Evaluation of landslide susceptibility of the Ya’an–Linzhi section of the Sichuan–Tibet Railway based on deep learning

The Qinghai–Tibet Plateau is an area with frequent landslide hazards due to its unique geology, topography, and climate conditions, posing severe threats to engineering construction and human settlements. The primary purpose of this paper is to map the landslide susceptibility of the Ya’an–Lin branch of the Sichuan–Tibet Railway using two deep learning (DL) algorithms, convolutional neural network (CNN) and deep neural network (DNN). Initially, a geospatial database was generated based on 587 landslide hazards determined by Interferometric Synthetic Aperture Radar (InSAR) Stacking technology and field geological hazard surveys; thus, 18 landslide-influencing factors were selected. Subsequently, the landslides were randomly divided into training (70%) and validation data (30%) for model training and testing. Next, a Pearson correlation coefficient and information gain (IG) method were used to perform the correlation analysis and feature selection of the 18 influencing factors. Afterward, landslide susceptibility maps were generated for the two models. Finally, the performance of the model is validated using the receiver-operating characteristic (ROC) curve and confusion matrix. The results show that the CNN model (AUC = 0.88) provided better performance in both the training and testing phases compared to the DNN model (AUC = 0.84). In addition, the high landslide susceptibility is primarily distributed in the Jinsha, Lancang and Nu River basins along the railway. The slope, altitude and rainfall are the main factors for the formation of the landslides. Furthermore, the two deep learning models can accurately map the landslide susceptibility, providing important information for landslide risk reduction and prevention.


Introduction
Landslides are one of the most destructive natural disasters and are caused by a combination of natural and human factors (Kjekstad and Highland 2009). Under the influence of extreme climatic events, an increasing number of landslides are occurring worldwide resulting in significant economic and human losses (Huang et al. 2013). The Sichuan-Tibet Railway was built on the eastern margin of the Qinghai-Tibet Plateau, where the plates collide and are structurally active. It is one of the regions where crustal deformation and tectonic activities are still extremely intense Lan et al. 2021). The Sichuan-Tibet Railway traverses the most complex geological, topographic, and topographical areas in the world, and the plate structures along the line are the most active. The active faults are dense within this region, topographic changes are significant, and geohazards such as landslides, collapses, and debris flows are the most developed (Guo et al. 2017). Preliminary investigation 1 3 250 Page 2 of 18 results indicate that a total of 3043 geohazards such as collapses, landslides, and debris flows have been found in the Ya'an-Linzhi section of the Sichuan-Tibet Railway (Guo et al. 2021). In 2000, a substantial high-speed landslide occurred on the Yigong Zangbo River in Bomi County, Tibet blocking the river and forming a barrier lake of 3.0 × 10 8 m 3 ; the debris flow caused by the dam break washed away the Tongmai Bridge about 17 km downstream, resulting in huge casualties and property losses (Zhou et al. 2016). Two large sequential landslides formed a dam and the resulting lake along the Jinsha River on October 11 and November 3, 2018. Approximately 24 and 9 × 10 6 m 3 of material collapsed and rushed into the river (Ouyang et al. 2019). Due to the large inflow rates at the time of damming, the barrier lake level rose rapidly, destroying the Jinsha River Bridge and other downstream coastal transportation facilities, posing considerable risks to the downstream residents and properties ). All of these disasters result in enormous damage to roads and the surrounding environment, causing serious threats and impacts to the construction and safe operation of major cross-river transportation projects such as the Sichuan-Tibet Railway. However, landslide susceptibility mapping could provide estimations on where the landslides are likely to occur, effectively improving disaster forecasting efficiency (Dahal et al. 2008;Skilodimou et al. 2019;Wu et al. 2020). Landslide susceptibility is an assessment of the spatial distribution of the probability of a landslide occurrence within a certain area based on local geological environmental factors (Azarafza et al. 2018;Kadavi et al. 2018). Therefore, it is imperative to evaluate the geological disaster susceptibility of the Sichuan-Tibet Railway Ya'an-Linzhi section, which can provide a scientific basis for the construction of the Sichuan-Tibet railway as well as disaster prevention and mitigation in future safe operations.
Landslide susceptibility evaluation techniques are divided into knowledge-driven and data-driven techniques (Aditian et al. 2018). Knowledge-driven depends on the experience and knowledge of experts to evaluate landslide susceptibility, which has certain subjectivity and the limitation of a small evaluation range (Rozos et al. 2011;Bathrellos et al. 2012). With the rapid development of computer technology and geographic information systems, data-driven methods have been widely used in regional landslide susceptibility evaluation, such as the weight of evidence model (Ilia and Tsangaratos 2016), logistic regression model (Ayalew and Yamagishi 2005;Zhao et al. 2017), neural Network model (Yilmaz 2009;Lucchese et al. 2021), support vector machine model (Yao et al. 2008;Huang and Zhao 2018;Yu et al. 2019), and the integrated learning algorithm (Chen et al. 2014b;Hong et al. 2019;Wang et al. 2021;Sun et al. 2021;Hussain et al. 2022). However, with continuous in-depth research on machine learning, it was found that deep learning algorithms have more levels of non-linear operations than "shallow learning" methods such as single hidden layer neural networks, and support vector machines. In addition, the ability to build advanced features encourages the discovery of the deepest connection between the parameters, which generally obtains a robust performance for non-linear processing Reichstein et al. 2019). DL models, especially the convolutional neural network (CNN) model and the deep neural network (DNN) model, have gradually been applied to natural disaster susceptibility assessment, and have achieved excellent results (Wang et al. 2020a;Mahdi et al. 2021). Both the adopted model and the completeness of the landslide list in the study area affect the accuracy of the landslide sensitivity evaluation. However, the complete landslide hazard in the study area cannot be obtained through the field survey of geological hazards and remote sensing image interpretation, which has an impact on the results of the landslide sensitivity evaluation. In this study, the InSAR Stacking technology is used to identify hidden landslide points in the study area; thus, a more complete list of landslides can be obtained and the evaluation results are more accurate.
In this study, the primary purpose is to map the landslide susceptibility of the Ya'an-Lin section of the Sichuan-Tibet Railway using two deep learning algorithms, CNN and DNN. This study innovatively applies the landslides identified by the InSAR Stacking technology to landslide susceptibility evaluation and explores the potential application of the CNN and DNN deep learning models within the landslide susceptibility evaluation of the Ya'an-Linzhi section of the Sichuan-Tibet Railway. In addition, a comparison of their overall performance was made. This paper provides new ideas and valuable information for related research on landslide susceptibility evaluation in other regions.

Study area
This study area is identified by the 25 km buffer of the Ya'an-Linzhi section of the Sichuan-Tibet Railway. The area is 50957 km 2 , which is approximately 1011 km long (Fig. 1). This area is primarily affected by the warm and humid air currents of the Pacific and Indian Oceans. The regional differentiation of climate along the route is apparent. Along the line, it transitions from a mid-subtropical climate zone in the Sichuan Basin to the plateau sub-temperate humid-sub-temperature-humid zone and the plateau temperate sub-humid-sub-arid zone. The annual average temperature and annual rainfall decreases from east to west as the altitude increases. The vertical zoning characteristics of the climate zone of the Qinghai-Tibet Plateau are obvious, with significant temperature differences between winter and summer, day and night, as well as strong freeze-thaw weather.
The topography and geomorphology along the Sichuan-Tibet Railway are complex and highly variable. It passes through 5 geomorphological regions, the Sichuan Basin, West Sichuan Alpine Canyon, West Sichuan High Mountain Plain, Hengduan Mountains in Southeast Tibet, and Southern Tibet. The railway traverses the Hengduan, Nyainqentanglha, Himalayan Mountains, and other mountains, across the Dadu, Jinsha, Nu, and Yarlung Zangbo Rivers. Active faults and strong earthquakes along the line are frequent, such as the Longmenshan, Xianshuihe, Jinshajiang, Lancangjiang, and Nujiang fault zones. The active fault zone controls this area's topography and geomorphology and plays a vital role in controlling the distribution of earthquakes. The formations along the route are diverse and are controlled by geological structures. Except for the Cambrian, it is distributed from the Quaternary to the Sinian. The main lithologies are sedimentary and metamorphic rocks dominated by sandstone, slate, and phyllite, dominated by granite, and soluble rock dominated by limestone.

Materials and methods
There are four stages using the CNN and DNN models for landslide susceptibility mapping: (1) the establishment of a spatial database, including InSAR Stacking technology and a field survey to generate a list of landslides, as well as selecting the landslide impact factors; (2) assessing data accuracy and the removal of noisy data with null prediction power; (3) use CNN and DNN models to generate landslide susceptibility maps; (4) validation and comparison of the two models (Fig. 2).

Landslide inventory map
Landslide inventory maps are prepared for multiple scopes, which is the first step toward modeling landslide susceptibility (Guzzetti et al. 1999;Rosi et al. 2018). This study combines the results of InSAR Stacking deformation , using Google Earth satellite images and field surveys to prepare a landslide inventory map (Fig. 1). Consequently, a total of 587 landslides were identified in the inventory map. The area of landslides in the study area is 691 km 2 , and the largest and smallest landslides are 19688941 m 2 and 1152 m 2 , respectively. The types of landslides are divided into loose accumulation layers and rocky landslides. The loose accumulation layer landslides are primarily smallscale shallow landslides, the rocky landslides are made up of large and medium-sized deep landslides, and the distribution in the region consists of small loose accumulation layer landslides. Figure 3 shows the hidden danger points of the  Using machine-learning methods to model landslides is typically a binary classification (Wang et al. 2020a). Therefore, it is necessary to use positive samples (landslides) and negative samples (non-landslides) for modeling. Within the inventory map, 587 pixels of landslide occurrences have been extracted. In this study, to avoid the error rate of non-landslide selection, an equal number of non-landslide points were randomly selected out of the landslide buffer area. According to previous research, the results of the sample ratio of 7:3 for training and testing are the most reasonable (Juliev et al. 2019;Pham et al. 2020;Fang et al. 2021). Therefore, the landslide and non-landslide samples were randomly selected 410 (70%) for model training, and the remaining 177 (30%) were used for model testing.

Landslide-influencing factors
The cause of landslide occurrence is complex, and their mechanisms are still under debate. Generally, landslides result from a combination of internal geological conditions and external environmental factors (Zezere et al. 1999;Xiang et al. 2010). Internal factors include topography, stratigraphic lithology, geological structure, and tectonic movement (Yin et al. 2010). The external factors of landslides can be divided into human and natural factors. Natural factors include meteorological hydrology, hydrogeology, weathering, and new tectonic movements (Capitani et al. 2013). Human factors refer to human engineering activities, including constructing roads, buildings, factories, and the mining of minerals. In general, the external factors are the inducing factors of landslide occurrence (Huang 2007;Zhuang and Peng 2014).
To produce a reliable susceptibility map, selecting appropriate landslide conditioning factors is essential. Previous studies confirmed the relationship between the conditioning factors and landslide occurrence (Reichenbach et al. 2018). In the study area, the tectonic activity is intense, the large-scale tectonic faults are densely distributed, the surface is severely cut and broken, the landform features of deep mountains and canyons are widely distributed, near-upright cliffs and "V"-shaped valleys are common, and the landslide disasters are densely distributed, mainly in shallow small landslides dominate. Therefore, the formation of landslides is affected by factors such as topography, stratigraphic lithology, geological structure, hydrological conditions, vegetation coverage, and human engineering activities; therefore, 18 factors affecting landslides have been selected, including elevation, slope, aspect, plan curvature, profile curvature, terrain surface convexity (TSC), terrain ruggedness index (TRI), surface cutting degree (SCD), topographic wetness index (TWI), stream power index (SPI), sediment transport index (STI), lithology, distance to faults, land use, normalized difference vegetation index (NDVI), rainfall, distance to rivers, and distance to roads (Fig. 4). Since the 18 factors are represented on different intervals or scales, all factors are converted into a grid with a DEM resolution of 30 m × 30 m for unification. Furthermore, all factor datasets can be divided into either continuous or discrete datasets. A continuous dataset of each factor was reclassified into discrete subclasses with data in each specific interval using a manual method; discrete datasets of the rest factors were classified using original natural grouping. The detailed information of the classes of each landslide conditioning factor is shown in Table 1.
Each factor has a different effect on the occurrence of landslides. Elevation has a significant impact on landslide development and determines the potential energy of the landslide; it also affects the movement characteristics of the landslides (Pourghasemi et al. 2012a). An increasing slope angle will cause an increase in the size of the free face and shear strength on the potential slide surface, resulting in slope failure . The slope aspect determines the illumination time received by the slope surface. There are differences in surface humidity, vegetation coverage, and different slope aspects, which affect the distribution of pore water pressure and the physical and mechanical characteristics of the rock and soil masses (Pourghasemi et al. 2012b). Plan curvature affects convergence and divergence of flow. Profile curvature has great significance on the acceleration and deceleration of flow providing valuable information about erosion and deposition (Chen et al. 2014a;Wu et al. 2020). TSC describes the relief characteristics of the terrains surface (Tolga 2019). TRI is a measure of the roughness and brokenness of the ground. The larger the roughness means the ground is broken, and the loose deposits are richer, which is conducive to the occurrence of a landslide (Korzeniowska et al. 2018). SCD refers to the differences between the average and minimum value of the elevation of a point on the ground within a specific area, reflecting the degree to which the ground surface is cut (Zhang et al. 2020). TWI comprehensively analyzes the influence of topographical features on the spatial distribution of soil moisture (Mousa et al. 2018). SPI indicates the erosion power of streams which might affect landslide occurrences (Moore and Burch 1986). STI describes topographic variables of water and sediment transport in landslides (Pourghasemi et al. 2012a). Lithology is one of the basic factors affecting the occurrence of landslides (He et al. 2012;Nsengiyumva et al. 2019). According to the hardness and type of lithology, it is divided into the following eight groups, A (Harder sandstone, siltstone), B (Weaker gneiss, phyllite, mudstone), C (Soft and hard limestone interbedded with sandstone), D (Harder quartz sandstone, feldspar quartz sandstone), E (Hard basalt, ophiolite, syenite), F (Hard granite, diorite), G (Soft and hard silty slate, conglomerate sandstone), and H (Weak loose deposits). Faults have a significant influence on the strength of the rock mass, the development of the terrain structure, and the slope's stability (Pham et al. 2015). Land use influences slope stability by changing land use and disturbing the slope stability conditions (Pham et al. 2016). NDVI represents vegetation coverage and groundwater content, which may affect the development of landslides ). Rainfall causes a large amount of rainwater to infiltrate, saturating the soil layer on the slope, increasing the weight of the sliding body, thus causing the occurrence of landslides (Zhuang et al. 2016). When building roads, natural slopes must be excavated and repaired, which will inevitably interfere with the balanced conditions of the original slope, often leading to unstable slopes and landslides (Pham et al. 2017). Distance to rivers is one of the conditioning factors that has an effective role in landslide stability. The wet saturated water of the river acting on the sliding area and part of the sliding body may reduce the shear strength of the soil and weaken the layers, thus, reducing the stability of the landslide (Ding et al. 2017;Erener et al. 2017).

Correlation analysis
When evaluating landslide susceptibility, it is essential that the influencing factors maintain mutual independence. If there appears to be a strong linear correlation in the factors mentioned above, then the predisposing factors are assumed to exist within a multicollinearity problem. The multicollinearity problem will affect the accuracy of the training model and may lead to errors in the prediction results Chen and Chen 2021). This paper uses a Pearson correlation coefficient to analyze the correlation between the influencing factors. Its value ranges from − 1 to 1. − 1 means that the two variables are completely negatively correlated, 1 means that the two variables are completely positively correlated, and 0 means that they are not correlated. When the absolute value of the correlation coefficient between two factors is greater than 0.7, it is considered to have a high correlation (Martín et al. 2012;Hong et al. 2020).

Information gain
In this study, the feature selection method of information gain (IG) was used to select an optimal subset to improve the prediction performance in the evaluation of landslide susceptibility (Dash and Liu 1997). The information gain is determined by calculating the entropy reduction of the output category y to which the input factor x i corresponds: where E(y|x i ) is the conditional entropy and E(y) is a priori Shannon entropy; therefore, they are calculated as follows: The average merit (AM) derived from this method uncovers the importance between conditioning factors and landslide occurrence. The greater the weight, the greater the contribution of the corresponding factors to the occurrence of landslides. If this value is less than or equal to 0, then this influencing factor has nothing to do with the occurrence of landslides and should be excluded when making predictions.

Frequency ratio analysis
The frequency ratio (FR) method can be employed to evaluate the correlation between landslide occurrence and the influencing factors. In the landslide susceptibility analysis, it is perceived that future landslides will occur under the same conditions as past landslides. The FR can be calculated as follows (Chen et al. 2014a): where N is the number of each factor's landslide; N ′ is the number of total landslides; A is the number of pixels in a particular class; and A ′ is the number of total pixels.

Landslide susceptibility models CNN
A CNN model exhibiting robust performance in visual image analysis is a class of feed-forward neural network whose artificial neurons respond to a portion of the surrounding elements (Girshick 2015). The general CNN model structure includes an input layer, convolutional layers, maximum pooling, fully connected layers, and an output layer (Shin et al. 2016), as shown in Fig. 5. The convolutional layer uses a sliding convolution window method to extract features from the input layer. The first convolutional layer where N is the number of factors affecting the landslide, f represents a non-linear activation function and * denotes the convolutional operator, k is the number of convolutional kernels, and w j and b j denotes the weight and bias, respectively. (5) The pooling layer is used to realize the sample processing of the feature map, which can reduce the amount of data while retaining useful information, preventing over-fitting and improving the generalization ability of the model. Next, these local representations extracted by the convolutional and pooling operations are reorganized through the fully connected layers. Finally, the fully connected layer is connected to the output layer, which consists of two neurons representing landslide and non-landslide. The parameters in the CNN layer are optimized using a back-propagation algorithm.  . The general DNN model structure includes an input layer, several hidden layers, and an output layer. In this architecture, the neurons (nodes) in the previous layer are completely connected to all the neurons in the next dense layer. Afterward, more dense layers are added to extract hidden information in the learning process. Figure 6 presents the architecture of DNN. The basic processes of deep NN mechanisms are as follows: (1) the network correctly assigns inputs to their associated targets, (2) introduce a loss function to calculate the prediction and true target of the network, and (3) the training loop is repeated enough times to generate weights that minimize the loss function.
In this study, the DNN model was applied to the evaluation of landslide susceptibility. The impact factor became the input signal received in the first layer and analyzed in the hidden layer. Finally, the prediction results are displayed in the output layer as landslide and non-landslide. The structure of the DNN model was determined through several trial-anderror methods, which consisted of a model of three hidden layers, including 16 neurons, 2 output neurons, and 3 hidden layers of 64 neurons.

Model evaluation methods
In landslide susceptibility mapping, it is essential and necessary to validate the model's performance. In this paper, the receiver-operating characteristic curve (ROC) is used to evaluate the model's training and prediction accuracy. The ROC curve is an indicator of the continuous variables of data specificity and sensitivity. The area under the ROC curve (AUC) represents the accuracy of the model; the closer the AUC is to 1, the better the model performance (Wu et al. 2020;Sameen et al. 2020). Simultaneously, a confusion matrix was used to evaluate the performance of the two models. Statistical indices including accuracy (ACC), recall, precision, and F-measure (F 1 ) were acquired from the confusion matrix. The calculation is as follows: where TP (true positive) and TN (true negative) are the numbers of correctly classified landslides, and FP (false positive) and FN (false negative) are the numbers of landslides incorrectly classified. For ACC, recall, precision, and F 1 , these values are between 0 and 1. With increasing numbers, the model's performance improves.

Selection of landslide-influencing factors
The Pearson correlation coefficient method was used to analyze the correlation of 18 landslide-influencing factors within the study area. The results are shown in Table 2. It can be observed that the correlation coefficient between the factor SCD and TSC is 0.94, the correlation coefficient between the factor TRI and slope is 0.92, and the correlation coefficient is greater than 0.7; thus, there is a high correlation between the factors. Therefore, the SCD and TRI were removed from the initial factors to improve the data quality.
The remaining 16 factors with less correlation were evaluated for landslide susceptibility within the study area. Landslide susceptibility is affected by various influencing factors. In this study, the information gain method was used to predict the contribution weight of 16 factors to the occurrence of landslides. The invalid influencing factors (with a weight of 0) do not participate in landslide susceptibility evaluation. The predicted results are shown in Fig. 7, showing that the AM of the selected factors is all greater than 0, which contributes to the occurrence of landslides in the study area. The highest AM factor is slope (0.242), which is the dominant factor to induce landslide occurrence. Second, the AM values of elevation, rainfall, and topographic relief are between 0.1 and 0.2, which influence the landslide. The remaining AM values are between 0 and 0.1, indicating that they have little contribution to the landslide occurrence.

Evaluation of landslide susceptibility
In this study, the TensorFlow framework of Python was used to construct the CNN model. The main parameters of the model are set as (optimizer: AdamOptimizer; number of iterations: 1300; activation function: ReLU; number of convolutional layers: 2; Dropout: 0.5), and the rest of the parameters are default values. Input the training sample data to train the CNN model and select the mean square error (MSE) as the loss function. After approximately 1300 epochs, the MSE value changes very little, and the CNN model converges, as illustrated in Fig. 9. Finally, the trained model is used to predict the landslide susceptibility index within the study area, based on the areas' visual and easy interpretation and comparison. The susceptibility was classified into five categories: 10%, 10%, 10%, 20%, and 50% (from high to low), corresponding to very high, high, middle, low, and very low susceptibility regions, respectively (Pradhan and Lee 2010;Sun et al. 2020). Figure 10 shows the landslide susceptibility map of the Ya'an-Linzhi Section of the Sichuan-Tibet Railway. It was observed that high susceptibility areas are primarily distributed in the Jinsha, Lancang, and Nujiang River Basins along the railway, which is consistent with the historical landslide distribution.
The DNN-based approach was implemented using the Python package Keras with Tensorflow as the backend. The main parameters of the model are set as (optimizer: RMSProp, number of iterations: 45, activation function: ReLU, hidden layers: 3, neurons: 64, batch_size = 40, verbose = 1), and the rest of the parameters are default values. Similarly, input the training sample data set into the constructed DNN model for training, and select the mean square error (MSE) as the loss function. After approximately 45 epochs, the DNN model converges, as illustrated in Fig. 9. Input 16 factors into the trained DNN model for prediction and obtains the landslide susceptibility index within the study area. The susceptibility was classified into five categories: 10%, 10%, 10%, 20%, and 50%. Figure 11 shows the landslide susceptibility map produced by the DNN model for the Ya'an-Linzhi Section of the Sichuan-Tibet Railway, overlaid with landslides. The distribution results of high and very high susceptibility areas are more consistent with those predicted by the CNN model. This indicates that the landslide susceptibility map matched well with the distribution of the actual historical landslides.

Model validation and comparison
The results of the ROC curve evaluating the goodnessof-fit and the prediction performance of the two models are shown in Fig. 12, and the confusion matrix is shown in Table 3. It can be observed that the AUC of the training data of the CNN and DNN models are 0.99 and 0.98, Fig. 9 The variation of MSE values during the training process respectively. The two deep learning models have a significantly better performance on the goodness-of-fit to the training data (success rate); the CNN model has the best performance (99%), followed by the DNN (98%). Using test data to verify the prediction performance of the two deep learning models, the AUC of the CNN and DNN models are 0.88 and 0.84, respectively. It can be observed that both models show higher predictive power; however, the predictive power of the CNN model is higher than that of the DNN model. In addition, the ACC, recall, precision, and F1 of the confusion matrix are used to validate the test data of the two models. The results are shown in Table 3; these metrics ACC, recall, precision, and F 1 of the CNN model were 84.68,84.48,84.39,and 84.64,respectively,while those of the DNN model was 79.48,81.88,75.72,and 78.68,respectively. All the metrics revealed that although both models demonstrated reasonable goodness of fit, the CNN model performed better in terms of the training and test datasets. Therefore, the CNN model had a better prediction than the DNN model in this case.
The two models were compared to the landslide density (number/km 2 ) quantitative analysis based on the predicted landslide susceptibility zoning map and historical landslides. The results are shown in Table 4, it can be observed that in the landslide susceptibility maps generated by the two models, as the landslide susceptibility increases, the

Discussion
Landslide susceptibility maps are essential for decisionmakers to formulate reasonable policies and reduce the impact of landslides (Wang et al. 2020b). Therefore, it is of great significance to obtain high-quality landslide susceptibility maps (Guzzetti et al. 2012;Chang et al. 2019). However, with insufficient data, these machine-learning models often suffer from generalizing to areas other than the training area. Especially in landslide susceptibility mapping, gathering inventory data is expensive, and it is difficult to collect a complete list of landslides. To solve this problem, this paper uses InSAR Stacking technology to identify early landslide hazards in the Ya'an-Linzhi section of the Sichuan-Tibet Railway and uses the identified landslide and historical landslide data as modeling data to evaluate the landslide susceptibility. Zhao et al. (2019) used a combination of landslide data identified by InSAR Stacking technology and historical landslides to map landslide susceptibility. The study found that the optimized results of InSAR Stacking technology were more reliable than the results of only the historical landslides. The slopes deformation identified by the InSAR Stacking method is usually a precursor to the occurrence of landslides. In a time series analysis, the slope deformation rate is an accelerating process, and indicates the occurrence of a landslide. Therefore, InSAR Stacking deformation monitoring results can provide an important basis for early identification and susceptibility evaluation of landslides, making the results of model predictions more reliable. The choice of a prediction model has an important influence on the results of landslide susceptibility evaluation. Many scholars have conducted extensive research on the application of deep learning in landslide susceptibility evaluation, and have achieved numerous results (Bui et al. 2020;Dong et al. 2020;Mandal et al. 2021). Fang et al. (2020) integrated the CNN model with traditional machinelearning classifiers, and their results show that integrating CNN techniques can effectively improve the performance of the above machine-learning classifiers. Through research investigating the affected size of the model accuracy under different sampling strategies, it was found that the prediction accuracy of deep learning is not affected, while the prediction accuracy of other traditional machine-learning models is greatly affected by the sampling strategy, and the prediction accuracy of the deep learning DNN model highest precision ). In addition, many scholars have used deep learning to map landslide susceptibility in other regions, and compared with other traditional machinelearning models found that the deep learning model has better performance Azarafza et al. 2021;Nguyen and Kim 2021).
In this study, the CNN and DNN models were used to evaluate the landslide susceptibility in the study area, and there ROC curves were used to verify the training accuracy (success rate) of 0.99 and 0.98, respectively, and the prediction accuracy of 0.88 and 0.84, respectively. The results show that the two models have good predictive ability and can obtain high-quality landslide susceptibility maps. Although the two deep learning models have achieved good results, there are still some limitations. Due to the lack of theoretical basis, the selection of hyperparameters and network design is also a considerable challenge, the key parameter values of the model are determined by a trial-and-error method, and the parameters determined by this method may not be the best model. There are also other scholars who have the same problem when using deep learning models for landslide sensitivity mapping, and therefore, have not obtained optimal parameter values (Bui et al. 2020;Yao et al. 2020). In addition, when using CNN and DNN in landslide susceptibility mapping, the problem of over-fitting usually occurs due to the insufficient training data for landslides, which requires a certain data volume Ngo et al. 2021). In future research, we can explore the method of optimizing the parameters in the deep learning model to further improve the accuracy of the model.

Conclusions
In this study, two well-known deep learning algorithms, the CNN-and DNN-based models, were applied to generate a landslide susceptibility map of the Ya'an-Linzhi section of the Sichuan-Tibet Railway combined with the application of InSAR Stacking technology identifies hidden danger points of landslides. A complete list of landslides improves the accuracy of landslide susceptibility evaluation. The results show that the two models have a higher success rate and prediction performance in this study area; however, the CNN algorithm showed a 4% higher performance than DNN. According to the analysis of the landslide-influencing factors within the study area, it was found that slope, elevation, and rainfall are the main influencing factors that affect the occurrence of landslides. High and very high landslide susceptibility were primarily distributed in the Jinsha, Lancang, and Nujiang River Basins along the railway, which better reflects the distribution of landslide susceptibility in the study area; therefore, providing a scientific basis for the disaster prevention and mitigation work of the Ya'an-Linzhi section of the Sichuan-Tibet Railway.