides.

The data sources for each indicator are shown in

Table 1. Data sources for evaluation factors.

Factors	Source Data Description	Source
Elevation	GDEMV3 30 m	https://www.gscloud.cn/ (accessed on 22 April 2010)
Slope	Derivative dataset	Extracted from DEM
Aspect	Derivative dataset	Extracted from DEM
Plane curvature	Derivative dataset	Extracted from DEM
Profile curvature	Derivative dataset	Extracted from DEM
TRI	Derivative dataset	Extracted from DEM
TWI	Derivative dataset	Extracted from DEM
Stream density	River network distribution vector data	http://www.geodata.cn (accessed on 2 April 2010)
NDVI	Remote sensing image	https://www.gscloud.cn/ (accessed on 20 April 2010)
NDWI
Land cover type	30 m resolution global land cover	http://www.resdc.cn/ (accessed on 20 April 2010)
Road density	Road network data	http://www.webmap.cn (accessed on 20 April 2010)
Distance to fault	National active fault vector data	https://data.earthquake.cn/ (accessed on 2 April 2010)
Fault density
Rainfall	CHIRPS Pentad	https://developers.google.cn/earth-engine/ (accessed on 22 April 2010)

.

2.3. Factor Selection Based on Multicollinearity Diagnosis

The initially selected landslide factors may exhibit statistical linear correlations, potentially impacting the accuracy of the model results. A collinearity diagnosis was performed on the evaluation factors specified in Section 3.2 using IBM SPSS Statistics 27 to mitigate these errors. Factors with a tolerance more significant than 0.1 and a Variance Inflation Factor (VIF) less than 10 indicate weak collinearity between factors. The diagnosis results are presented in

Table 2. Collinearity diagnosis results of landslide evaluation factors.

Factors	Tolerance	VIF
Profile curvature	0.088	11.304
TRI	0.089	11.261
Elevation	0.263	3.808
Land cover type	0.414	2.417
Rainfall	0.469	2.134
Slope	0.469	2.133
NDVI	0.491	2.038
NDWI	0.607	1.648
Road density	0.746	1.341
Plane curvature	0.799	1.251
Fault density	0.815	1.227
TWI	0.865	1.156
Stream density	0.960	1.041
Aspect	0.978	1.022
Distance to fault	0.996	1.004

, revealing that the profile curvature and terrain roughness do not meet these criteria, indicating strong collinearity. Consequently, these factors were excluded from further analysis.

2.4. Data Representation Method for Three-Dimensional Deep Learning Models of Landslides

The primary challenge of integrating deep learning models into landslide risk prediction is constructing a suitable representation of landslide data for the model. This study adopted a three-dimensional representation of landslide data, which is more intuitive and less likely to lose data accuracy.

First, factors closely related to landslides, such as slope, aspect, and elevation, were selected. All raster grid rows and columns were aligned to 6737 × 10,053 pixels, with a uniform resolution of 30 m. The various factor layers were stacked and placed to form an array with a shape of (6737, 10,053, 15), as illustrated in Figure 4.

3. Methodology

Figure 5 depicts the technical route of this study. Firstly, relevant data in the Zayu area were collected and organized to form a factor dataset. Then, from the landslide dataset in the Zayu area, landslide factors required for this study were selected, and collinearity diagnosis was conducted to screen out the landslide evaluation factors suitable for susceptibility assessment. The factors were normalized and standardized as the input set.

This study employed the Inception model for landslide susceptibility and the “April 25, 2010, Tibet Zayu heavy rain” landslide risk assessment. The BPNN, ResNet, CNN, and VGG-16 models were introduced for comparative analysis [31]. The fitting effects of these models in this scenario were evaluated, and the prediction results of each model were assessed for accuracy. SHAP was used for model interpretation further to understand the decision-making process and logic of the models.

3.1. Inception Model

3.1.1. Model Introduction

The Inception model is an innovative deep convolutional neural network architecture proposed by the Google Brain team in 2014. Its core design concept is to explore how to maximize a neural network’s ability to capture features of different scales while maintaining computational efficiency. Its main advantages are as follows [32]:

• Multi-scale feature learning: The multi-scale convolutional structure of the Inception model can simultaneously extract fine-grained local features and coarse-grained global features from landslide-related data. This capability helps the model comprehensively capture the critical factors of landslides, improving prediction accuracy.
• Avoiding overfitting: The Inception model is a multi-path parallel structure, which helps to disperse the model’s learning ability, reduce excessive reliance on a single type of feature, and, to some extent, suppress overfitting.
• High computational efficiency: Landslide prediction tasks often involve large-scale spatial data analysis. Using 1 × 1 convolutions and modular design, the Inception model can effectively reduce the model’s parameter volume and computational complexity while maintaining high prediction performance.
• Its structure is shown in Figure 6.

3.1.2. Structure of Inception for Landslide Susceptibility Assessment

This study constructed an Inception model based on the structure of deep learning models, as shown in Figure 7. This model included an input layer, hidden layers, fully connected layers, and an output layer. The input layer converted the original image data into a digital form that the model could process. The hidden layers consisted mainly of convolutional layers and max-pooling layers. The fully connected layers integrated all the features extracted from the previous layer into a high-level abstract feature representation proper for landslide classification. The model output was binary classification using softmax, outputting “landslide” or “non-landslide” results.

3.1.3. Comparative Models

This study uses the Inception model to assess the susceptibility and risk of landslides in the study area. Additionally, it compares the fitting effects of four models, including BPNN, ResNet, CNN, and VGG-16.

• BPNN (Backpropagation Neural Network) is a neural network model that uses the “backpropagation algorithm” for learning and optimization. After the model completes a forward propagation calculation to obtain the output result, it compares the actual target output to calculate the error between them. Each layer of neurons adjusts its weights and biases based on the received error signal to minimize the loss function of the entire network. This weight update method based on the gradient descent strategy ensures that the network continuously optimizes and gradually improves the model’s prediction accuracy with each iteration.
• ResNet (Residual Neural Network) introduces the concept of residual blocks. Instead of directly mapping the input signal to the output through a series of layers in a residual block, it learns a residual mapping from the input to the production. Therefore, even if the network structure is intense in ResNet, it can effectively alleviate the gradient vanishing problem and avoid degradation in network performance as the depth increases [33].
• CNN (Convolutional Neural Network) is a deep learning model designed for processing image data. It is widely used in computer vision tasks, such as image classification and object recognition. Its core idea is to simulate certain characteristics of the human eye’s retina and brain cortex, automatically extracting features from raw pixel information [34].
• VGG-16 (Visual Geometry Group-16) is a deep convolutional neural network model proposed by researchers from the Oxford University Visual Geometry Group in 2014, characterized by its deep network architecture. All convolutional layers in VGG-16 use 3 × 3 convolutional kernels, a design choice that maintains a relatively small parameter while achieving a receptive field equivalent to larger kernels by stacking multiple 3 × 3 convolutions. This design increases the depth and non-linear expressive power of the network. Stacking small convolutional kernels is akin to performing multiple acceptable local feature extractions, which helps capture richer detail information. It has shown excellent accuracy in some landslide susceptibility studies [35].

3.1.4. Model Construction and Training Parameter Settings

This study used the Tensorflow platform for model training and prediction, with the evaluation units being grid units. Tensorflow has a rich neural network code library provides a complete technical solution for model training, prediction, error analysis, and other tasks. Grid units are like raster data structures in GIS, making it more intuitive and straightforward to import factor data into the model and facilitate fast reading and computation by computers.

Model training requires a dataset of positive and negative samples, where positive samples are historical landslide points and negative samples are points where landslides have not occurred. The positive samples in this study were directly obtained from the “Global Landslide Points and Landslide Areas Dataset (1915–2021)”, totaling 221 landslide points in the study area. The sampling method for negative samples mainly followed the technique for negative samples [36] in regional landslide risk assessments, randomly sampling 221 negative sample points outside a specific range buffer of the positive samples.

Before training the model, the landslide factor dataset was preprocessed to generate data that met the model’s input requirements. According to the First Law of Geography, objects closer together have a stronger correlation. In this study, we selected a region measuring 47 by 47 units centered around each sample point, both positive and negative, for cropping purposes. The landslide factor dataset was divided into multiple datasets of shape (47,47,15), labeled with corresponding “landslide” and “non-landslide” labels, shuffled, and divided into training and testing sets at a ratio of 8:2. These datasets were then imported into the model for multiple rounds of training and fitting validation. By setting different training parameters, the optimal parameter combination was eventually obtained. The data processing process is shown in Figure 8 below:

Through repeated experiments to improve model accuracy [37], the parameters for all four models were ultimately set as follows: learning rate of 0.001, number of training epochs of 25, and batch size of 16. The specific parameters are shown in

Table 3. Model parameters.

Model Name	BPNN	Inception	ResNet	CNN	VGG-16
Input Shape	(47,47,15)
Hidden Layer Structure	None	Convolutional Layer + Pooling Layer
Convolution Kernel Size	None	5 × 5			3 × 3
Pooling Kernel Size	None	3 × 3			2 × 2
Activation Function	ReLu
Learning Rate	0.001
Epoch	25
Optimizer	Adam
Loss Function	Categorical Crossentropy
Classification Function	Softmax

.

Based on the parameter configuration mentioned above, 150 independently trained models were batch-generated for each model type (Inception, BPNN, ResNet, CNN, VGG-16) to explore the adaptability and generalization performance of the different architectures for the “landslide susceptibility” task under the same training conditions.

3.2. Accuracy Evaluation Metrics

This study used five statistical metrics, including Overall Accuracy (OA), Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Sensitivity (SEN), and the ROC curve, to assess the feasibility of this study and select the best-performing model. The descriptions of each metric are shown in

Table 4. Descriptions of accuracy evaluation metrics.

Evaluation Metric	Variable Meaning	Formula Number
$OA=\frac{TP+TN}{TP+FP+TN+FN}$	TP (True Positive) represents the number of landslide samples correctly classified; FP (False Positive) represents the number of landslide samples incorrectly classified; TN (True Negative) represents the number of non-landslide samples correctly classified; FN (False Negative) represents the number of non-landslide samples incorrectly classified.	(a)
$TPR=\frac{TP}{TP+FN}$		(b)
$FPR=\frac{FP}{FP+TN}$		(c)
$MAE=\frac{1}{n}{\displaystyle \sum _{i=1}^n}{(T}_i-P_i)$	$T_i$ represents the true value of the $i$-th sample, and $P_i$ represents the predicted value of the $i$-th sample.	(d)
$RMSE=\sqrt{\frac{1}{n}{\displaystyle \sum _{i=1}^n}{{(T}_i-P_i)}^2}$		(e)

.

OA and SEN (TPR) have values between 0 and 1, with values closer to 1 indicating better model accuracy; RMSE and MAE values closer to 0 indicate more minor model errors.

The sensitivity TPR and False Positive Rate (FPR) are the coordinates of the ROC curve, and AUC is the area under the ROC curve used to measure the model’s performance. AUC values range from 0.5 to 1.0, with values closer to 1.0 indicating higher model accuracy.

3.3. Model Interpretation

The significance of model interpretation is to improve the transparency and credibility of deep learning models, facilitate understanding of the model’s decision-making process and logic, detect whether the model is influenced by biases in the training data, and judge the reasonableness of the results.

In this study, SHAP is used to explain the Inception model. The SHAP method aggregates local explanations into global explanations, which can better explain the contributions of factors to landslide susceptibility [38]. SHAP draws on the concept of Shapley values from economics, which quantifies the marginal contributions of features to model predictions. The SHAP library in Python provides rich visualization functions, such as “summary_plot” and “force_plot”, which intuitively display the contributions of different landslide factors to the model output.

4. Experimental Results and Model Interpretability Analysis

4.1. Susceptibility Evaluation

After 150 repeated trainings, the results of the best-performing models in terms of their accuracy metrics were selected for landslide susceptibility evaluation. The accuracy and sensitivity of the chosen models are shown in

Table 5. Accuracy and sensitivity of selected models.

Index	Dataset	Inception	BPNN	ResNet	CNN	VGG-16
OA/%	Training set	92.5	90.0	89.8	88.3	86.1
	Validation set	87.5	88.3	89.6	87.4	85.4
Sensitivity/%	Training set	95.0	90.0	92.5	90.0	87.5
	Validation set	90.0	87.5	87.5	92.5	87.5

.

4.1.1. Analysis of Inception Model Results

The Inception model was used to predict the landslide occurrence probability for each pixel in the study area, obtaining a landslide susceptibility evaluation map for the Inception model (Figure 9). The susceptibility levels were classified into five levels using the natural break method: low, moderately low, medium, moderately high, and high.

In the prediction results provided by the Inception model, most of the landslide points fall within the moderately high and high susceptibility levels, while other areas where landslides have not occurred are predicted to have low susceptibility levels. This demonstrates the model’s high level of fit accuracy. The Inception model exhibits strong discriminative ability, effectively identifying “high-risk areas” from “low-risk areas”, thus reducing the occurrence of false alarms and missed alarms. In conclusion, the Inception model demonstrates outstanding performance in landslide susceptibility assessment.

4.1.2. Comparison of Other Model Results

BPNN, ResNet, CNN, and VGG-16 models were used to predict landslide susceptibility in the study area, yielding the landslide susceptibility assessment results shown in Figure 10.

To visually demonstrate the fitting effect of each model on landslide susceptibility prediction, the number of historical landslide points in different risk-level zones and the proportion of landslide points in moderately high-risk and high-risk areas were calculated and are summarized in

Table 6. Number of historical landslide points in different risk-level zones.

Model	Low	Moderately Low	Medium	Moderately High	High	Landslide Ratio (%)
Inception	6	5	10	20	180	90.5%
BPNN	8	7	10	16	180	88.7%
ResNet	6	10	9	20	176	88.7%
CNN	6	10	6	35	164	90.0%
VGG-16	8	12	29	73	100	78.3%

.

Landslide predictions were made for positive and negative sample points using the four models’ results. The ROC curves were plotted with True Positive Rate (TPR) on the y-axis and False Positive Rate (FPR) on the x-axis (see Figure 11). The Area Under the Curve (AUC) of the ROC curve indicates the model’s accuracy, with a larger AUC indicating higher accuracy.

Combining Figure 10 and

Table 7. Accuracy evaluation of five models.

Index	Dataset	Inception	BPNN	ResNet	CNN	VGG-16
OA/%	Training set	80.213	72.326	68.568	80.014	70.431
	Validation set	78.951	71.610	66.427	78.277	68.869
Sensitivity/%	Training set	85.017	80.752	80.634	83.626	75.341
	Validation set	82.667	78.233	78.933	81.233	73.250
1/MAE	Training set	2.324	2.745	1.806	2.268	1.907
	Validation set	2.199	1.931	1.753	2.170	1.802
1/RMSE	Training set	5.023	3.812	3.107	4.862	3.598
	Validation set	4.751	3.522	2.979	4.603	3.212

, as well as the ROC curve in Figure 11, it can be seen that the Inception model performed the best in predicting landslide susceptibility. A total of 90.5% of historical landslide points fall into the high and very high susceptibility areas predicted by the Inception model. In areas where landslides have not occurred, the susceptibility levels predicted by the Inception model are generally lower, demonstrating high fitting accuracy. Additionally, the Inception model has the highest AUC value, at 0.986.

The ResNet model also exhibited excellent predictive performance, but more landslide points were predicted as low or moderate susceptibility, in contrast to the Inception model, indicating some shortcomings. The CNN and BPNN models performed slightly worse, with sound fitting effects, but still showing poor predictive accuracy in some areas.

Among the four models, the VGG-16 model had the poorest fitting effect. It predicted large areas of moderate to high susceptibility in areas where landslides have not occurred, and only 78.3% of the landslide points fall into high and very high susceptibility areas. There are also many landslide points in the low to moderate susceptibility areas predicted by this model, resulting in the smallest AUC value.

4.2. Model Training Accuracy Evaluation

Statistical evaluation of the accuracy metrics (OA, SEN, MAE, RMSE) was carried out for each of the four models in 150 independently trained instances. The average value of each metric was calculated and uniformly quantified: the reciprocals of the MAE and RMSE were taken so that the larger the value of the four accuracy metrics, the better. This facilitated a more intuitive comparison. The accuracy evaluation of the five models is shown in Table 7. In order to investigate whether there was an overfitting problem, the differences between each accuracy index of the training and validation sets were compared. It is evident that the differences between the training and validation sets are slight, which indicates that none of the models had an overfitting problem and that the results of this study are highly credible.

Among the five models, the Inception model had the highest scores in all metrics, indicating the best overall performance and stability. The CNN and BPNN models performed slightly worse than Inception but still show an acceptable overall performance. The VGG-16 and ResNet models had lower accuracy and 1/MAE and 1/RMSE scores, indicating poorer model stability with more significant fluctuations and lower sensitivity. There may be cases where they incorrectly identify some non-landslide areas as landslides. In general, the experimental results prove the feasibility and suitability of the Inception model for landslide susceptibility assessment.

4.3. Hazard Assessment

Based on the landslide susceptibility assessment results from the Inception model, a landslide hazard assessment was conducted for the “25 April 2010, heavy rainfall in Zayu, Tibet.” The rainfall dataset used was the “1961–2022 national daily precipitation grid dataset.” Rainfall data for the corresponding dates were extracted and converted into a unified resolution of 30 m, aligned with the row and column sizes of the model training factor dataset. The maximum, minimum, and average daily rainfall within the study area is shown in

Table 8. Details of rainfall from 22 to 25 April.

Date	Minimum Value/mm	Maximum Value/mm	Average Value/mm
22 April	3.09	26.41	8.21
23 April	5.08	118.60	80.74
24 April	5.26	94.78	46.33
25 April	4.65	50.49	18.75

. The daily rainfall data were accumulated to obtain the cumulative rainfall from 22 to 25 April.

The threshold for rainfall-induced landslides was divided into five levels: does not occur, beginning to occur, sporadic occurrence, group occurrence, and massive occurrence, corresponding to cumulative rainfall thresholds of 100 mm, 140 mm, 170 mm, and 200 mm [39]. The reclassified results of the cumulative rainfall grid are shown in Figure 12.

Based on the landslide susceptibility results predicted by the Inception model, a landslide hazard classification table (

Table 9. Landslide hazard classification table.

Landslide Susceptibility (Cumulative Rainfall)	High	Moderately High	Medium	Moderately Low	Low
Massive occurrence	H	H	H	MH	M
Group occurrence	H	H	MH	M	M
Sporadic occurrence	H	MH	M	M	ML
Beginning to occur	MH	M	M	ML	L
Does not occur	M	M	ML	L	L

) was established, categorizing landslide risk into five levels: low, moderately low, moderate, moderately high, and high. These are abbreviated as L, ML, M, MH, and H in the table. Landslide susceptibility was scored from top to bottom as 5–1; the cumulative rainfall was scored in the same way, and the two scores add up to 8–10 as high (H), 7 as moderately high (MH), 5–6 as medium (M), 4 as moderately low (ML), and 2–3 as low (L).

After processing the data in ArcMap, the landslide hazard distribution map for Zayu County from 22 to 25 April was obtained (see Figure 13).

Figure 13 shows that as the cumulative rainfall increases daily, the areas with moderate, moderately high, and high landslide hazard levels in Zayu County also increase. The areas with higher landslide hazard levels are mainly distributed in the southern part, where the cumulative rainfall is highest, and there are many areas with high and moderately high landslide susceptibility levels.

Overall, the risk of landslides in Zayu County began to increase on 22 April, with a maximum rainfall of 26.41 mm in the area. The maximum rainfall on 23 April was 118.60 mm, an increase of 92.19 mm compared to the 22 April. The maximum rainfall on April 24 was 94.78 mm, when the cumulative rainfall reached 239.79 mm, making the risk of landslides extremely high. After 25 April, the maximum rainfall showed a clear downward trend; the maximum rainfall on 25 April was 50.49 mm, a decrease of 44.78 mm compared to the day before, and the risk of landslides was significantly reduced. However, because of the lagging effect of rainfall on landslides, the landslide hazard remained in the “moderately high” or “high” category in most areas. According to a report from the China Weather Network, “Record-breaking heavy rainfall in Zayu, Tibet, resulted in geological disasters in some townships”, most landslides in Zayu County occurred between 24 and 25 April, consistent with the evaluation results.

5. Discussion

5.1. Individual Sample Model Driver Explanation

This study explains the landslide susceptibility assessment of the Inception model by selecting one positive sample point and one negative sample point and drawing different landslide factors to model the impact of the predicted results (Figure 14).

In the positive sample point, the Terrain Wetness Index (TWI) had the greatest impact on the model predicting positive results (landslides), followed by slope and elevation. Road line density had the greatest impact on the model predicting negative results (non-landslides). Overall, the driving force for predicting positive results was much greater than predicting negative results, so the model output for this sample point was a landslide. Similarly, as shown in Figure 14b, the factors with the greatest impact on positive prediction were slope and plan curvature, while road line density had the greatest impact on negative prediction. Therefore, the model output is non-landslide.

5.2. Global Sample Explanation

5.2.1. Sample Feature Analysis

The SHAP value of each feature was plotted for each sample. The results are shown in Figure 15. Each dot represents a sample, the X-axis represents the SHAP value, and the samples are stacked in the Y-direction. The dots are colored from red to blue, representing feature values from high to low, respectively. The larger the width of the red area (the length of the Y-axis direction), the more samples affecting the output of the result as “positive”, which means the stronger the positive impact of the feature, while the larger the width of the blue area, the stronger the negative impact. High values of slope affected the model’s positive output, and low values of most factors had a greater impact on the model’s output, which is consistent with the real landslide pattern.

5.2.2. Feature Importance Ranking

The importance level of each feature was obtained by calculating the mean absolute SHAP values for each feature. In this study, the “plot_type” parameter of the “summary_plot” function in the SHAP library was set to “bar” to generate a standardized bar chart for intuitive display (Figure 16). The most important features in the landslide susceptibility in Zayu County were TWI, NDWI, road line density, and slope, which had the greatest impact on the model’s output. In contrast, the contributions of the water system line density and distance to fault to the model output were not high.

6. Conclusions

This study used the Inception model for landslide susceptibility assessment and combined the daily rainfall-disaster factor to complete the landslide hazard assessment for the “25 April 2010 heavy rainfall in Zayu, Tibet.” The BPNN, ResNet, CNN, and VGG-16 models were introduced for comparative analysis, evaluating the fitting effects of these models for landslide susceptibility analysis. An accuracy evaluation index system was established to assess the prediction results of each model. Finally, SHAP was used to perform an interpretability analysis of the Inception model. The following conclusions were drawn:

• Regarding landslide susceptibility assessment, the Inception model had the highest accuracy and the best fitting effect, followed by the CNN and BPNN models, with ResNet and VGG-16 showing an average performance.
• The results of the landslide hazard assessment for the “25 April 2010 heavy rainfall in Zayu, Tibet” showed that the highest risk areas were mainly distributed in the southern part of the study area. The overall landslide hazard in the area started to increase on 22 April and peaked on 24 April. According to relevant news reports, most landslide events in Zayu County occurred between 24 and 25 April, consistent with the evaluation results.
• From the model interpretation results, in the Inception model, the most important influencing factor was TWI, followed by NDWI, road line density, and distance to fault, which made lesser contributions to the model output.
• Using advanced deep learning models to predict the potential risks of landslides in specific areas accurately is a highly feasible approach that combines scientific value with practicality. This method fully utilizes deep learning technology’s powerful data processing capabilities and advantages. This technology is expected to play a vital role in landslide prevention and control, as well as in improving disaster management capabilities. However, there are still many uncertainties in predicting the occurrence of landslide disasters. Therefore, it is necessary to combine various methods, such as remote sensing interpretation, to improve the overall accuracy of prediction.