Enhanced Landslide Susceptibility Assessment in Western Sichuan Utilizing DCGAN-Generated Samples

Abstract

The scarcity of landslide samples poses a critical challenge, impeding the broad application of machine learning techniques in landslide susceptibility assessment (LSA). To address this issue, this study introduces a novel approach leveraging a deep convolutional generative adversarial network (DCGAN) for data augmentation aimed at enhancing the efficacy of various machine learning methods in LSA, including support vector machines (SVMs), convolutional neural networks (CNNs), and residual neural networks (ResNets). Experimental results present substantial enhancements across all three models, with accuracy improved by 2.18%, 2.57%, and 5.28%, respectively. In-depth validation based on large landslide image data demonstrates the superiority of the DCGAN-ResNet, achieving a remarkable landslide prediction accuracy of 91.31%. Consequently, the generation of supplementary samples via the DCGAN is an effective strategy for enhancing the performance of machine learning models in LSA, underscoring the promise of this methodology in advancing early landslide warning systems in western Sichuan.

1. Introduction

Landslides represent a prevalent form of geological hazard, characterized by their extensive destructiveness and widespread occurrence. They constitute a significant threat to public safety, inflicting substantial socioeconomic losses [1]. Western Sichuan is affected by the uplift and compression of the Qinghai–Tibet Plateau, as well as the erosion of water systems. The geological structure is complex and active, leading to frequent earthquakes. Landslides, as one of the secondary geological hazards triggered by strong earthquakes, constitute the predominant geological hazard type in western Sichuan [2,3]. Landslide susceptibility assessment (LSA) involves analyzing the spatial distribution of historical landslides along with environmental factors, such as topography, geomorphology, geological formations, and hydrometeorology, to predict the risk of landslides in the study area [4]. Validating landslide occurrence mechanisms and conducting precise landslide susceptibility predictions serve as a crucial reference basis for disaster prevention and engineering mitigation.

Numerous scholars have conducted research on regional landslide susceptibility, with methodologies evolving from qualitative analysis to quantitative and intelligent assessment. The quantitative methods can be classified into deterministic and non-deterministic approaches. Deterministic methods employ geomechanical models alongside spatial geological information to predict the probability of landslide occurrences in specific regions, such as the Shallow Landslide Stability Model (SHALSTAB) proposed by Dietrich et al. [5] and the Stability Index Map (SINMAP) proposed by Pack et al. [6]. Non-deterministic methods are divided into knowledge-driven methods based on expert experience and data-driven methods based on data analysis. Their critical disparity lies in the determination of factor weights [7].

Since the 1990s, the field of artificial intelligence has undergone significant development, witnessing the emergence of various machine learning models such as random forests (RFs), support vector machines (SVMs), artificial neural networks (ANNs), and convolutional neural networks (CNNs). These models have showcased robust data mining capabilities, adept at handling large volumes of heterogeneous data from diverse sources. Their wide applications span across fields such as geological hazard research, medical research, economics, etc. In recent years, various machine learning methods have been applied to LSA; Abdelaziz and Bahare, respectively, conducted literature reviews and studies based on landslides in Iran, comparing various algorithm models, and both achieved favorable results [8,9]. Compared to traditional LSA methods, deep neural networks extract features from various influencing factors in landslide areas more efficiently, thus exhibiting superior prediction accuracy [10,11]. Nevertheless, they also require a larger amount of data samples for training [12]. Hussin et al. found that when limiting the amount of training data below a certain threshold, the model showed significantly poor performance [13]. Numerous studies demonstrated that methods such as geometric transformations, color space transformations, and adding noise can partially address the issue of sample scarcity [14,15,16]. However, research by Shorten and Khoshgoftaar suggests that these methods pose risks of losing crucial information and contaminating the dataset, consequently leading to a decline in model performance [17]. The emergence of generative adversarial networks (GANs) has provided a new way to address these challenges [18]. Gu et al. introduced an enhanced SinGAN method for generating remote-sensing images [19]. Their study verified that the generated images exhibited high structural similarity and logical consistency with real images. These images effectively serve as training data for the model, yielding satisfactory results. Husam and Biswajeet generated additional landslide samples for training through GAN and demonstrated that additional samples can effectively improve the performance of LSA models constructed using ANN and RF [20].

GANs show great potential in addressing the issue of insufficient training samples. However, GANs suffer from training difficulty and instability [19,21]. Therefore, this study proposes a data augmentation technique built upon a DCGAN, leveraging its convolutional structures for enhanced stability and expedited convergence during training. Subsequently, the augmented samples are used to improve the effectiveness of machine learning methods in landslide sensitivity assessment [21]. The remaining sections of this article are organized as follows: Section 2 introduces the overview of the study area and the sources of data. Section 3 describes the experimental process, highlighting the machine learning models, LSA modeling methodologies, and the evaluation metrics of these models. Then, Section 4 presents the results of each step in the experimental process. Moreover, Section 4 verifies the feasibility of enhancing the performance of LSA models through data augmentation using a DCGAN. Section 5 discusses the environmental factors and additional sample numbers. Finally, Section 6 summarizes this article.

2. Materials

2.1. Study Area

The western Sichuan region is shown in Figure 1a, and it is situated in the transitional zone between the eastern edge of the Qinghai–Tibet Plateau and the Chengdu Plain. In terms of administrative divisions, the study area includes Ya’an City, Panzhihua City, Aba Prefecture, Garze Prefecture, and Liangshan Prefecture. The terrain of the study area is rugged, with elevations ranging from 300 to 7500 m (shown as Figure 1b). The topography exhibits significant variations, mainly consisting of mountainous and hilly areas, transitioning gradually from plain areas in the east to hilly regions and high mountain valleys in the west. Under the influence of topography, the study area is characterized by vertical climatic zones. In the valley areas, the climate is mainly subtropical, warm–temperate, and temperate, while in the mountainous plateau regions, it is primarily sub-frigid and frigid. Annual rainfall is unevenly distributed, primarily concentrating between May and October, with precipitation ranging from 400 to 1400 mm [22]. Its spatial distribution shows a decreasing trend from east to west. The geological structure of the western Sichuan region is shown in Figure 1c, with stratigraphic fault zones distributed both vertically and horizontally; the predominant rock types in the study area include clastic rock, metamorphic rock, continental deposit, carbonatite, and basalt [22]. Among them, sandstone, mudstone, and siltstone within the clastic rocks collectively form the easily erodible and prone-to-landslide strata [23].

Affected by natural and anthropogenic factors, geological disasters such as collapses, landslides, and debris flows frequently occur in western Sichuan. These disasters are characterized by their widespread occurrence, extensive impact, and rapid onset, causing severe losses to the safety of people’s lives and property. Therefore, western Sichuan is an ideal region for validating LSA methods [24].

2.2. Inventory of Landslides

The historical landslide point data catalog was established based on survey data of historical geological disasters in Sichuan and combined with remote sensing images. A total of 3515 historical landslide points was identified in western Sichuan (shown as Figure 1b). After filtering out repeated occurrences of landslides at the same location and consolidating multiple landslides within small areas, 2000 historical landslide points were selected as the base data for the study. Among them, small-to-medium-sized landslides prevail, accounting for 95% of the total number of landslides. The patterns of landslides include sliding–cracking, creeping–shearing, and tilting–fracturing [22]. In addition, in LSA, the selection of negative samples significantly influences the prediction results. Negative samples are typically selected from areas outside of the landslide region randomly, making it difficult to ensure their reliability and consequently leading to a decrease in model performance [25]. Therefore, efforts should be made to select sample points from areas with lower landslide risk. This study utilizes the information value (IV) model to select non-landslides by analyzing the spatial distribution of landslides and their relationship with influencing factors, thereby quantifying the contribution of each factor to landslide occurrence [26]. The methodology involves reclassifying the factors, calculating the proportion of landslide and non-landslide pixels within each category, and determining the IV for each category. Subsequently, the IVs of all factors are aggregated and reclassified to produce a preliminary landslide risk map for the study area. Finally, a landslide dataset was constructed by randomly selecting non-landslide samples from areas with low landslide risk at a 1:1 ratio, as shown in Figure 2.

The validation dataset was constructed based on the aerial image interpretation dataset of landslides in Sichuan and surrounding areas published by Zeng et al., as well as high-definition historical imagery from Google Earth [27]. Thirteen large landslide events that occurred in western Sichuan from 2017 to 2022 were identified based on information gathered from the relevant literature and news sources.

2.3. Landslide Factors

Selecting landslide factors is crucial for constructing LSA models, typically encompassing terrain, land cover, hydrometeorological conditions, engineering geology, and human activities. Their selection is often based on the relationship with historical geological hazards.

In the context of landslide triggers in western Sichuan, a total of 18 factors were chosen as inputs for the LSA models. The data sources are shown in

Table 1. Data sources of landslide factors.

Data Name	Sources	Type	Resolution
Elevation	NASA SRTM DEM	raster	30 (m)
Geological faults	Active tectonic map of China (1:4,000,000)	vector	1:4,000,000
Road	National Basic Geographic Database	vector	1:250,000
River system	National Basic Geographic Database	vector	1:250,000
Lithology	National 1:200,000 digital geological map	vector	1:200,000
Landform	National 1:200,000 digital geological map	vector	1:200,000
Soil	1:1 million soil map of the People’s Republic of China	vector	1:1,000,000
Precipitation	China Meteorological Administration (CMA)	station	-
Land use	GLOBELAND30	raster	30 (m)
NDVI	Google Earth Engine	raster	30 (m)

[28,29,30]. All data were resampled to a spatial resolution of 30 m, and continuous variables such as elevation, slope, and the NDVI were reclassified using the Jenks natural breaks classification method (shown in Figure 3) [31].

Topographic factors: Topographic factors include elevation, aspect, plan curvature, profile curvature, slope, relief amplitude, the topographic wetness index (TWI), and the landform. Except for the landform, which is obtained from the national 1:200,000 digital geological map, the remaining factors are calculated from 30 m resolution elevation data. These factors play an essential role in the development of landslides. For example, elevation and aspect have important control functions. Elevation differences can lead to variations in vegetation and potential energy within a region. Moreover, different aspects are affected by different sunlight levels, resulting in spatial differences in soil moisture, thereby influencing the development of landslides. The frequency ratio (FR) analysis of partial typical factors indicates that when the elevation is below 2700 m, the landslide frequency consistently exceeds 1 (

Table 2. Frequency ratios of some typical landslide factors.

Landslide Factors	Value	Landslides	Pixel Number	Frequency Ratios
Elevation (m)	<1500	762	12,521,106	10.701
	1500~2100	566	22,803,722	4.364
	2100~2700	378	32,203,639	2.064
	2300~3200	152	31,160,160	0.858
	3200~3600	66	43,476,058	0.267
	3600~4000	51	59,416,050	0.151
	4000~4300	14	54,487,458	0.045
	4300~4600	8	58,221,426	0.024
	>4600	3	37,378,450	0.014
NDVI	<0.1	26	11,539,489	0.396
	0.1~0.3	147	36,268,491	0.713
	0.3~0.5	788	159,049,040	0.871
	0.5~0.7	886	112,372,150	1.386
	>0.7	153	32,438,852	0.829
Precipitation (mm)	<550	155	81,219,497	0.336
	550~650	256	93,794,711	0.480
	650~750	490	106,941,315	0.806
	750~1000	798	65,695,362	2.136
	>1000	301	4,017,184	13.175
Distance to a river (km)	<1	519	29232080	3.122
	1~3	585	56,067,215	1.835
	3~6.3	402	85,449,923	0.827
	6.3~9.6	220	70,887,607	0.546
	9.6~13	143	52,398,139	0.480
	13~17	80	34,924,928	0.403
	>17	51	22,708,218	0.395
Lithology	A (Clastic rock)	724	107,987,222	1.179
	B (Granite)	34	6,525,398	0.916
	C (Metamorphic rock)	269	133,115,327	0.355
	D (Continental deposit)	317	49,941,742	1.116
	E (Urban)	239	9,216,852	4.560
	F (Lake)	3	111,410	4.735
	G (Carbonatite)	251	31,535,669	1.400
	H (Outcrop)	0	660,414	0
	I (Basalt)	163	12,401,449	2.311
Distance to road (km)	<0.5	964	32,899,167	5.152
	0.5~1.5	458	47,514,033	1.695
	1.5~3.5	314	71,330,026	0.774
	3.5~6	136	61,810,401	0.387
	6~10	63	58,396,257	0.190
	10~15	31	36,685,427	0.149
	>15	34	43,032,799	0.139

), suggesting a higher likelihood of landslide occurrences within this elevation range [32].

Land cover and hydrometeorological factors: Land cover and hydrometeorological factors have important influences on landslide occurrence, including the stream power index (SPI), sediment transport index (STI) [33], distance to rivers, precipitation, land use, and the normalized difference vegetation index (NDVI). For example, the NDVI reflects the land cover types and vegetation growth status in the study area. Regions with dense vegetation tend to have more stable soil, resulting in lower landslide probability. Conversely, factors such as precipitation and watercourses increase soil moisture and slope erosion, leading to decreased shear strength of the soil and slope stability. In the study area, regions with NDVI values ranging from 0.5 to 0.7 have frequency ratios over 1. Similarly, areas with precipitation exceeding 750 mm and located within 3 km of rivers also have higher frequency ratios than others. This indicates that landslides are more likely to occur in these environments.

Engineering geology and human activities: The frequency and magnitude of landslides vary depending on the geological formations present. Historical records of geological hazards in western Sichuan indicate that landslides and collapses are particularly prevalent in debris rock layers primarily composed of sandstone and mudstone. Moreover, the construction of roads can exacerbate land degradation and soil displacement, consequently increasing slope instability. Notably, within a 1.5 km radius of roads in the study area, the frequency of landslides is significantly higher compared to other regions.

3. Methods

In this study, additional samples were generated using a DCGAN to augment the dataset, facilitating LSA in western Sichuan through an SVM, CNN, and ResNet. The experimental procedure is outlined as follows (refer to Figure 4): (1) compilation of the historical landslide inventory in the study area and acquisition of 18 fundamental environmental factors. (2) Conducting multicollinearity analysis and employing GeoDetector to eliminate irrelevant factors associated with landslides. Principal component analysis (PCA) was utilized to mitigate redundant information and streamline model complexity. (3) Expansion of the landslide dataset via a DCGAN, followed by the training of LSA models (SVM, CNN, ResNet) using both the original and augmented datasets. (4) Generation of landslide susceptibility prediction maps for western Sichuan using the trained models, followed by validation using large-scale landslide data.

3.1. VIF Analysis and GeoDetector

This study quantitatively evaluates the contributions of 18 landslide factors to landslides by calculating the variance inflation factor (VIF) and GeoDetector (GeoDetector_2015_example in Excel). The VIF value quantifies the degree to which the variance of a regression coefficient is inflated due to correlations among independent variables. A higher VIF value and a lower tolerance (TOL) value indicate higher multicollinearity for the factor. The GeoDetector, proposed by Wang et al. [34], is a statistical method based on spatial autocorrelation theory. Its core idea is that if a factor has a significant impact on a geographical event, then there will be a similarity in the spatial distribution of the factor and the geographical event. In this study, the differentiation and factor detection models of the geographic detector are applied to quantitatively analyze the strength of factors in explaining spatial differentiation of events [35,36]. Factor detection is measured using the q statistic, which is in the range of [0, 1]. A higher q-value indicates a stronger explanation of the factor for landslides.

3.2. Data Augmentation Based on DCGAN

A DCGAN, proposed by Radford et al., represents a generative adversarial network incorporating deep convolutional structures [37]. It is employed for unsupervised learning and the generation of similar images. A DCGAN amalgamates CNN architecture into traditional GANs with specific modifications (refer to Figure 5). Notably, the generator and discriminator dispense with the pooling layers of the CNN, opting instead for batch normalization (BN) layers following each convolution (transpose convolution) operation. The model is described by the following mathematical equation [21]:

$$\underset{\mathrm{G}}{\min }\underset{\mathrm{D}}{\max }\mathrm{V}\left(\mathrm{D},\mathrm{G}\right){=\mathit{E}}_{{\mathit{x}~\mathit{P}}_{\mathit{data}}\left(\mathit{x}\right)}\left[\log \mathrm{D}\left(\mathit{x}\right)\right]{+E}_{{z~P}_z\left(z\right)}\left[\log \left(1 - \mathrm{D}\left(\mathrm{G}\right(z\left)\right)\right)\right]$$

(1)

where $\underset{\mathrm{G}}{\min }\underset{\mathrm{D}}{\max }\mathrm{V}\left(\mathrm{D},\mathrm{G}\right)$ represents the convergence direction of the generator and discriminator, $\mathit{x}$ denotes real samples, and $\mathrm{G}\left(\mathit{z}\right)$ represents samples generated from random vectors. ${\mathit{P}}_{\mathit{data}}$ and ${\mathit{P}}_{\mathit{z}}$ denote the real data distribution and the generated data distribution identified by the discriminator. The ultimate objective of the model is to find $\underset{\mathrm{G}}{\min }$ that produces synthesized data resembling real data.

3.3. LSA Machine Learning Models

3.3.1. Support Vector Machine (SVM)

An SVM is a machine learning algorithm rooted in statistical theory, designed to minimize structural risk [38]. Within the field of landslide susceptibility research, SVMs find extensive application, particularly with the radial basis function (RBF) kernel, favored for its ability to explore complex, nonlinear mappings [39]. In this study, RBF-SVM is employed as one of the LSA models. The performance of RBF-SVM is influenced by parameters such as the kernel width parameter γ and the regularization parameter C. We utilize the original dataset as input and identify the optimal parameters through grid search [40,41], in which γ = 0.01, C = 1.

3.3.2. Convolutional Neural Network (CNN)

A CNN is a widely used feed-forward neural network. In this study, a CNN is selected to construct an LSA model, exploring the relationship between landslides and environmental factors. The model maps the probability of landslide occurrence onto the interval [0, 1], treating the task as a regression problem to achieve accurate landslide prediction [42]. The CNN architecture includes three convolutional layers and BN layers for training. A maximum pooling layer is inserted between the convolutional layers to filter redundant information. Finally, a fully connected layer is employed to propagate features and adjust them to the appropriate size. The output layer employs 2 neuron outputs to represent landslide (output of 1) and non-landslide (output of 0) categories [43,44].

3.3.3. Residual Neural Network (ResNet)

A ResNet is an improved convolutional neural network proposed by He et al. [45] that addresses the problem of gradient disappearance by introducing the concept of residual learning. The residual mapping layer is superimposed onto the shallow network to create the residual block through skip connections. This innovation significantly improves accuracy and mitigates gradient loss [46]. The output of the function mapping layer can be expressed as [23]

$$\mathit{y}=\mathrm{H}(\mathit{x},W_i)+\mathit{x}$$

(2)

where $\mathit{x}$ and $\mathit{y}$ refer to the input and output of the residual block, respectively. $\mathrm{H}(\mathit{x},W_i)$ is the residual mapping parameter. In this study, the ResNet18 model with 4 stages and a total of 8 base residual blocks is employed.

3.4. Evaluation Indicators

To evaluate model performance, five metrics are calculated based on the confusion matrix, including Accuracy (ACC), Precision (PRE), True Positive Rate (TPR), F1-score, and the Matthews Correlation Coefficient (MCC). Among them, ACC, PRE, and TPR fall within the range [0, 1], with values closer to 1 indicating superior performance. These metrics can be calculated as follows:

$$\mathrm{ACC}={\displaystyle \frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}}$$

(3)

$$\mathrm{PRE}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}}$$

(4)

$$\mathrm{TPR}={\displaystyle \frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}}$$

(5)

$$\mathrm{F}1-\mathrm{score}={\displaystyle \frac{2 \times \mathrm{PRE} \times \mathrm{TPR}}{\mathrm{PRE}+\mathrm{TPR}}}$$

(6)

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

(7)

Furthermore, the performance of classification models can be evaluated using receiver operating characteristic (ROC) curves, where the horizontal axis represents the false positive rate (proportion of correctly predicted non-landslide samples) and the vertical axis represents the true positive rate (proportion of correctly predicted landslide samples). The ROC curve reflects the classification effectiveness of the model, quantified by the area under the curve (AUC). A higher AUC value indicates a higher classification accuracy of the model.

4. Results

4.1. Landslide Factor Screening

The multicollinearity among initially selected landslide factors can distort the LSA model during training, leading to convergence difficulties or random predictions. Additionally, an excess of landslide factors inevitably introduces redundant information, thereby affecting the accuracy of predictive results. The VIF results of landslide factors are presented in

Table 3. Multi-collinearity analysis of landslide factors.

Landslide Factors	VIF	TOL
Aspect	1.00676	0.99329
Lithology	1.09788	0.91085
Profile curvature	1.1011	0.90818
Plan curvature	1.22913	0.81358
Distance to fault	1.28592	0.77765
Landform	1.35468	0.73818
NDVI	1.41616	0.70613
Land use	1.45153	0.68892
Distance to a river	1.45996	0.68495
Soil	1.8564	0.53868
Relief amplitude	1.99312	0.50173
Precipitation	2.34496	0.42645
Distance to road	2.64131	0.37860
SPI	4.15778	0.24051
TWI	4.18707	0.23883
Elevation	5.88457	0.16994
STI	9.73319	0.10274
Slope	11.0813	0.09024

, where VIF values closer to 10 indicate more severe collinearity. Specifically, the VIF values of the STI and slope factors are 9.733 and 11.081, respectively, indicating collinearity among these factors. Consequently, the STI and slope factors are excluded from subsequent analyses. The q-value results are shown in Figure 6. Factors such as the SPI, profile curvature, the TWI, plan curvature, and aspect have q-values less than 0.01, indicating a weak explanation for landslide events and thus are considered redundant information. Consequently, a total of 11 factors are selected for subsequent LSA model construction.

4.2. Quality Analysis of Additional Landslide Samples

This study augmented the landslide dataset to enhance the performance of three landslide susceptibility models, namely an SVM, CNN, and ResNet18. The model hyperparameters are shown in

Table 4. Model hyperparameters.

Hyperparameter	DCGAN	CNN	ResNet18
Kernel	4 × 4	3 × 3	3 × 3
Pooling	-	2 × 2	3 × 3
Activation Function	D: LeakyReLU G: Tanh,ReLu	Tanh	ReLU
Optimizer	Adam	Adam	Adam
Loss Function	BCELoss	CELoss	CELoss
Learning Rate	D: 0.0002 G: 0.001	5 × 10−4	5 × 10−4
Epoch	6000	150	150

. When employing generated samples for LSA, an excessive number of additional samples may lead to issues such as indistinct features or uniform sample characteristics, potentially resulting in decreased model performance. Therefore, the optimal number of samples needs to be determined. In this study, six augmented datasets were generated by the DCGAN, with the number of additional samples increasing incrementally from 100 to 600 (100, 200, 300, 400, 500, 600). Subsequently, predictions were made using LSA models trained on these datasets. Training accuracy, testing accuracy, and the AUC were utilized as evaluation metrics to assess the accuracy and classification capabilities of the models, aiming to determine the optimal number of additional samples. The experimental results are as follows (Figure 7):

(1)

When the number of additional samples is 100, 200, and 300, the testing accuracies of the SVM are improved by 2.56%, 2.66%, and 2.3%, respectively. However, those of the CNN and ResNet show no significant changes, and there is a difference of approximately 5% to 10% between the accuracy of the training and testing datasets, indicating the presence of overfitting.

(2)

With 400 and 500 additional samples, the testing accuracies of the SVM are improved by 1.88% and 2.18%, respectively. The testing accuracies of the CNN are improved by 1.54% and 2.57%, respectively. Moreover, those of the ResNet are improved by 3.37% (400) and 5.28%, respectively.

(3)

When there are 600 additional samples, compared to 200 additional samples, the testing accuracy of the SVM is decreased by 2.31%. However, it is improved by 0.35% compared to the original dataset. Additionally, compared to 500 additional samples, the testing accuracies of the CNN and ResNet were decreased by 2.11% and 4.3%, respectively, although compared to the original dataset, they were improved by 0.46% and 0.98%, respectively.

The simple structure of the SVM determines that it does not require a large amount of data during training, and it is less sensitive to the number of additional samples. Consequently, the improvement in prediction accuracy compared to the original dataset is always limited. On the other hand, deep neural network models like the CNN and ResNet can leverage larger datasets more effectively during training, enabling them to better explore the underlying relationships within the data. Therefore, as the number of additional samples increases, the performance of these models also improves. However, the utilization of pseudo samples generated from real samples may lead to the dataset becoming less diverse, thus reducing its quality. According to the test result, both the CNN and ResNet show significant improvements when training with 400 and 500 additional samples. Specifically, in the case of 400 additional samples, the CNN exhibits the best landslide classification performance (AUC = 92.42%), while the ResNet performs best with 500 additional samples (AUC = 93.61%). Taking all factors into account, when there are 500 additional samples, the SVM, CNN, and ResNet show the greatest improvement. Therefore, the final LSA models are trained on datasets with 500 additional samples in this study.

4.3. LSA Results

4.3.1. Effectiveness of LSA Machine Learning Models

The quantitative evaluation results of the LSA models trained with the original dataset and the DCGAN-augmented dataset (500 additional samples) are shown in

Table 5. Evaluation of landslide susceptibility assessment models.

Model	ACC%	PRE	TPR	F1-Socer	MCC
SVM	77.08	0.7627	0.7607	0.7604	0.5236
CNN	86.69	0.8631	0.8731	0.8681	0.7339
ResNet	86.50	0.8718	0.8401	0.8557	0.7294
DCGAN-SVM	79.26	0.7927	0.7906	0.7826	0.5853
DCGAN-CNN	89.26	0.9074	0.8741	0.8899	0.7856
DCGAN-ResNet	91.78	0.9149	0.9094	0.9121	0.8257

. On the original dataset, the SVM performs poorly with an accuracy of 76.08%, while both the CNN and ResNet achieve similar accuracies of over 86%, demonstrating excellent performance in landslide prediction. Based on comprehensive evaluation metrics, the SVM shows the lowest performance. The CNN has a lower PRE value than the ResNet but higher TPR, F1-score, and MCC values than the ResNet. These results suggest that the CNN achieves superior overall prediction accuracy, albeit with a tendency to misclassify some non-landslides as landslides. From a landslide prevention perspective, the prediction pattern of the CNN can effectively reduce the risk of landslides in densely populated and built-up areas, while the prediction pattern of the ResNet is more suitable for sparsely populated and high-altitude mountainous areas.

Training with the DCGAN-augmented dataset substantially enhances the performance of all LSA models, with the ResNet demonstrating optimal performance. Compared to training on the original dataset, the ACC of the CNN is increased by 2.57%, while other metrics are increased by 0.03 to 0.05, indicating higher overall landslide prediction accuracy and greater sensitivity in identifying landslides and non-landslides. The improvement of the ResNet is more significant, with an ACC reaching 91.78%, marking an impressive increase of 5.28%. Moreover, metrics such as PRE, TPR, and F1-score all surpass 0.9, and the MCC is increased by approximately 0.1. The results suggest that training the ResNet model with the DCGAN-augmented dataset leads to enhanced overall prediction accuracy and reduced misclassification of landslides and non-landslides.

The ROC curve reflects the model’s prediction accuracy and error rate (Figure 8). In the training of the original dataset, the classification performance of the CNN (AUC = 80.23%) is lower than the SVM (AUC = 82.17%) and ResNet (AUC = 81.09%), despite the CNN having the highest accuracy. However, the SVM exhibits a more stable landslide classification. With an additional 500-sample dataset, both the CNN (AUC = 91.01%) and ResNet (AUC = 93.61%) show significant improvements in classification performance. Compared to training on the original dataset, the CNN’s AUC increases by 10.78%, while the ResNet demonstrates the best classification performance, with an AUC increase of 12.52%.

4.3.2. Analysis of Landslide Susceptibility Mapping (LSM)

The CNN and ResNet, which perform better, were selected as classifiers for LSM. Figure 9 shows the LSM results in western Sichuan based on the models trained with the original dataset and DCGAN-augmented dataset, as well as detailed maps. For the original LSM results, the distributions of landslide-prone areas classified by the two LSA models are generally similar, with high-risk areas concentrated in the eastern and southern parts of the study area, and the probability of landslides decreasing as the altitude increases towards the west. However, an in-depth analysis of regions (a) and (b), where landslides are more frequent, reveals that the CNN-predicted map shows a larger area of very high landslide probability compared to the ResNet. The CNN shows a tendency to over-predict the landslides, indicating a greater inclination to classify some moderate and high-risk areas as very high-risk areas. Region (a) is characterized by high-altitude mountains, canyons, and grasslands with sparse populations, resulting in limited economic losses from landslides. In landslide prevention and control in this area, prediction from the ResNet is more feasible. In contrast, region (b) encompasses several cities characterized by low altitudes and dense populations. As mentioned earlier, the “over-prediction” of the CNN is more in line with the practical need for landslide prevention in these areas. According to the LSM results using the models trained with the DCGAN-augmented dataset, the distribution of risk areas still follows the trend of high-risk areas being concentrated in the eastern and southern parts of the study area, which are characterized by low altitudes and high human activity. As the altitude increases towards the west, the probability of landslides decreases. Compared to the original LSM results, the main difference in predictions lies in the very low-risk areas. Combining the percentage of landslide sensitivity zones depicted in Figure 10, it is observed that the proportion of very low-risk areas predicted by the DCGAN-ResNet is significantly higher. Additionally, compared to the ResNet model trained on the original dataset, the predicted moderate-risk areas of the DCGAN-ResNet model are decreased noticeably.

4.3.3. Validation Based on Large-Scale Landslide Images

This study validated the LSM results by visually interpreting 13 large landslide areas that occurred in western Sichuan from 2017 to 2022 with Google’s historical high-resolution imagery and high-resolution aerial orthophotos.

Table 6. Validation of landslide susceptibility mapping.

Model	Very Low	Low	Moderate	High	Very High
CNN	0	0	12.24%	28.68%	59.08%
ResNet	0	0	19.62%	60.42%	19.96%
DCGAN-CNN	0	0	16.37%	22.65%	60.98%
DCGAN-ResNet	0	4.15%	5.79%	24.79%	66.5%

presents the prediction results of these 13 landslides based on area statistics, with six major landslides shown in Figure 11. It is evident that in the validation of the original LSA results, the CNN is significantly superior to the ResNet. When the model predicts landslide areas as high and very high-risk areas, it is considered a successful prediction. The success rate of the CNN prediction reached 87.76%, which is 7.38% higher than the ResNet. Particularly, in the case of the Jiuzhaigou landslide group, while the CNN successfully predicted the majority of landslide areas, the ResNet classified most of them as moderate-risk areas. The validation of the DCGAN-LSA results led to the following conclusions: (1) The DCGAN-CNN shows a decrease of 4.13% in the prediction success rate. However, there is a 1.9% increase in its ability to predict landslide areas as very high-risk areas. (2) The DCGAN-ResNet exhibits a significant improvement compared to the ResNet, with a prediction success rate of 91.31%, representing an increase of 10.93%. Additionally, the DCGAN-ResNet shows greater sensitivity in landslide recognition, successfully predicting 66.52% of the areas with occurred landslides as very high-risk areas, making it the best performer among the four models.

5. Discussion

In landslide susceptibility modeling, eliminating irrelevant factors helps reduce model complexity and improve training efficiency. After factor selection, it was found that factors such as slope and the STI have high multicollinearity, while factors like aspect, curvature, the SPI, and the TWI have low contributions for landslide events. Therefore, both types of factors should be excluded in the subsequent modeling process. Historical landslides occur more frequently in low-altitude areas with high precipitation and frequent human activities. The prediction results of the CNN and ResNet are also similar to the distribution of historical landslides, with landslide risk significantly higher in urban areas with dense populations at low altitudes compared to other areas.

The evaluation of original LSA models and DCGAN-LSA models indicates that the DCGAN can effectively enhance the performance of LSA models. When trained with the augmented dataset, the SVM, CNN, and ResNet all exhibit improved accuracy. Furthermore, deep neural networks are sensitive to the quantity of additional samples, with a shortage leading to overfitting and an excess resulting in underfitting. Therefore, before incorporating additional samples, it is essential to consider the number of samples as a model parameter and conduct repeated testing to determine the optimal outcome.

In the validation process of LSA mapping results, traditional validation methods typically involve cataloging a portion of landslide points separately when constructing the dataset. These points are excluded from model training and testing, serving solely for the final validation. However, this paper argues that the primary application of LSA is to predict and mitigate future landslide risks in the region. Therefore, in the final validation phase, accuracy is evaluated by conducting a zonal area analysis using recent landslide surface data from the past few years. This approach aims to accurately assess the practical performance of the LSA model in real-world applications.

This study demonstrates the viability of supplementing training samples with a DCGAN to enhance the performance of LSA machine learning models, addressing the issue of insufficient samples. However, there are still limitations in this research. (1) Obtaining landslide data for certain areas may be challenging, potentially leading to omissions in the analysis of landslide factors. (2) While generating synthetic samples with the DCGAN, some unrealistic samples were observed. Although manual screening was performed to remove them during the experiments, there is still potential for improvement in the application of DCGANs for generating landslide samples. The focus should primarily be on improving the trainability and stability of GAN models when learning multi-factor landslide data, with the goal of achieving high-quality and efficient landslide sample generation to better address the issue of insufficient data in LSA.

6. Conclusions

In the traditional LSA model construction process, when the data are insufficient, it is common to either manually select input data to make the features more distinct or use improved algorithm models to accommodate small sample data in order to enhance the model’s prediction accuracy [47,48]. This study introduces a DCGAN-based data augmentation method to enhance the efficacy of three machine learning methods for LSA in western Sichuan. Experimental results indicate that (1) the SVM, CNN, and ResNet demonstrate the ability to extract the intrinsic connections among various landslide factors, achieving satisfactory classification performance. (2) The DCGAN-based data augmentation method is effective in enhancing the performance of various machine learning models for LSA. It not only increases sample diversity but also helps the model better capture the potential distribution of landslides, thus improving the generalization capability of the LSA models. Notably, the CNN and ResNet are sensitive to the quantity of additional samples. (3) Training with the augmented dataset, the SVM and CNN show relatively modest improvements. The success rate of CNN predictions is decreased during validation. However, the ResNet exhibits significant enhancement.