Landslide Susceptibility Mapping Based on Resampling Method and FR-CNN: A Case Study of Changdu

Abstract

Deep learning can extract complex and high-dimensional characteristic information with its deep structure, effectively exploring the complex relationship between landslides and their numerous influencing factors, and ultimately, more accurately predict future landslide disasters. This study builds a landslide susceptibility mapping (LSM) method based on deep learning, compares the frequency ratio (FR) sampling method with a buffer random sampling method, and performs resampling operations of landslide and non-landslide samples to explore the applicability of deep learning in LSM. In addition, six indices, precision, accuracy, recall, ROC, and the harmonic mean F1 of accuracy and recall were selected for quantitative comparison. The results show that both the resampling method proposed in this paper and the non-landslide sample selection method based on FR can significantly improve the accuracy of the model, with the area under curve (AUC) increasing by 1.34–8.82% and 3.98–7.20%, respectively, and the AUC value can be improved by 5.32–9.66% by combining the FR selection and resampling methods. Furthermore, all the deep learning models constructed in this study can obtain accurate and reliable landslide susceptibility analysis results compared to traditional models.

1. Introduction

The Qinghai–Tibet Plateau region is a high-risk area for landslide hazards in Asia. The complex geology of the Qinghai–Tibet Plateau region, coupled with global warming and the frequency of extreme weather events in recent years, has led to an increasing frequency of landslide hazards [1,2]. Landslides cause serious human, social, environmental, and economic losses [3,4]. Therefore, reliable landslide prediction is essential. Landslide susceptibility mapping (LSM) assumes that the occurrence of landslides is closely related to their natural surroundings and engineering activities [5], and furthermore, that areas where landslides have occurred in the past are more likely to experience landslides in the future [6]. The probability of landslides can be predicted by analyzing the spatial distribution of landslides that have already occurred and combining the underlying geological conditions, topography, and hydrology of the study area [7,8].

Numerous approaches have been proposed to assess landslide susceptibility, among which non-deterministic models are the most widely used. The non-deterministic models can be divided into knowledge-driven and data-driven models according to the way in which the weights of the factors influencing landslides are taken [9]. Analytical hierarchy process [10,11,12], expert scoring [13], and fuzzy logic [14,15] are typical representatives of knowledge-driven models. These LSM models require the researcher to have a very thorough understanding of the causes of landslides and to assign different weights to the influencing factors, which are highly subjective [16]. Statistical analysis methods and machine learning are both data driven, and include informative methods [17], frequency ratio model (FR) [18,19], logistic regression [20,21], and weight of evidence methods [22]—which are are widely used in landslide susceptibility assessment [23]—and machine learning such as support vector machines(SVM) [24], random forests (RF) [25], and artificial neural networks [26], which also play an important role in LSM. Traditional statistical methods and general machine learning have a relatively shallow structure that makes it difficult to cope with the task of analyzing complex linear or non-linear relationships between landslide hazard data and environmental factors [27]; therefore, it is increasingly important to analyze these massive amounts of landslide hazard data accurately and effectively.

As computer technology has been successfully applied in various fields [28,29], convolutional neural networks (CNN), a representative of deep learning, have gradually started to be used in landslide susceptibility analyses [30,31]. The deep structure of CNN can characterize the sensitive information of landslide factors, thus better extracting the deep features of landslide influencing factors. On the other hand, the multi-scale feature extraction capability of CNN can make fuller use of the environmental information around the landslide, and further improve the accuracy of landslide susceptibility analysis by combining the multi-scale features of the landslide. However, little research has focused on the effectiveness of CNNs in LSM applications [32]—the key to which is the issue of training dataset selection and data representation. To overcome the problem of data representation in CNN in LSM, Wang et al. [33] used one-dimensional, two-dimensional, and three-dimensional data representations to build a CNN framework and found that CNN–2D is more robust than other dimensional data. Recently, Yi et al. [34] proposed a multi-scale sampling strategy for this problem and analyzed the effect of different landslide sample sizes on the accuracy of CNN models.

Non-landslide samples in the training dataset can suppress overestimation of predicted landslide susceptibility [35], but the above studies do not take into account that both the spatial extent of non-landslide sample selection [36] and the manner of selection [37,38] can have an impact on the results of landslide susceptibility modeling.

In this study, we used the Changdu region as an example to construct a landslide susceptibility model based on CNN, aiming to investigate the rationality of non-slip sample selection and the applicability of deep-learning methods in LSM. Non-landslide samples were selected using the FR model, and based on this, resampling operations were performed on landslide and non-landslide samples to generate model training data. In addition, samples selected using the random sampling method are used to build models by applying FR, multilayer perceptron (MLP) neural networks, and SVM to verify the feasibility and effectiveness of the resampling operation based on FR selection in deep learning.

2. Study Area and Data

2.1. Study Area

As the research area selected for this study, Changdu (as shown in Figure 1) is located in the eastern part of Tibet, with a total area of 109,816 square kilometers, 93°6′–99°2′ east longitude, 28°5′–32°6′ north latitude, looking east to Sichuan, southeast bordering Yunnan and Qinghai to the north, and located in the Hengduan Mountains and the basins of the Jinsha River, Lancang River, and Nujiang River, at the junction of the Eurasian plate and the Indian Ocean plate. Neotectonic movement has caused significant landform differences in Changdu, with complex terrain and large elevation differences, and generally with a relative elevation difference of more than 2000 m [39]. The rock and soil in this area are often exposed, with broken and loose deposits formed under the action of freeze-thaw weathering [40], coupled with the strong geological movement of external forces, resulting in frequent and large-scale geological disasters in the area, which are extremely serious [41].

2.2. Landslide List

The landslide list is the first step in—and basis of—the LSM. In this study, the top centroid of the initial surface of the landslide to represent the landslide, and the historical landslide point data catalog of the study area, was established based on survey data of historical geological hazards in Changdu and combined with remote sensing images. A total of 425 historical landslides occurred in the study area. Figure 1 shows the distribution of landslides in the study region, the largest volume in these landslides is 3600 × 104 m³ [42]. Models using only landslide samples usually extend the range of landslide hazard zone predictions due to the lack of constraints of non-landslide samples [43]. Moreover, the way in which non-landslide samples are selected can also affect LSM results [44]. For example, the random sample selection method may result in the selection of non-landslide sites that are not of low landslide susceptibility, reducing the quality of the training dataset, and thus, affecting the accuracy of the LSM results. Therefore, the non-landslide samples were selected based on the LSM classification results of the FR model in this study. The sample was randomly divided into 2 parts: 70% of the dataset was used for training the model and the remaining 30% for validation.

2.3. Landslide Evaluation Factor

The base data for this study consisted mainly of a 1:50,000 topographic map (published by the National Geographic Information Centre in 2021), a 1:200,000 geological map (completed by the Geological Survey of China in 2002), a 1:1,000,000 soil map of the People’s Republic of China (prepared and published by the National Soil Census Office in 1995) and a 30 m SRTM DEM with GlobalLand-30 (the source is shown in

Table 1. Experimental data and sources.

1st-Order Factor	2nd-Order Factor	Data Source
Geological conditions and topography	Lithology	http://geocloud.cgs.gov.cn (6 April 2022)
	Distance to faults
	Elevation	http://dwtkns.com/srtm30m/ (5 April 2022)
	Slope	30 m SRTM DEM
	Aspect
	Profile curvature
	Plane curvature
	Degree of relief
	Surface roughness
Hydrological conditions	Distance to streams	https://www.webmap.cn (1 April 2022)
	the topographic wetness index (TWI)	30 m SRTM DEM
Human activities	Distance to roads	https://www.webmap.cn (1 April 2022)
	Land-use	http://www.globallandcover.com/ (2 April 2022)
Surface cover	the Normalized Difference Vegetation Index (NDVI)	GEE (Landsat 8 OLI)
	Soil	1:1 million Soil map of the People’s Republic of China

). In this study, the topography, geology, and hydrological conditions of the study area are taken into consideration, along with other factors. In total, 15 landslide evaluation factors were finally selected, as shown in Figure 2, where the lithological categories are described as shown in

Table 2. Description of lithological categories.

Lithology Classification	Description
A	Other lithology
B	Hard rock intrusive formations
C	Soft, thinly bedded sand, gravel and mudstone formations
D	Soft and hard detailed laminated, thinly bedded sand, sand and gravel, mudstone formations
E	Hard, harder lava rock formations
F	Clay and sandy soils
G	Harder laminated, thinly bedded medium to shallow metamorphic formations
H	Thinly laminated schistose sand shale formations with soft and hard intervals
I	Hard, harder laminated tuffs, dacite, dolomite formations
J K	Soft and hard interbedded carbonates, clastic rock formations
	Permafrost
L	Hard, more rigidly laminated, thinly bedded carbonate formations

. This study performs image alignment, vectorization, and geo-correction on existing base data using ArcGIS and the Google Earth Engine (GEE) platform, as different evaluation units are suitable for different scales and types of landslide susceptibility studies [45]. Considering the wide scope of the study area and the large amount of data, a regular grid was chosen as the evaluation unit for this paper. To ensure spatial consistency, the factor data were resampled to a raster size of 30 m×30 m resolution.

3. Methodology and Model

In this study, the landslide susceptibility of the study area was assessed based on the CNN model, which mainly included the following 6 processes: (1) select the landslide evaluation factor after feature selection as the basic evaluation factors and stack all factors into a factor data set; (2) use the FR model for landslide susceptibility in the study area, which is divided into 5 categories and randomly select non-landslide grid cells from the low and very low susceptibility areas; (3) resample operation for landslide and non-landslide raster cells of the training dataset; (4) train the network using CNN to predict the final landslide susceptibility partition; and (6) analyze landslide influencing factors and model performance. In addition, FR, SVM, and MLP neural networks were used for comparisons. Figure 3 shows a flow chart of this study.

3.1. FR Model

The FR model quantitatively analyzes the relationship between landslide points and various factors that affect landslides [46]. Landslides are affected by different factors, each of which plays a different role in landslide development, and the nature of each factor will also be different. By calculating the ratio of the probability of landslide occurrence to no landslide occurrence under different classification intervals of each evaluation factor (i.e., the frequency ratio FRI) [47], the frequency ratios of all evaluation factors were added to obtain the landslide susceptibility index (LSI) [48]. The frequency ratio calculation formula is as follows:

$$FRI=\frac{P_{ij}/P_r}{A_{ij}/A_r}$$

(1)

where FRI is the frequency ratio, $P_{ij}$ is the number of landslide points in the j-th category of the i-th landslide evaluation factor, $P_r$ is the sum of the landslide points in the study area, $A_{ij}$ is the area of the j-th category of the i-th landslide evaluation factor, and $A_r$ is the area of the study area.

The frequency ratio indicates the sensitivity of the evaluation factor to the landslides. An FRI value of 1 (FRI = 1) indicates that the evaluation factor has no apparent impact on landslides. An FRI value greater than 1 (FRI > 1) indicates that the evaluation factor promotes landslides. When the FRI value is less than 1 (FRI < 1), the factor degree of correlation with landslides is low [49].

The importance of each evaluation factor in the landslide susceptibility evaluation was different. The weight of each evaluation factor can be obtained by performing a relief-f importance analysis on each factor relative to the landslide—that is, the importance of the evaluation factor relative to the landslide. In this study, combined with the relief-f method, the frequency ratio of different evaluation factors is multiplied by the corresponding weight value and then added to obtain the weighted frequency ratio:

$$LSIw=\sum {W_i\times FRI}_i$$

(2)

where $LSIw$ is the index of the weighted landslide susceptibility, $W_i$ is the weight of each factor, and ${FRI}_i$ is the frequency ratio of each factor.

3.2. Landslide Resampling Method R Model

CNN can perform convolution and pooling operations on input data of different dimensions; therefore, landslide susceptibility modeling using CNN needs to consider the expression of landslide data [16,33]. Usually, we obtain historical landslide data in terms of the top centroid of the initial surface of the landslide to represent the landslide, thus lacking information on the environment surrounding the landslide [35]. However, a landslide is multi-dimensional data, including multi-dimensional data of its geographical location and environment. A one-dimensional representation for convolutional operations is not conducive to the training of CNN models. Although the planar two-dimensional data, to a certain extent, gives the relative spatial position of evaluation factors and considers the association between a single environmental factor and landslides, landslides are the result of multiple factors acting together; therefore, the association between multiple influencing factors in the landslide area needs to be considered in LSM [16].

The assessment of landslide susceptibility is based on the first law of geography. The closer the ground objects are, the more significant the correlation between them; therefore, the occurrence of landslides is closely related to the surrounding environment [50]. In this study, considering the availability of point landslide data and the irrationality of expression, a resampling operation of landslide point data was performed based on geospatial autocorrelation [34,51]. First, the screened nine landslide evaluation factors were stacked into a factor dataset in the form of three-channel data form. Second, this study used the low and very low landslide susceptibility areas obtained from the FR, and based on this, randomly selected the same number of non-landslide samples as the landslide samples. The sample points and their surroundings are then extracted from the superimposed factor dataset to produce a sub-dataset at a scale of 3 × 3 pixels to replace the original 1-dimensional sample dataset. Finally, the sub-dataset was used to build the LSM method for the CNN, CNN–squeeze-and-excitation networks (CNN-SE) and residual networks (ResNet) (referred to as FR–CNN–resample, FR–CNN–SE–resample, and FR–ResNet–resample in later text). In addition, the buffer random sampling method (randomly selecting the area beyond a certain distance from the landslide point) and landslide resampling were used to build a sample dataset with equal sample sizes to construct the CNN, CNN–SE, and ResNet models (hereafter referred to as CNN–resample, CNN–SE–resample, and ResNet–resample later in this paper) for comparison.

3.3. CNN-SE Model

3.3.1. CNN Theory

A convolution neural network (CNN) is a classic and widely used deep learning algorithm. CNNs effectively reduce the learning complexity of the network through the characteristics of local connections, weight sharing, and pooling operations of the network, making the model more robust, fault-tolerant, and easier to train and optimize [52]. CNN can learn image features through the training, extraction, and classification of image features [16].

The structure of the CNN is divided into three parts: a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer realizes the feature extraction function; the pooling layer is used to reduce the dimension of the output of the convolutional layer and reduce the complexity of the model; and the fully connected layer is used for classification. Among them, the convolutional layer is a unique layer of the CNN that is different from other networks, and it is also the core layer of the CNN [53]. The convolutional layer sequentially performs convolution operations on the input images at different positions using the convolution kernel to extract the image features. The formula is given in Formula 3.

$$C_j=\sum _i^{N}f\left(w_j\times x_i+b_j\right), j=\mathrm{1,2},\dots ,k$$

(3)

where k is the number of convolution kernels, $C_j$ is the output of the j-th convolution kernel, $f$ is the non-linear activation function, i is the location of the convolution operation, $x_i$ is the input data corresponding to the convolution kernel, $w_j$ is the weights of $j$, and $b_j$ is the bias.

3.3.2. SE-Net

squeeze-and-excitation networks (SE) were proposed by Jie et al. [54]. The SE module improves the quality of representations generated by the network by learning the interrelationships between the channels of convolutional features and improves the sensitivity of the model to channel features [55]. Experiments show that the SE module can add an attention mechanism to the network and significantly improve the network performance of CNNs at the cost of micro-computation [56].

The SE module is mainly composed of squeeze and excitation. The compression part compresses the original input features from two dimensions to one. The one-dimensional parameter obtains the previous two-dimensional global field-of-view, and the perception area is more expansive. This part is realized by global average pooling, where the excitation part consists of two fully connected layers generated for each feature channel through parameter w weights. Finally, the weight value of each channel calculated by the SE module was multiplied by the two-dimensional matrix of the corresponding channel of the original feature map to obtain the resulting output. The formula is as follows:

$$z_c=\frac{1}{H\times W}\sum _{i=1}^H\sum _{j=1}^W{u}_c(i,j)$$

(4)

where $z_c$ is the c-th statistic, $u_c$ is the feature map of the c-th channel of, size is $H\times W$, and (i, j) is the value of this position on the feature map.

3.4. Comparative Methods and Model Evaluation Indicators

In this study, the precision, accuracy, recall rate, harmonic mean F1 of precision and recall rate, Matthews correlation coefficient (MCC), and receiver operating characteristic curve (ROC) were selected. Six indicators were used to evaluate the performance of a deep learning model based on FR with the resampling method, combined with the proportion of historical landslide points falling in areas with high landslide susceptibility. The area under the ROC curve (AUC), represents the prediction accuracy. The higher the AUC value, the higher the model’s prediction accuracy; the closer the curve is to the upper left corner, the higher the model’s prediction accuracy. The calculation formula for each evaluation index is as follows:

$$Accuracy=\frac{TP}{TP+FP}$$

(5)

$$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$

(6)

$$Recall=\frac{TP}{TP+FN}$$

(7)

$$F1=\frac{2\times Precision\times Recall}{Precision+Recall}$$

(8)

$$MCC=\frac{TP\times TN-FP\times FN}{\sqrt{(TP+FP)(TP+FN)(TN+FP)(TN+FN)}}$$

(9)

where TP is the number of correctly classified landslide samples, FP is the number of misclassified non-landslide samples, TN is the number of correctly classified non-landslide samples, and FN is the number of mis-classified landslide samples.

In addition, this study selected the FR model, MLP neural network, and SVM model for comparison with the deep-learning model constructed in this study to verify the model’s accuracy further, and conducted deep learning landslide susceptibility analysis using samples selected by the buffer random sampling method (without resampling operation) to verify the effectiveness of the landslide resampling strategy of landslide samples using the FR selection method.

4. Results

4.1. Filtering of Landslide Evaluation Factors

Usually, there may be different degrees of correlation among the initially selected landslide susceptibility evaluation factors, resulting in redundancies or those that are unrelated to landslides [57]. Excessive input variables increase the construction cost of the landslide susceptibility model and reduce the model’s performance to a certain extent [58]. The characteristic factors were screened to reduce the modeling cost and improve the model performance before the landslide susceptibility model. Therefore, Pearson correlation analysis and the relief-f method were used to screen landslide evaluation factors in this study.

The Pearson product–moment correlation coefficient is widely used in landslide susceptibility evaluation to measure the strength of the correlation between evaluation factors [59,60]. The results of the correlation analysis of this study are shown in Figure 4. Among the 15 evaluation factors initially selected, the correlation coefficient between elevation and soil type was 0.630, indicating a relatively strong correlation. The correlation coefficients of degree of relief were 0.833 and 0.839, respectively, which were strongly correlated. The correlation coefficients between elevation and distance to roads and between profile curvature and plan curvature are 0.522 and −0.515, respectively, with absolute values greater than 0.5, showing a high degree of correlation. Therefore, elevation, profile curvature, degree of relief, and slope were excluded.

In addition, the relief-f algorithm was used to measure the importance of the remaining 10 factors. The relief-f algorithm is a feature weighting algorithm that assigns weights to evaluation factors based on the correlation between the evaluation factors and landslides [61]. The importance of landslide evaluation factors is shown in Figure 5, where the NDVI should be eliminated because its importance is negative. Finally, 10 evaluation factors, including distance from the road, distance from the water system, and land-use type were screened to build a landslide evaluation model.

4.2. Influencing Factors Analyses Using FR Model

The relationship between landslide occurrence and related influencing factors using the FR model is shown in Figure 6. Essential landslide evaluation factors have varying degrees of influence on landslide development. In this study, the frequency ratio and relative point density of landslides were used to quantify the landslide development in the study area. The frequency ratio reflects the sensitivity of different grading intervals to different evaluation factors for landslides [62]. In contrast, the relative point density of landslides can reflect the spatial distribution characteristics of the landslide points and each evaluation factor. It can be analyzed whether the evaluation factors in this classification state are conducive to landslide development.

(1)

Stratum lithology

The frequency ratios of mudstone rock groups, sandy soils, carbonate rocks, and clastic rock groups are greater than 1, and the relative point density of landslides is relatively large (Figure 6a), indicating that rock groups with soft, soft, and hard rock groups are more conducive to landslides development. This is because the lithology of the stratum determines its mechanical properties. The more complicated the lithology category, the stronger the mechanical properties of the stratum are. On the contrary, the softer the lithology, the lower the mechanical strength, the worse the geological engineering conditions, and therefore, the more prone it is to landslides.

(2)

Distance from fault

Faults are a type of structure that developed in the crust during tectonic movements. Faults are generally considered essential factors in landslide development. At a distance of 0–1959.38 m from the fault zone, the frequency ratio is greater than 1, and both the FR and the landslide relative point density (LRPD) values show an opposite trend to the distance value from the fault (Figure 6b). This indicates that landslides are more likely to occur within a distance of 0–1959.38 m, and that the closer the distance to the fault, the higher the landslide susceptibility. Because the rocks on the fault are often broken and easily weathered, the rocks near the fault are generally more broken, and the rocks farther away from the fault gradually transition to complete rocks, making it easier to accumulate and cause geological disasters such as landslides near the fault zone.

(3)

Topography

The altitude of the study area has a direct influence on landslide development. The analysis showed that landslides are more likely to occur in the study area in the elevation range of 2023–4141 m (Figure 6c). This range only accounted for 25% of the study area, but more than 85% of the landslide disaster points were distributed. Among them, the density below an elevation of 3592 m was the highest, reaching 0.024/km². Landslides rarely developed above an elevation of 4141 m. From the frequency ratio and relative point density of landslides, it can be seen that they rarely occur in the area above an elevation of 4500 m; however, this may be related to other influencing factors.

Landslides are more likely to occur on the northeast, south, and southwest slopes, or areas with slopes below 60°, and the distribution density of landslide points in the range of 45°–60° is the highest (Figure 6d,e). This indicates that steep slopes are key conditions for landslides.

The landslide development in this study area was most favorable in the area where the topographic relief was less than 37 or the surface roughness was 1.48–2.00 (Figure 6h,i). Curvatures also affect the landslide development. Plane curvature affects the convergence and diffusion of surface water systems, and it directly controls the speed of water flow and thus erosion [63]. The curvature of the section affects the accumulation of surface rock mass or deposits; therefore, the convex section (profile curvature > 0) is more conducive to the accumulation of deposits required for the development of landslides, while the concave plane (plane curvature < 0) is more conducive to the accumulation of water systems. The same result is obtained from Figure 6f,g: study areas with section curvature greater than 0.5, or plane curvature less than 0.5 have higher frequency ratios and are more prone to landslides.

(4)

Hydrological conditions

Hydrology is an essential factor affecting the strata of landslides. Under gravity, the water system always gathers and flows downhill—that is, the trend of water accumulation and the TWI can reflect the magnitude of this trend. Therefore, in this study, the TWI and the distance to the water system were chosen to assess the hydrological conditions of the study area. Landslides are prone to occur when the topographic humidity index is more significant than the 11.76 range, or when the area is closer to the water system with a greater value of frequency ratio to landslide point density (Figure 6m,l).

(5)

Human activity

Human activities, especially engineering are important factors affecting landslides, which cannot be ignored in the factors leading to landslides. Road construction usually causes the lower part of the slope to lose its support, and landslides occur. The land-use type reflects the degree of human use or damage to the land, so the type of land use and the distance from the road were selected as the evaluation factors for landslides in this study. According to the analysis in Figure 6, landslides mainly occur in areas closer to the road (within 1961.23 m), where human activities are more frequent. The frequency ratio of landslides in this area was the largest, reaching 2.537, and the density of landslides was also the largest (Figure 6j). Landslides were more likely to occur in areas with frequent human activities such as cultivated land, grassland, wetland, and artificial surfaces (Figure 6k).

(6)

Land cover

Soil and vegetation coverage are factors that influence landslide development. Different types of soil have different degrees of stability. The higher the soil stability, the more stable the slope and the less prone it is to landslides. Vegetation functions in water and soil conservation, and areas with high vegetation coverage can retain water and sand. This effect is noticeable and reduces the slope’s deformation tendency and the probability of landslide occurrence [64]. In this study, the normalized difference vegetation index (NDVI) was used to measure the vegetation growth status. The NDVI reflects the background influence of the plant canopy. Landslides are more likely to occur in the areas of dark brown soil, fuller gray soil, yellow-brown soil, and black felt soil, among which the density of landslides in the fuller gray soil area was the highest (Figure 6o), In areas where the NDVI value was less than 0.36, the landslide frequency ratio was greater than 1 (Figure 6n). This means that areas with low vegetation coverage or ice and snow coverage (NDVI less than 0) were more conducive to landslide development.

4.3. Mapping and Verification of Landslide Susceptibility by Deep Learning

In this study, the CNN, CNN–SE and ResNet models were built based on PyTorch, a deep learning library for Python. CNNs are used for susceptibility analysis in a pixel-by-pixel binary classification process. The convolution kernel size used in this study was 3× 3. The last fully connected layer uses a dual-node output, the Softmax function as the activation function, and Cross Entropy. The initial learning rate was set to 0.01, and the CNN–SE model converged after approximately 300 epochs (Figure 7).

The landslide susceptibility index for each grid unit was obtained using the established model. The natural breakpoint method was used in ArcGIS to reclassify the susceptibility index into five levels [7]. The landslide susceptibility grading map and grading proportion map obtained by the different models (FR–CNN–resample, FR–CNN–SE–resample and FR–ResNet–resample model) are shown in Figure 8 and

Table 3. Percentages of different landslide susceptibility classes.

Susceptibility Class	FR-CNN-Resample	FR-CNN-SE-Resample	FR-ResNet-Resample
very low	47.68	45.47	56.38
low	18.79	18.98	13.31
moderate	10.49	13.25	6.91
high	7.93	7.95	5.11
very high	15.1	14.35	18.29

. The distributions of landslide susceptibility obtained using the different models were similar. Susceptible landslide areas were widely distributed throughout the study area. More than 50% of the study area has low landslide susceptibility, and no more than 20% of the area is extremely sensitive. The susceptible area, the area close to the water system, is the most vulnerable to landslides. In addition, the high susceptibility regions generated by the ResNet (Figure 8c) method are significantly larger than those generated by the CNN (Figure 8a) and CNN–SE (Figure 8b) methods, and the high susceptibility regions based on the ResNet method are relatively concentrated. The results of the three methods for the low-susceptibility regions are essentially the same.

FR < 1 indicates that the state is unfavorable to landslides, whereas FR > 1 indicates that it is favorable to landslides. The FR values of the high-sensitivity area and the extremely high-sensitivity area shown in Figure 9a are more significant than 1, and the FR value of the extremely high-sensitivity area is much higher than 1, which is consistent with the theory. As many landslide hazard points as possible were located in highly prone areas. Combined with the LRPD in Figure 9, it can be seen that the LRPD value and FR value increase significantly with the increase of the susceptibility level. The FR and LRPD values were the highest in the extremely high-sensitivity areas, and the LRPD value in the very low-sensitivity region was zero. This shows that most landslides were located in high-sensitivity and extremely high-sensitivity areas, and very few were located in medium-sensitivity areas. No landslides appeared in low-sensitivity areas, which strongly proves the learning and prediction ability of the deep learning model constructed in this study.

4.4. Model Verification and Comparison

In addition, FR, MLP neural networks, and SVM models were selected for comparison with deep-learning models to further validate the effectiveness of deep learning for LSM applications.

Visually, the landslide susceptibility grading maps obtained by these three methods were similar to the distribution of the results obtained by the deep-learning model, and the landslide highly sensitive areas were nearly uniformly distributed throughout the study area. From Figure 10, it can be seen that the landslide susceptibility values obtained by the FR model are smaller than those obtained by the other models, the highly sensitive areas obtained by the deep learning and the other two models are significantly larger than those obtained by the FR model, and the area of the very highly sensitive area obtained by deep learning is the largest. This indicates that the deep learning model is more practical for landslide prevention and management.

In terms of accuracy, the AUC values of 78.03%, 86.34%, and 87.27% for the FR model (Figure 10a), MLP artificial neural network model (Figure 10b), and support vector machine model (Figure 10c), respectively, were lower than those obtained using the deep-learning method (

Table 4. Accuracy evaluation of different models.

Model	Sampling Method	Precision	Accuracy (%)	Recall	MCC	F1 Score
CNN	Random	0.7903	81.10	0.8167	0.6219	0.8033
	FR	0.8796	86.05	0.8293	0.7217	0.8537
	Random-resample	0.9189	91.05	0.8939	0.8209	0.9063
	FR-resample	0.9217	92.98	0.9395	0.8597	0.9305
CNN-SE	Random	0.8205	82.28	0.8000	0.6443	0.8101
	FR	0.8823	90.52	0.9312	0.8118	0.9061
	Random-resample	0.9187	93.15	0.9420	0.8633	0.9302
	FR-resample	0.9617	94.38	0.9246	0.8883	0.9428
ResNet	Random	0.8071	86.61	0.9417	0.743	0.8692
	FR	0.9577	92.72	0.8910	0.8561	0.9231
	Random-resample	0.9632	95.74	0.9483	0.9148	0.9557
	FR-resample	0.9769	97.01	0.9632	0.9404	0.9700

). This indicates that deep-learning-based methods have excellent goodness-of-fit and strong predictive ability.

5. Discussion

In this study, non-landslide samples were selected using the FR model, and resampling was performed to obtain more accurate results. The influence of landslide influencing factors on landslide development was analyzed using FR and landslide point density. Not only is the environmental information around the landslide more fully utilized, but the data representation problem and non-landslide sample selection problem of the CNN model in LSM application are also solved. Compared with traditional methods, CNN achieves more accurate results, and the effects of FR selection methods and resampling operations on CNN modeling are further explored below.

Based on the accuracy, precision, recall, and MCC value, the model based on the FR selection method (FR, FR–resample in Table 4) outperformed the model based on the buffer random sampling method (random, random–resample in Table 4) when the selected base deep learning and resampling methods were the same. Overall, the accuracy of the deep learning models obtained from the two non-slippery slope sample selection methods ranged from 81.10 to 97.01%, indicating satisfactory model performance. Specifically, the accuracy, precision, recall, and MCC values of each model based on the FR sample selection method were higher than those of the same model based on the buffer random sampling method. The results of the accuracy analysis are shown in Table 4, which shows that the FR non-landslide selection method is feasible and effective, and the buffer random sampling method has a higher subjectivity and lack of adequate verification.

However, with the same base deep-learning model and resampling operations performed, the FR-based selection method does not provide a significant improvement in accuracy over the buffer random sampling method, with only 0.0028–0.043 improvement in precision, 1.23–1.93% improvement in accuracy (Table 4), and 0.35–0.78% improvement in AUC values (Figure 11a–c). This is due to the fact that deep neural networks have a strong expressive power compared to traditional models, thus requiring more data to avoid the occurrence of overfitting. The resampling operation increases the sample size to a certain extent. For deep learning, which requires a large number of samples, sufficient samples after resampling can make the deep-learning model fit more accurately and reduce the impact of the accuracy of non-landslide slope sample selection on the accuracy of the model.

When the base deep-learning model and non-landslide sample selection methods were selected in the same way, the models constructed after the resampling operation (Random–resample, FR–resample in Table 4) performed better than the models constructed without resampling (random, FR in Table 4), with MCC 0.172–0.219 higher and accuracy 9.13–10.87% higher. This indicates that the CNN, CNN–SE and ResNet models constructed after resampling have higher performance. The resampling method makes the landslide point data have a three-dimensional data form, utilizes the environmental information around the landslide, makes full use of the multi-dimensional information of the data, and is more conducive to the model learning the complex relationship between the landslide and its influencing factors.

To evaluate the generalization ability of the deep learning models constructed in this study, the ROC (Figure 11a–c) and PR (Figure 11d–f) of each model were plotted. The three deep-learning models based on the FR selection and resampling methods had the highest AUC values of 98.18%, 98.55%, and 99.64%, respectively. The PR curve is drawn with the precision rate and recall rate as the horizontal and vertical coordinates, and the BEP value is the value of precision when precision is equal to recall. The larger the value, the higher the performance of the model. The results show that the model based on FR selection and resampling outperforms the buffer random sampling method. In addition, the F1 values of CNN, CNN-SE, and ResNet models based on FR non-landslide sample selection and resampling method are 0.9305, 0.9428, and 0.9700, respectively, which are higher than those of CNN, CNN–SE, and ResNet models with other sampling methods, which further proves the feasibility and effectiveness of FR-based non-landslide sample selection method and resampling operation, and the deep–learning model constructed has excellent goodness-of-fit and strong prediction ability.

6. Conclusions

In this study, considering the insufficient samples of landslide data, the problem of data representation in the application of CNN to LSM, and the subjectivity of traditional landslide and non-landslide sample selection methods. Eleven evaluation factors, such as elevation, distance from the road, distance from the water system, and land-use type, establish different sample data sets through different non-landslide sample selection methods and data expression methods, and apply them to the constructed CNN, CNN–SE, and ResNet models. Finally, the landslide susceptibility classification results obtained from the different models were analyzed. The main findings of this study are as follows:

(1)

Deep-learning structures such as CNN can analyze the relationship between complex landslide influencing factors and landslide occurrence and are practical for analyzing landslide susceptibility. The deep-learning models constructed in this study were successful in both the training and validation phases, with the CNN, CNN–SE, and ResNet models all achieving AUC values above 85%, which is higher than other traditional models including FR models and support vector machines.

(2)

The non-landslide samples selected by the FR-based method are more accurate and objective than those selected by the traditional buffer random sampling method. Compared with the traditional buffer random sampling method, the accuracy of the model constructed by the FR-based non-landslide sample selection method was significantly improved, the AUC was increased by 3.98–7.20%, and the obtained landslide susceptibility classification was more reasonable.

(3)

The resampling operation can effectively use the information of the surrounding environment of the landslide, which is conducive to the model learning deeper information and improving its generalization ability and robustness. This also helps, to a certain extent, solve the problem of insufficient CNN learning owing to inadequate samples. Compared with the model constructed from the unsampled dataset, the model constructed from the resampled sample dataset had higher accuracy, with an improved AUC of 1.34–8.82%.

In summary, this study proves that CNN has excellent potential in the analysis of landslide susceptibility and provides countermeasures for the lack of samples and the expression of sample data in CNN applications. Although the resampling operation increases the number of samples to a certain extent, the application of deep learning in LSM still requires more samples. In the future, we should explore more effective solutions to the problem of insufficient sample size.