Landslide Risk Evaluation in Shenzhen Based on Stacking Ensemble Learning and InSAR

Construction activities of accelerated urbanization in Shenzhen have increased the landslide risk area, which has intensified the potential threat to human and natural environment. However, the risk of landslides in Shenzhen is poorly evaluated. In this article, a landslide risk evaluation (LRE) model is constructed using landslide susceptibility map (LSM) and landslide vulnerability. In the experiment, a stacking ensemble learning (SEL) model is constructed based on convolutional neural network (CNN), multilayer perceptron, gated recurrent unit (GRU), and support vector machine regression to generate LSM by using topography, geology, human engineering activities, time-series precipitation, and time-series normalized difference vegetation index. Road network, building distribution density and annual average precipitation data are used to evaluate landslide vulnerability based on entropy weight method. In this article, multiple statistical indicators are used to evaluate the performance of the LSM model, and Interferometric Synthetic Aperture Radar (InSAR) deformation data are utilized to verify the LRE results in Shenzhen. The results show that the SEL method has more refined results for LSM, with a best overall evaluation accuracy, especially in the receiver operating characteristic curve, where the accuracy is improved by nearly 8%. In LRE of Shenzhen, very high, high, moderate, low, and very low risk areas account for 0.283%, 0.451%, 0.859%, 36.890%, and 61.517%, respectively. In most of very high-risk area, InSAR deformation results show a clear concentrated deformation trend with a large deformation rate. Research results can provide technical and data support for landslide disaster prevention in Shenzhen.


I. INTRODUCTION
L ANDSLIDE is one of the major hazards prevalent around the world, causing serious economic losses, especially in highly dense urban areas and in landslide susceptible area [1]. Landslide risk evaluation (LRE) is the basis of landslide risk management, which conducts a comprehensive evaluation and analysis of the risks combining landslide hazards and potential value loss vulnerability [2]. Reasonable and effective landslide risk management can reduce the economic losses caused by landslide disasters [3]. In Shenzhen, due to a series of high-intensity engineering and human activities, coupled with the influence of unique climatic conditions, numerous potential landslides with high risk are formed, which brings great difficulties to the local disaster prevention and management. Therefore, a reasonable and effective LRE is extremely urgent for Shenzhen.
LRE is a systematic approach to identify and evaluate the potential losses and risks caused by landslides, and to provide a scientific basis for decision-makers. LRE is a complex and systematic exercise, the core of which consists of establishing scientific evaluation models. The purpose of LRE is to identify landslide potential hazards so that appropriate measures can be taken to reduce the risk. At present, there have been many studies on LRE, including fuzzy comprehensive evaluation [4], gray correlation analysis [5], and neural network [6]. Although these methods are widely used in practice, there are still some shortcomings. For example, the fuzzy comprehensive evaluation method is subjective and uncertain [7], leading to low accuracy of evaluation results, while the gray correlation analysis method cannot consider the interaction relationship between different factors and has low utilization of data.
LRE needs to consider potential hazards, such as threats to buildings, traffic routes, and human life [8], [9], [10]. LRE is determined by a combination of landslide vulnerability and landslide hazard. Landslide vulnerability is usually expressed in terms of economic and heuristic metrics [11], [12]. Landslide hazard analyses the geology, hydrology, topography, vegetation, and other factors of the potential landslide area, determines the risk level of the potential landslide, and classifies and evaluates the potential landslide area [13], [14]. There is a closed relationship between landslide susceptibility and landslide hazard. Therefore, landslide susceptibility is the basis of LRE, and only after an accurate evaluation of landslide susceptibility can we better carry out LRE.
In landslide susceptibility studies, machine learning methods can effectively capture the nonlinear relationship between landslides and geography [15], such as random forests [16], artificial neural network [17], support vector machines [18], [19], and deep learning [20], [21]. However, these methods classify the samples directly and cannot fully explore the underlying characteristics of landslides, resulting in insufficient generalization ability of single model, information redundancy, and overfitting when facing complex scenes and complex multisource data [22]. To address these issues, scholars use an integrated learning model strategy [23], [24], [25], [26] to compensate for the deficiencies existing in single models, optimize model performance, and greatly improve evaluation accuracy as well as generalizability. There are already integrated learning models for weak classifiers mostly using traditional machine learning models or homogeneous structural deep learning models. [27] used convolutional neural network (CNN), recurrent neural network (RNN) and logistic regression using stacking ensemble learning (SEL), blending, simple averaging, and weighted averaging integrated methods to achieve landslide susceptibility. Lv et al. [28] constructed integrated learning models based on single models such as CNN, residual network, and deep belief network as weak classifiers using stacking, bagging, and boosting methods achieve landslide susceptibility. All of them have achieved good results.
This article takes Shenzhen as the study area, ten static landslide conditioning factors (LCFs) and two time series dynamic LCFs are selected as the base data for LSE; A SEL model is constructed based on CNN, multilayer perceptron (MLP), gated recurrent unit (GRU), and support vector machine regression (SVR) as weak classifiers to generate high-precision landslide susceptibility map (LSM) for replacing landslide hazard. Based on road network, building distribution density and annual average precipitation data, landslide vulnerability weight values are established using the entropy weight method; Finally, a LRE model in Shenzhen is constructed using LSM and landslide vulnerability, the LRE map of Shenzhen is obtained and the regional hazard is verified and evaluated by InSAR deformation data.

A. Study Area
Shenzhen is located in the south of Guangdong Province, China, and is an important window city. Shenzhen city has Futian, Luohu, Nanshan, Yantian, Baoan, Longgang, Longhua, Pingshan, and Guangming nine administrative districts (see Fig. 1). The rapid urbanization of Shenzhen, the high density of construction activities, the high intensity of human activities, Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply. and a large number of artificial slope cutting and slope piling form unstable elements. In recent years, landslide hazards have occurred frequently, seriously threatening the regional engineering construction process and the safety of people's lives and properties, especially a strong precipitation is very likely to induce landslides, resulting in normalized difference vegetation index (NDVI) changes.

B. Landslide Conditioning Factors
1) Static Factors: Landslides are mainly influenced by topography, geology, hydrology, and humanities. In this article, selected LCFs include elevation, slope, aspect, terrain ruggedness index (TRI), distance from fault (DFFA), distance from river (DFRI), land use and land cover (LULC), distance from road (DFRO), soil types (ST) and soil content (SC), and ten static factors [16], [29] based on the environmental characteristics of Shenzhen. Used specific data of this article are as follows.
Digital elevation model (DEM) data with 30 m spatial resolution is used for elevation of Shenzhen city (http://earthdata. nasa.gov). Slope, aspect and TRI are calculated by DEM for topographic factors. Among them, slope is the key topographic factor that directly affects landslide stability [30]. Generally, the slope affects the infiltration process and the distribution of stress fields in slopes [29]. Aspect indirectly induces landslides by responding to the absorption of solar radiation [20]. TRI [31] is the ratio of the surface area of the ground unit and its projected area on the horizontal surface. It is used to quantify the variation of topographic relief and the degree of soil erosion in the study area. The magnitude value represents the ground fragmentation rate. The more fragmented the ground surface, the more serious the erosion.
In our study area, fault (http://www.resdc.cn) main characteristics are the general fault in Guangdong Province and the extrusion breakage consisting of two fault bundles with a medium-high angle of track hedging at 30-50°. At a certain distance, when DFFA is closer, the lithology is more discontinuous, the soil is looser, and the probability of landslides is higher [31]. DFRI reflects the influence of surface water on landslide evolution [32] in terms of water erosion and soil erosion. LULC and DFRO are used to measure the influence of human engineering activities on the land surface [33]. Different STs show differences in hydraulic parameters, such as inner SC, density, porosity, shear strength, and permeability coefficient [34]. Among them, an increase in pore water pressure will cause an increase in effective load stress, which will reduce the shear strength of the slope components and, thus, lead to landslides (http://www.resdc.cn). In this article, the selected ten static LCFs are shown in Fig. 2.
2) Dynamic Factors: a) Time-Series precipitation: Precipitation is one of the main LCFs for Shenzhen that causes slope instability [35]. Precipitation varies in different seasons, and short-term heavy precipitation increases the water content of the ground surface, which enhances the pore water pressure, reduces slope stability, and increases the possibility of landslides [36]. In this article, a total of 48 months of time-series precipitation data are selected from 2018 to 2021 as dynamic factors (http://www.resdc.cn). Fig. 3(a) indicates the temporal variation of the minimum and maximum precipitation in the selected study time. It can be seen that precipitation in the study area shows periodic changes. The spatial resolution of precipitation is resampled to 30 m by the inverse distance weighting method to ensure the same spatial resolution as the static factors and the time resolution is 30 days.
b) Time-Series NDVI: Vegetation is rich and covers a large area in Shenzhen. Vegetation affects soil and water conservation, and landslides lead to local NDVI reduction in a short period [37]. NDVI value will increase for abundant precipitation and mild climate [see Fig. 3(b)]. As shown in Fig. 3(b), three historical landslide zones are selected to monitor their NDVI values over time. Among them, NDVI shows an increasing trend with time and has a periodic change trend. In this article, the MODIS13Q product is selected for NDVI data with a spatial resolution of 250 m. The spatial resolution of NDVI is resampled to 30 m to ensure the same spatial resolution as the static factors and 32 days are selected to ensure that the two temporal factors have a similar time resolution.

III. METHODOLOGY
In this article, multicollinearity and Pearson correlation analysis are used to determine a multicollinearity problem and a strong correlation among LCFs. Geo-detector method is used to analyze the importance of LCFs on landslides in Shenzhen. Selected LCFs are divided into a training set (70%) and a testing set (30%). SEL method is constructed based on CNN, MLP, GRU, and SVR models to implement LSM for landslide hazard of Shenzhen, and accurate evaluation of LSM result and model performance is performed using various evaluation metrics. In addition, road network, building distribution density and average annual precipitation data are used as vulnerability. LRE model of Shenzhen is established by LSM and vulnerability. Finally, LRE model reliability is evaluated using InSAR deformation data and LRM of Shenzhen is obtained. The overall study flow chart is shown in Fig. 4.

A. Deep Learning Model
1) Convolutional Neural Network: CNN was originally proposed by [38] for handwritten number recognition. CNN mainly consists of convolutional layer, pooling layer and fully connected layer, main advantage of CNN is in processing data with spatial relationships and its ability to automatically learn image features [20]. In recent years, CNN has been widely used in landslide classification and LSM [22]. In this article, a CNN model is constructed based on convolutional layer, dropout layer and fully-connected layer for LSM (see Fig. 5). In convolutional layer, input LCFs are done convolutional operations by convolutional kernels to extract low-level features of each factor, and extract higher-level features by combination of multiple convolutional kernels. In dropout layer, it is mainly used to inhibit model overfitting. In fully-connected layer, the landslide feature map is unfolded into a one-dimensional (1-D) vector for final output. In this article, we stack LCFs in the channel dimension to form the original data set with dimension size  (Row, Column, 106), where row and column denote the row and column of LCF.
2) Multilayer Perceptron: MLP is a deep learning model based on feedforward neural networks [39]. MLP consists of multiple neural network layers, each layer contains multiple neurons, main advantage of MLP is achieving the modeling of nonlinear features by stacking multiple neurons with strong fitting ability. MLP is used to generate LSM [40]. In this article, MLP model with an input layer, three hidden layers and an output layer is constructed for LSM (see Fig. 6). The input layer can receive inputs from LCFs. The hidden layer can abstract features containing landslides from the output of the previous layer as input to the next layer through a combination of linear transformations and activation functions [41]. The output layer outputs probabilistic results of landslide susceptibility prediction. In this article, a same data structure as CNN (Row, Column, 106) is used for the input data set of MLP.
3) Gated Recurrent Unit: GRU is an excellent variant of the traditional RNN model for longer time sequences [42], [43]. The main advantage of GRU is it can effectively solve the gradient disappearance and gradient explosion problems of RNN in the training process of long sequence data [44], thus improving the prediction performance and generalization ability of the model. In this article, a three-layer GRU network structure is established to capture more stable landslide features to generate LSM. In GRU structure (see Fig. 7), the gating mechanism is introduced to control the flow and preservation of information through a gating unit with a reset gate and an update gate [45]. Reset gate can fuse the input temporal LCFs and the hidden state of the previous time step. Update gate receives the fused data, which can control the relative weight of the input data of the current time step and the hidden state of the previous time step, thus controlling the flow and preservation of information.
In GRU data processing, the input data set is divided into static and time series LCFs (see Fig. 8). Static LCFs are stacked in channel dimension based on the importance of each LCF, and the importance is calculated by Geo-detector [28], to form a feature dataset of size (Row, Column, 10), where 10 denotes the number of channels. The static LCFs are input into a fullyconnected layer for sequence processing. Time series LCFs are performed dimensional expansion in temporal dimension and channel dimension to form the original dynamic data set of size (48, Row, Column, 2), where 48 denotes the time series and 2 denotes the number of channels. Finally, the two parts of the data features are fused using concatenate structure.

1) Stacking Ensemble Learning Based on Single Deep Learn-
ing Model: Although CNN, MLP have shown good performance in LSM, the model performance is often inadequate when dealing with large time series dynamic data sets, which easily leads to model overfitting and information redundancy. The ensemble model idea is of great importance in model combination and model improvement. SEL is a model combinationbased approach that can take advantage of multiple underlying models to improve predictive performance and robustness [24]. It centers on using the predicted results of different classifiers as new features, and then using a high-level classifier to make final predictions based on these new features [23].
In this article, an SEL model is established to generate LSM. CNN, MLP, and GRU network models are used as single classifiers in the first layer and SVR model is used as a high-level classifier in the second layer. The main construction process is shown in Fig. 9. First, the landslide dataset is divided into a training set and a testing set. Second, the training set is divided into five parts, and four parts are used to train in base classification models, and the remaining one part is used as the cross-validation set. After five-fold cross validation, the predicted results of each base model on the cross-validation set are obtained as the new landslide features of the second layer classifier. In the second layer classifier learning process, new features learned from the first layer CNN meta-classifier prediction are stacked     sequentially to form 1-D column vectors. The new features obtained from MLP and GRU models are subjected to the same operation, and finally N × 3 vector matrices (A1, A2, A3) with rich landslide features are formed as the new training set for the second layer classifier. Testing results obtained from the first layer testing dataset are subjected to the same stacking operation to form a new N × 3 testing set (B1, B2, B3). New training set is fed into the SVR model for final fitting, and the new testing set is used to validate the metamodel predicted performance.
2) Model Accuracy Evaluation: Accuracy evaluation of predicted results is crucial for the predictive power of models [16]. In this article, overall accuracy (OA), specificity, Matthews correlation coefficient (MCC), F1 score, and Kappa statistical metrics are used to evaluate the performance of the model. The specific formula is as follows: where TP denotes the number of positive classes predicted as true positive, TN denotes the number of negative classes predicted as true negative, FP denotes the number of negative classes predicted as false positive, and FN denotes the number of positive classes predicted as false negative.
In addition, the deep learning model framework used in this experiment is based on Google's TensorFlow 2.8, implemented using python 3.10 programming, and the SEL framework is implemented based on scikit-learn. The model parameters are obtained through multiple experimental training.

C. Landslide Risk Evaluation
LRE is mainly determined by landslide hazard and vulnerability [9]. Landslide hazard refers to the probability of landslide occurrence in a certain area, which is related to the geology, topography, climate, and human factors of the landslide area [46]. Vulnerability indicates the degree of influence on human and natural environment, such as economic loss, damage to buildings, and roads after the landslide disaster [47]. Shenzhen city has been developing rapidly in recent years, with expanding population and urban construction. In this article, LSM is selected as landslide hazard indicator, road network, building distribution density, and average annual precipitation data are selected as vulnerability factors. LRE calculation formula is expressed in (8) where R denotes LRE, H L denotes landslide hazard, V L denotes vulnerability, D r denotes road density, D b denotes building density, D p denotes average annual precipitation, and w1, w2, and w3 denote the weight values assigned to three vulnerability factors using the entropy weight method.

A. LCFs Analysis of Shenzhen
Multicollinearity aims to determine whether there is a severe linear relationship among independent variables. Multicollinearity analysis can help to avoid problems of model instability and unreliability affected by relationship among independent variables [48]. In this article, we conducted a multicollinearity analysis on the selected 10 static LCFs (see Table I). The values of variance inflation factors (VIF) and tolerances (TOL) of all LCFs meet the independent variable threshold requirement (VIF < 10 and TOL > 0.1) (He et al., 2021). It indicates that there is no significant linear multicollinearity among LCFs. However, whether there is a nonlinear relationship among LCFs cannot be revealed by multicollinearity analysis.  Pearson correlation method is suitable for analyzing nonlinear relationships among factors [49]. Therefore, we use Pearson correlation method to further verify the nonlinear multicollinearity problem of selected 10 static LCFs (see Fig. 10). The correlation coefficients are relatively low in all. The maximum correlation coefficient is 0.49 between DFRI and DFRA, which indicates that selected ten static LCFs in this article have weak correlations with each other. Therefore, the selected LCFs in this article can be used for LSM.

B. LSM of Shenzhen
In the process of generating LSM based on the proposed SEL model, the training and testing sets are generated using a standard sampling strategy [27], in which 70% of the landslide samples (21281) are randomly selected for training, and the remaining 30% (9120) for testing. In addition, an equal number of nonlandslide areas are randomly selected for the training and testing datasets. In the training of the proposed SEL model, we employed a five-fold cross-validation method to search for the optimal hyperparameters of the model [27]. Moreover, we employed a slide-cut method to clip the selected LCFs pixel and pixel to obtain the dataset of the study area. Finally, LSM of Shenzhen city based on the proposed SEL model are obtained. Based on ArcGIS 10.7 software, we applied the natural break method [31], [34] to classify the LSM into very high, high, moderate, low, and very low susceptibility five leaves (see Table II, Fig. 11). Table II can be observed that the majorities of the landslide susceptibility areas obtain based the SEL method have low susceptibility, occupying a relatively large proportion of the entire study area. The proportions of very high, high, and moderate landslide susceptibility areas are 1.074%, 0.781%, and 1.003%, respectively, with extremely high susceptibility being the dominant category. As shown in Fig. 11, the areas with very high susceptibility are mainly concentrated in the central part of Shenzhen, with a few distributed along the coastal areas of Yantian district and the northern part of Longgang district. Historical landslides and potential landslides in development were mainly distributed in areas with very high, high, and moderate susceptibility. The very low susceptibility areas are mainly distributed in low altitude and low slope. The transition effect between different susceptibility levels is relatively smooth, and the center of the landslide body have the highest susceptibility intensity, which gradually weakens from the inner to the outer side. This is consistent with the varying degrees of impact of landslide bodies on the surrounding environment observed during field investigations. That was, as one moves away from the center of the landslide body, the strength of its impact decreases, leading to a lower probability of landslide occurrence and, thus, a reduction in landslide susceptibility [50]. It also indicated that the LSM generated by the proposed stacking model is reasonable.

C. LRM of Shenzhen
We used the LSM, combined with road network density, building distribution density, and annual average precipitation vulnerability indicators to generate LRM of Shenzhen. The entropy method [12] is used to adaptively calculate the weight for vulnerability indicators based on the size of the information entropy of each factor, the weight values are 0.342, 0.334, 0.324 for the LRM in each region, respectively. The LRM of Shenzhen is generated in Fig. 12. Based on natural break method [46], the LRM is divided into very high risk, high risk, moderate risk, low risk, and no risk areas (see Table II, Fig. 12).
It can be found from Table II that the proportions of areas with very low and low landslide risk are relatively high. Among them, the proportion of areas with low risk is 36.890%, which  may be due to the higher density of road networks and buildings in these areas, which elevate the risk level on the basis of very low susceptibility. The area of very high risk is 0.283%, which is relatively lower than that in the LSM. This could be due to the lower density of roads and buildings in some of these areas. The proportions of high and moderate risk areas are 0.451% and 0.859%, respectively. As can be seen from Fig. 12, areas of very high and high landslide risk mainly concentrate in the southwest of Longgang district and the central part of Baoan district, which means that these areas have higher risk of landslide.
Slope and altitude are the major factors affecting landslide [51]. By exploring the relationship between altitude and slope and landslide risk can we better understand the distribution characteristics of landslide risk in Shenzhen. In this article, we calculate the distribution of very high and high-risk areas of Shenzhen in slope and altitude, and analyze the relationship between them (see Fig. 13). As shown in Fig. 13(a) and (b), the average altitude of the very high and high-risk areas was mainly distributed between 91.86 m and 109.41 m, with a median altitude of 88.00 m to 104.00 m. Fig. 13(c) and (d) reveals that the average slope of the very high and high-risk areas is mainly distributed between 16.49°and 17.72°, with a median slope of 14.10°to 16.73°. In summary, our analysis reveals a clear correlation among altitude, slope and risk levels of landslide hazard areas in Shenzhen. By providing these insights, our research can help inform landslide risk management and planning strategies, such as targeted implementation of monitoring and landslide hazard altitude within potential landslide risk zones characterized by relevant altitude and slope, to mitigate future landslide risk.

A. Evaluation of SEL Model Performance
Landslide susceptibility analysis is crucial for LRE [52]. It is of great significance to obtain more accurate LSM [27]. CNN, MLP, and GRU models have good performance in LSM, however, single model has exposed some shortcomings due to the large amount of remote sensing data and complex scenes, such as overfitting and information redundancy. Therefore, in this article, we used SEL technique for LSM, at the same time, it is compared with the single model (CNN, MLP, GRU, and SVR).
LSMs obtained using a single classifier model (CNN, MLP, GRU, and SVR) are presented in Fig. 14. Fig. 14(a), (b), (c), and (d) represents LSM results obtained based on CNN, MLP, GRU, and SVR models, respectively. It can be seen all single neural network classifiers can obtain relatively accurate LSM results. Intuitively, the very high susceptibility areas overlap  with the historical landslide and landslide susceptible points. The LSM results of CNN and GRU models are similar with that of the SEL model. However, CNN model LSM result show poor transition between different susceptibility zones. This may be due to the powerful feature extraction ability of the convolutional layers in the CNN model [48]. In contrast, GRU model considers the temporal dynamic features through recurrent layers, which enhances the efficiency of GRU. LSM result of GRU model show better transition between different susceptibility zones compared to CNN model. However, the areas with very high, high, and moderate susceptibility are distributed over a larger area in GRU model than in CNN and SEL models. The very high, high susceptibility area predicted by MLP model is widely distributed, spanning the central region of Shenzhen. The LSM result of SVR model is less reliable than those of deep learning models overall, which does not match reality. In addition, in some nonlandslide areas with small surface undulations and gentle slopes, the LSM results of MLP [see Fig. 14 [see Fig. 14(c)] models show high and moderate susceptibility. This phenomenon may be related to the model's difficulty in overcoming overfitting, which results in a weak generalization ability to identify effective features in these regions with small variations in terrain and gentle slopes. From the model susceptibility visualization effect, the proposed SEL model shows an overall better performance. Compared to the CNN, GRU, MLP, and SVR models, the SEL method integrates the advantages of each single classification model in processing data and has a better ability to extract features. SEL model overcomes the problem of poor generalization ability and overfitting of single models, and the overall performance of the proposed model is best.
To further evaluate the reliability of the proposed SEL model used in this article, we employed multiple quantitative indicators including Recall, Specificity, MCC, OA, Kappa, and F1 scores to assess the accuracy of the SEL model (see Fig. 15). As can be clearly seen from Using ROC curve, we further explore the accuracy and reliability of the SEL model [53]. As shown in Fig. 16, the SEL model has the highest area under the ROC curve, which is 0.992, followed by the CNN model (0.988), and the SVR model has the smallest area under the ROC curve. Compared with the SVR model, the accuracy of the SEL model has been improved by about 9.4%. Combined with the Recall, Specificity, MCC, OA, Kappa, F1 scores, and ROC accuracy indicators, it can be concluded that the SEL model has higher accuracy and reliability compared to single-classification models.

B. Landslide Risk Analysis
LSE is determined by a combination of landslide vulnerability and landslide hazard. The landslide vulnerability is usually expressed in terms of economic and heuristic metrics [54]. Landslide hazard determines the risk level of the potential landslide [13], while LSM indicates the potential location of landslides. Therefore, landslide susceptibility can be able to the landslide hazard indicator. Most landslides in Shenzhen are typical precipitation-induced landslides, which have suffered a lot of damage in the past few decades. Heavy precipitation is one of the causes of landslide occurrence and formation of unstable slopes [55]. Therefore, precipitation can be used as an indicator to evaluate landslide vulnerability.
Due to the rapid development of urbanization in Shenzhen, human activities have expanded to sloping terrain, disrupting the geological environment and increasing the risk of urban landslides. When a rapid flow landslide occurs, buildings along its path will affect the flowability of the landslide [54]. Luo et al. [56] evaluated the impact of building blockages on landslide flowability and related energy dissipation mechanisms by analyzing the large-scale landslide disaster in Shenzhen in 2015.
The results showed that building clusters need to be considered in urban landslide disaster mapping and risk assessment. In addition, landslides can cause severe damage to the road network in the affected areas, resulting in both direct and indirect losses [57]. It is crucial to identify the parts of the road network that are more prone to landslides to reduce the risk and economic costs of potential population impacts and losses [47]. Therefore, the distribution and density of the road network are also important indicators for measuring vulnerability. In summary, LSM obtained in this article is credible, and the vulnerability selected index is reasonable, therefore, the LRM result in Shenzhen is reliable.
InSAR technology has become an effective tool for conducting regional surface deformation monitoring and early identification of landslides, which provides a new perspective for accurate and near real-time landslide hazard research. Numerous studies have demonstrated that monitoring and identification of unstable landslides using InSAR technology is a reliable means of landslide prediction [58]. Therefore, InSAR results are used to verify the high-risk areas to further verify the reliability of LRM result in this article.
Based on LRM of Shenzhen, the very high-risk areas are extracted based on ArcGIS software, and are overlaid Google Earth with InSAR deformation data for visualization and analysis. As shown in Fig. 17, the selected 12 very high-risk areas show a maximum negative deformation rate of −47 mm/year and a maximum positive deformation rate of 17 mm/year, and all the very high-risk areas show a concentrated deformation funnel, mainly showing a subsidence trend. Buildings are sensitive to movements caused by ground deformation [59]. It can be clearly seen from Fig. 17 that the very high-risk areas such as R1, R3, R5, and R7 involve larger construction areas, including some buildings and human living areas, while R2, R3, R4, R5, R9, and R12 areas involve multiple road networks, which indicates that in these riskier areas, the buildings are more vulnerable in case of disasters, and the damage to them is relatively large, and the buildings around these areas are easily destroyed by the impact from outside.
Time series distribution of the InSAR deformation maps of the selected four very high-risk areas (R4, R6, R7, and R10) from the selected 12 typical very high-risk areas and then one subsidence point (A1, A2, A3, A4) was selected in each area, and the mean value of time series deformation within 10 m buffer around this subsidence point was calculated, finally, the time series curves were plotted with the error maps of the InSAR deformation data are shown in Fig. 18. Fig. 18(a), (c), (e), and (g) indicates the changes of cumulative deformation variables within the four risk areas, respectively. It can be clearly seen that the cumulative deformation trend line shows a negative increasing trend as time changes within the risk area. Fig. 18(b), (d), (f), and (h) shows their error distributions, respectively. InSAR deformation data error distributions are analyzed by the Bootstrap method [60] for standard deviation. The bootstrap method is a statistical technique for estimating the uncertainty of a parameter or statistic in a sample by generating many samples from the original data and calculating the parameter or statistic for each sample. In the region of Fig. 18(b), after 1000 random samples, the distribution of uncertainty estimates obtained for multiple deformation is roughly in the range of −15 mm, with an error within 2 mm. In the regions of Fig. 18(d), (f), and (h), after the same resampling strategy, the uncertainty estimates of deformation are distributed around −21 mm, −7 mm, and −11 mm, respectively, with the error distribution within 1 mm, which is a robust estimate of the trend line and error range obtained from the original sample data. It indicates that the InSAR deformation bias obtained in this article is small and the results are reliable.
Combined with Figs. 17 and 18, distribution of InSAR deformation data and the trend of time-series InSAR changes in the very high-risk areas of Shenzhen show that irregular deformation exists in these areas, and the deformation rate is large, which poses a greater threat to the safety of buildings, roads, and people's lives and properties in the surrounding areas.

VI. CONCLUSION
The rapid urbanization and construction activities in Shenzhen have greatly increased the landslide risk area, posing a serious threat to human and natural environments. This article develops an LRE model that integrates LSM and landslide vulnerability of Shenzhen. We constructed an SEL model using CNN, MLP, and GRU network models as the first layer weak classifier and SVR model as the second layer classifier to generate LSM using aspect, slope, TRI, DEM, DFFA, DFRO, DFRI, sand, LULC, ST, time series precipitation and time series NDVI. Recall, specificity, MCC, OA, Kappa, F1 score, and ROC statistical metrics are used to evaluate the performance of the LSM model. The LSE results show that the SEL method provides finer results for LSM with the best overall evaluation accuracy, especially on ROC, with an accuracy improvement of nearly 8%. Landslide vulnerability evaluation is carried out based on the entropy weight method, taking into account the road density, building distribution density, and average annual precipitation data in Shenzhen. The LRE results show that the very high, high, medium, low, and very low risk areas account for 0.283%, 0.451%, 0.859%, 36.890%, and 61.517%, respectively. Among them, the area of very high and high-risk zones reached 15.500 km 2 . Notably, significant concentrations of InSAR deformation regions with large deformation rates are observed in most of the identified very high-risk areas, the absolute value of deformation rate up to 47 mm/year in the selected 12 typical very high-risk area. Our research results can provide valuable technical and data support for effective landslide hazard prevention and mitigation in Shenzhen. The LRE model based on LSM and landslide vulnerability evaluation can be considered as a reliable method to evaluate landslide risk in areas with rapid urbanization and construction activities.