A Strategy for Neighboring Pixel Collaboration in Landslide Susceptibility Prediction

Abstract

Landslide susceptibility prediction usually involves the comprehensive analysis of terrain and other factors that may be distributed with spatial patterns. Without considering the spatial correlation and mutual influence between pixels, conventional prediction methods often focus only on information from individual pixels. To address this issue, the present study proposes a new strategy for neighboring pixel collaboration based on the Unified Perceptual Parsing Network (UPerNet), the Vision Transformer (ViT), and Vision Graph Neural Networks (ViG). This strategy efficiently utilizes the strengths of deep learning in feature extraction, sequence modeling, and graph data processing. By considering the information from neighboring pixels, this strategy can more accurately identify susceptible areas and reduce misidentification and omissions. The experimental results suggest that the proposed strategy can predict landslide susceptibility zoning more accurately. These predictions can identify flat areas such as rivers and distinguish between areas with high and very high landslide susceptibility. Such refined zoning outcomes are significant for landslide prevention and mitigation and can help decision-makers formulate targeted response measures.

1. Introduction

Human activities severely challenge nature’s tolerance toward increasing anthropogenic changes during societal development. Decades of rapid socioeconomic development in China took place with escalating conflicts between resources, energy, and the environment, resulting in geological hazards that harm society [1]. The movement of earth materials leads to the complex formation of geological hazards, and both human activities and natural variations can lead to changes in the geological environment [2]. As such changes accumulate, consequences resulting from geological disasters can occur [3]. Since the beginning of the new century, landslides and other geological disasters have occurred frequently due to expanding human activities and global climate change, leading to surging ecological and environmental issues, including soil erosion and reduction in vegetation [4,5]. Landslides involve various ground movements such as rock falls, deep slope failure, and shallow slope instability [6]. Gravity is the primary cause of landslides on excessively steep slopes. Sliding sand, rocks, and soil can block rivers and cause further disasters, posing direct threats to people’s lives and property [7]. Landslides are characterized by multiple occurrences, rapid development, and significant impacts. The development of landslides is also a complex process [8]. However, the probability of landslides occurring in an area can be predicted based on hazardous environmental conditions influenced by multiple factors. Predictions can help us better understand potential locations of instability that could lead to landslides and thus control their development in high-risk areas [9,10]. In this way, landslide prevention and mitigation can achieve better efficiency and help reduce the risk of hazards.

Landslide susceptibility analysis considers the geological environment, topography, meteorological conditions, etc., and evaluates landslides’ susceptibility to and danger in a given area [11]. Predicting potential landslide areas correlates with major hazard factors and historical landslide hazards and involves determining the functional relationships between the variable of hazard factors and the dependent variable of landslide occurrence [12,13]. Landslides are complex natural phenomena whose formation process can be influenced by many hazard factors, including geology, topography, hydrology, meteorology, vegetation, soil properties, and anthropogenic activities. In addition, the interactions between multiple hazard factors complicate landslide susceptibility predictions [14,15], not only due to the large number of input variables but also because of the complex and time-consuming analysis methods involved [16]. Challenges in landslide susceptibility research methods emerge from the requirement to thoroughly understand the interrelationships and mechanisms between hazard factors to accurately predict landslides [17]. In addition, since various factors influence landslide susceptibility, it is necessary to incorporate different data sources and technologies such as remote sensing, geographic information systems (GIS), and geological surveys that can obtain and integrate data [18,19]. In data processing and analysis, the latest statistical models and machine learning methods are needed to address complex non-linear relations and pattern recognition problems among factors [20].

Since the 1990s, the advancement of algorithms, remote sensing, and global positioning systems have made 3S technology (remote sensing, global position systems, and geographical information systems) an important tool for landslide susceptibility prediction. Beyond replacing conventional methods, this technology also provides more effective models for landslide susceptibility prediction, thus filling the gaps in previous studies [21,22,23]. However, from the perspective of geographic information science, problems still exist in current research on landslide susceptibility prediction. Due to the particularities of landslide hazards, determining their locations may require expert knowledge or on-site investigations. Moreover, the samples in landslide interpretation databases are limited in size [24], greatly restricting data-driven learning models. Previous studies on susceptibility prediction did not fully utilize the edge characteristics and spatial shape features of landslides. Moreover, the complex non-linear relationships between landslides and influencing factors make simple network structures unable to fully utilize feature information; thus, there remains space for improvements in relevant research [25,26].

Commonly used landslide susceptibility prediction methods in the past have utilized statistical methods (e.g., the certainty factors method and informativeness method) to generate landslide samples, which were then combined with machine learning or deep learning models to perform the assessment. These methods are mainly based on counting the number of landslides (area) within each classification by classifying the impact factors. The samples generated in this way are one-dimensional textual datasets, which lose the boundary information of landslides to a certain extent and cannot effectively incorporate the corresponding spatial information. In this study, a moving slider (size: 64 × 64) was used to traverse the landslide label and influence factor layers across the study area to produce a sample set, which was a 2D image dataset at this stage. Iterative pixel-by-pixel learning of 2D matrices using deep learning methods can maximize the use of edge characteristics and spatial shape features of landslides. This method can help further refine the final landslide susceptibility zoning results. The specific steps of this study are as follows: 1. We combine a topographic map, geological map, and remote sensing image of the area to interpret landslides in Wenchuan and build an interpretation database. 2. By traversing landslide vectors and evaluation factors at a fixed scale, a stable, diverse, and high-quality sample training set is constructed. 3. We achieve susceptibility predictions in Wenchuan by considering the spatial characteristics of landslides as area-based geographic objects (including spatial relation, spatial scale, and spatial shape features) and using deep learning models with a focus on spatial features (UPerNet, ViG, ViT). These predictions enable us to provide technical and theoretical support for the early identification, warning, and accurate prediction of landslides, thus strengthening environmental protection.

2. Materials

2.1. Study Area

The study area of Wenchuan is located in the northwestern Sichuan Basin, with coordinates of 30°45′N–31°43′N and 102°51′E–103°44′E, featuring a total area of 4084 km² characterized by highly developed faults and folds (Figure 1). The two major fracture zones of Maoxian–Wenchuan and Beichuan–Yingxiu extend across Wenchuan County from northeast to southwest, in which the Beichuan–Yingxiu fracture was associated with the “5.12” earthquake [27]. Before the 2008 earthquake, the Wenchuan region was already highly susceptible to landslides, with landslide hazards particularly prominent at riverbanks.

2.2. Experiment Data

Landslide susceptibility prediction based on the semantic segmentation of deep learning requires a landslide dataset for training. Since no landslide interpretation and influencing factor dataset for Wenchuan existed prior to this study, it was necessary to collect and build a dataset for the county.

2.2.1. Source of Data

Data were placed into two classes according to the needs of this study (

Table 1. Data source of influencing factors on landslide susceptibility prediction.

Name	Type	Spatial Resolution	Use	Source
Gaofen-1 satellite image	Raster	Panchromatic 2 m	Landslide visual interpretation	Gaofen Sichuan center for data and application
Sichuan geological hazard survey data	Point	-	Reference for landslide interpretation	Sichuan Provincial Department of Natural Resources
Fundamental geographic information	Line, polygon	1:50,000	Extracting factors of roadway, river, etc.	National Geomatics Center of China
Digital elevation model (DEM)	Raster	30 m	Extracting factors of slope, aspect, etc.	Geospatial Data Cloud
Geological data	Line, polygon	1:200,000	Extracting tectonic line factor	Sichuan Geological Survey Research Institute
Rainfall data from weather stations	Point	-	Extracting average annual rainfall factor	China Meteorological Administration

): (1) remote sensing image and regional investigation data for the sample set on susceptibility prediction and (2) original data for the feature dataset of influencing factors on landslide susceptibility.

2.2.2. Production of Landslide Interpretation Database

The goal of the present study on landslide hazards was to develop a detailed, objective, and accurate landslide interpretation database, including the patterns and characteristics of spatial distributions. As reported in this section, we performed a remote sensing interpretation of landslides in Wenchuan for the last 10 years, combining detailed landslide investigation point sources and historical high-resolution remote sensing images with field investigations and on-site validation. The interpretation relied primarily on landslide exposure, vegetation destruction, and accumulation recognition, ultimately identifying 5944 landslides, as shown in Figure 2.

2.2.3. Factors That Influence Landslide Susceptibility

Topography, landforms, and other factors influence the occurrence of landslides. Moreover, a scientific and rational understanding of the factors influencing landslide occurrence in the geological environment is a precondition for conducting landslide susceptibility prediction [28]. Different combinations of influencing factors have different effects on the prediction results of landslide susceptibility, so it is necessary to analyze and compare the importance of factors. Thus, this study extensively investigated the geological environment and conditions that cause landslides in Wenchuan. Considering the actual situation in the study area, the 12 factors were chosen to establish landslide susceptibility prediction metrics: elevation, slope, aspect, topographic relief, plan curvature, profile curvature, tectonic lines, lithology, rivers, rainfall, land use, and roadways [29,30,31].

Deep learning models are often regarded as black-box models, meaning their internal decision-making process is opaque to us. Due to the complexity and non-linear nature of deep learning models, it is difficult to parse out the specific contribution of individual feature factors to the final prediction result directly from the models. In deep learning models, the importance of features is not directly visible because they are embedded in the complex computational processes of multi-layer neural networks. Traditional machine learning models tend to provide direct information about the importance of features. In this study, the importance of the evaluation factors was calculated based on the decision tree, and the calculated factor weights were normalized before proceeding with the model construction. The order of factor importance is shown in Figure 3. Therefore, the two factors of lithology and land use were not selected for modeling in this study.

Elevation (Figure 4a) does not have a direct relationship with landslides; instead, the impact of elevation on landslides is usually reflected by other associated conditions. In addition, the landform and topography vary in different elevation ranges, greatly affecting the form of landslide faces [32].

Slope (Figure 4b) or slope angle refers to the angle between the slope and horizontal surfaces. Slope angle is one of the major factors influencing the stability of slopes and is directly related to the shear strength of sliding surfaces. When the slope exceeds the maximum stability angle of the gravitational stress on the soil, the rock and soil bodies will slide downwards [33]. The slope aspect reflects the direction corresponding to different slope bodies, and the difference between shaded and sunny slopes can play a different role in the development and distribution of landslides. As Wenchuan County is in the northern hemisphere, the convention is that the sunny slopes face south, and conversely, the shaded slopes face north. Sunny slopes have longer hours of sunlight than shady slopes, resulting in more solar radiation, more water evaporation from sunny slopes, and more weathering of slopes.

The slope aspect (Figure 4c) is the direction that the largest slope surface on the landform faces and is widely used in landslide susceptibility modeling due to its direct correlation with slope stability. This aspect is relevant to slope stability since it determines the slope’s exposure time to the sun and the direction of rainfall. These elements lead to different moisture absorption and evaporation results on different slopes and change the weathering intensity on rock and soil bodies [34]. The slope aspect reflects the direction corresponding to different slope bodies, wherein shaded and sunny slopes can play different roles in the development and distribution of landslides. Wenchuan is located in the northern hemisphere; thus, sunny slopes face south, while shaded slopes face north. Sunny slopes experience longer hours of sunlight than shady slopes, resulting in greater solar radiation, more water evaporation from sunny slopes, and increased slope weathering.

Topographic relief (Figure 4d) in slope elements usually refers to the difference between maximum and minimum elevation, which impacts the possibility of landslide occurrence on a slope [35]. Less than 1% of the study area features a topographic relief greater than 349 m. Landslides are mainly concentrated between 29 and 394 m, with the proportion of landslides increasing as the topographic relief increases.

Curvature is the degree of bending or the rate of change to neighboring points at a certain point and includes components along both the horizontal and vertical directions [35]. Areas with curvature greater than 0 are convex slopes, while areas with curvature less than 0 are concave slopes. The component on the horizontal axis is extracted to determine plan curvature (Figure 4e,f). The density of landslide distribution rises overall with an increase in plan curvature. The curvature component on the vertical axis is known as profile curvature, which is the changing rate of the surface slope at that point. Similar to landslide distribution in ranges of plan curvature, the percentage of landslide distribution grows overall with an increase in profile curvature. Curvature quantitatively describes the degree of bending and deformation at a point on a terrain surface. Curvature breaks down the degree of distorted deformation into vertical and horizontal directions. The component of ground curvature in the vertical direction is called the profile, which measures the change in elevation in terms of the rate of change in the direction of the maximum drop in slope. The component of curvature in the horizontal direction is called plane curvature. Convex slopes are more prone to landslides because the bending deformation of convex slopes causes stress concentration within that slope, leading to slope instability. The proportion of landslide area in the study area is higher within the area of convex slopes than in the area of landslide hazards within concave slopes.

The rupture and fracture caused by tectonic movements result in overhanging surfaces and valleys, which provide conditions for landslide development and occurrence. In areas affected by tectonic and seismic movements, the spatial distribution of landslides is closely related to the tectonic pattern [36].

Rivers (Figure 4h) also impact landslides, mostly in terms of changes in the slope stress state and landslide mechanics. Dynamics related to river erosion and corrosion weaken the slope’s stability and catalyze the occurrence of landslides [37]. The bank slopes of rivers, lakes (reservoirs), seas, and ditches, as well as canyon areas with large differences in topographic elevation, are prone to landslides due to the scouring and erosion of water currents, which make the bank slopes steeper and higher. In terms of spatial distribution, a large number of landslides in this study are located relatively close to the stream and river system.

Rainfall (Figure 4i) can increase excessive water saturation and the overall weight of the slope. The lateral pressure along the slope surface due to gravity and the weakened shear strength of rock and soil bodies results in a sliding surface that can lead to landslide hazards [38].

Lastly, the distance to roads (Figure 4j) can reflect the regional intensity of human engineering activities. Potential side-slope excavations and the earth’s surface vibrations caused by vehicle transportation on roadways are additional potential factors influencing landslides. Existing studies suggest a centered landslide distribution along roadways and an inverse relationship between landslide density and distance to roads [39].

2.3. Production of Datasets on Susceptibility Prediction

Samples provide crucial data to support deep learning. Indeed, the quality of samples is critical to the practicality and precision of such models. Thus, the quality of samples should be strictly controlled. As reported in Section 2.2.2., a detailed landslide interpretation of Wenchuan was conducted and validated in the field. Landslide sample data were constructed from landslide attribute data (landslide or non-landslide data) and influencing factor attributes that affect landslide occurrence. The ultimate goal of the landslide susceptibility prediction model is to predict landslide attributes, with influencing factor attributes used as the input parameters. Since many factors in this study were derived from DEM (digital elevation model) data, all influencing factors used were uniformly sampled to a spatial resolution of 30 × 30 m [40,41], consistent with the DEM. This study is based on the interpretation results, which were transformed into raster data with the same spatial resolution as the influencing factor (a single pixel size of 30 × 30 m). Here, landslide data are labeled 1, and non-landslide data are labeled 0. To effectively utilize landslide margin and spatial shape features, a mobile block with 64 × 64 pixels [42,43] was introduced to build the prediction dataset; as can be seen from the cropping results of the block in Figure 5, the boundary information of the landslide can be well preserved.. In this way, we obtained a training set of 924 samples and a test set of 396 samples. In addition, 25% of the data in the training set was partitioned as a validation set to select training epochs and avoid overfitting.

Online data augmentation, also known as in-place data augmentation, primarily includes the following stages [44]: (1) starting model training; (2) extracting a small batch of data from the training set according to assigned or random rules (batch size); (3) sending original small-batch data to the data magnifier and using the magnifier to perform geometrical transformation according to the designated rules; (4) using the data augmentation module input to augment small batch data input into the training model and performing training; (5) repeating stages 2 to 4 for the required training times. The process of online data augmentation is illustrated in Figure 6. During each augmentation step, the new data input into the model are very similar to the original data. To gradually improve the small-batch data, we augmented data based on geometric and image color transformation through the online data augmentation module.

The geometric transformation of data usually involves flip, rotation, cropping, distortion, scaling, and other operations. Extensive experiments from previous studies suggest that flipping and distorting the image are valid data augmentation techniques for semantic segmentation [45]. Flipping is an operation similar to mirroring, while rotation involves a circular clockwise or counterclockwise movement. Moreover, it is important to rotate the image by 90 to 180 degrees to avoid inconsistent scaling among the image samples. Here, the flipped and rotated augmented images differed from the original images only in their directions and angles, as they retained the same size [46]. We selected random rotation and flipping for geometric data augmentation and expanded the sample dataset to three times its original size. A final training set of 2772 samples and a test set of 1188 samples were thus obtained. Labels were transformed according to the geometric transformation of the images to ensure consistency with the transformed images.

3. Methods

3.1. Description of the Applied Models

3.1.1. Unified Perceptual Parsing Network (UPerNet)

UPerNet [47] is based on the feature pyramid network (FPN) and uses a bottom-up process to sequentially extract image features, including small-scale detailed feature information and large-scale generalized feature information. The backbone network used in this study is the classical ResNet50. UPerNet connects high-level features that characterize the generalized information and low-level features that characterize the detailed information with top–down sidelobes through top–down paths, enabling features at all scales to contain rich information. The network architecture is shown in Figure 7. Since this susceptibility prediction study uses 10 impact factor layers, the input features correspond to 10 channels.

3.1.2. Vision Transformer (ViT)

The classical Transformer model [48] uses a one-dimensional embedded token as input. Dosovitski et al. [49] proposed using a Vision Transformer (ViT) to process 2D images. Considering time complexity and performance, ViT is able to directly input feature images into the Transformer network to perform image classification tasks without relying on the CNN network, which demonstrates for the first time the possibility of using the Transformer in the field of computer vision. In the natural language domain, the transformer accepts word tokens as sequence input. Thus, ViT cuts the input image into fixed blocks of P × P size without overlapping and then converts the segmented feature subgraphs into one-dimensional sequences, i.e., patch embedding. The category and location information are then fed into the Transformer encoder as a learnable vector along with the sequence data using a method similar to BERT [50], which performs a self-attention operation. The final category vector output is used for downstream classification tasks. The ViT network structure is shown in Figure 8.

3.1.3. Vision Graph Neural Networks (ViG)

The GNN is a neural network specialized in processing graph data. The GNN is most useful when handling complex graph data, including social networks, knowledge graphs, and molecular geometry. The core idea of the GNN is to update the output of nodes through iteration using the graph’s structural information [51]. In each iteration step, every node agglomerates information on its neighboring pixels and updates its state accordingly. Through this process, the nodes can capture global structural information from the graph. In reality, many important datasets exist as graphs or networks. Thus, combining a neural network with graph analysis could resolve numerous problems [52]. A graph, also known as a topology, is an important concept in computer science. The two primary components of a graph are its vertices and edges, represented as follows:

$$G=(V,E),$$

(1)

where G is the graph, V is the set of vertices, and E is the set of edges connecting vertices.

Vision GNN (ViG) is an innovative neural network architecture [53] that combines the advantages of computer vision with GNN. The core idea of this approach is to convert traditional image data into graph structures and utilize GNNs to capture the complex patterns and relationships in images. In ViG, an image is first segmented into multiple local regions (or “nodes”). These nodes are connected as a graph by considering the spatial relationship (e.g., adjacency) between them. Through this transformation, pixels in an image are no longer isolated but instead organized into a graph structure that expresses the overall form and local details of the image. ViG next utilizes the graph convolution operation in GNNs to aggregate and update the feature information of the nodes. Each node can capture information about its neighboring nodes through multiple iterations and updates, resulting in a rich representation that contains global and local information about the image. The network structure is illustrated in Figure 9.

3.2. Evaluation of Model Accuracy

To objectively validate the prediction results, we selected four validation factors, including pixel accuracy, overall accuracy, recall rate, and F1 score, based on the test images [54,55,56]. Then, the performance of different methods on landslide susceptibility prediction was assessed to choose the optimal prediction model. Here, pixel accuracy (PA) reflects the proportion of accurate predictions in all pixels predicted as landslides:

$$PA=\frac{TP}{TP+FP}.$$

(2)

Overall accuracy (OA) is the proportion of correct landslide and non-landslide predictions in all pixels predicted, as follows:

$$OA=\frac{TP+TN}{P+N}.$$

(3)

The recall rate (Recall) is the proportion of correctly predicted landslide pixels in all sample landslide pixels:

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

(4)

The F1-score considers both pixel accuracy and the recall:

$$F1\text{-}\mathit{score}=\frac{2\times Recall\times PA}{Recall+PA},$$

(5)

where $P$ is the number of landslide pixels; $N$ is the number of non-landslide pixels; $TP$ refers to landslide pixels predicted as landslide pixels; $TN$ refers to non-landslide pixels predicted as landslide pixels; FP refers to non-landslide pixels predicted as landslide pixels; and $FN$ refers to landslide pixels predicted as non-landslide pixels.

4. Results

4.1. Application of the UPerNet, ViT, and ViG Models

Landslide susceptibility predictions using deep learning models involve processing large amounts of sample data alongside high algorithmic and time complexity, which demand significant computational resources. Consequently, the present experiment used a 13th Gen Intel^® Core™ i7-13700KF CPU, an NVIDIA GeForce RTX 4090 GPU, and 64.0 GB of RAM. The software environment employed the deep learning framework pytorch-gpu-1.8.1 + CUDA 11.3 + CUDNN 8.2.1 for model construction, corresponding to Python version 3.7.

The augmented training set described in Section 3.2 was input into the UPerNet, ViT, and ViG models, with the best results obtained by optimizing the model parameters. The parameter settings of the different models are listed in

Table 2. Model Parameter Settings.

Models	Same parameters	Backbone
UPerNet	Epochs:100 Dropout: 0.5 Learning rate: 0.001 Batch_size: 16 Activation function: ReLU Optimizer: Adam Loss function: Binary CrossEntropy	7 × 7, 64, Conv 3 × 3, max pool ResNet50 _Block [2,2,2,2]
ViT		4 × 4 Patch 64 Embedding Transformer_Layer [2,2,6,2]
ViG		7 × 7, 64, Conv 3 × 3, max pool GNN_Layer [2,2,6,2]

, which also shows the optimal parameter weights.

All prediction samples were considered monochrome images and input into the established models. Landslide susceptibility index maps were generated by calculating the possibility of landslide occurrence in each pixel (Figure 10a, Figure 11a, and Figure 12a). Using the equal distance method, calculated indices were categorized into five levels: very low landslide-susceptibility areas (0–0.2), low landslide-susceptibility areas (0.2–0.4), medium landslide-susceptibility areas (0.4–0.6), high landslide-susceptibility areas (0.6–0.8), and very high landslide-susceptibility areas (0.8–1). The landslide susceptibility zoning maps are illustrated in Figure 10b, Figure 11b, and Figure 12b.

4.2. Validation of the UPerNet, ViT, and ViG Models

The network training performance on the test set is presented in

Table 3. Comparison of landslide susceptibility test sets from different models.

Models	PA	OA	Recall	F1-Score
UPerNet	0.78	0.98	0.77	0.77
ViT	0.66	0.97	0.72	0.69
ViG	0.72	0.97	0.74	0.73

. Compared with UPerNet and ViG, the overall test results for ViT lacked competitiveness and offered lower evaluation results for all criteria. Although the transformer (TR) model has excellent performance in the natural language processing (NLP) field, it is less competitive in handling image or graph data. The ViT extracts feature representations with context relevance, which is effective for text data. However, for images or graphs, ViT may not fully capture the data’s spatial–structural information or geometric features.

UPerNet performs better than ViG on the test set for all criteria, with 6% higher pixel accuracy, 1% higher overall accuracy, 3% higher recall, and a 4% higher F1 score than ViG. UPerNet provides increased performance because it specializes in processing image data. By performing convolution on input images with the convolutional layer, UPerNet can effectively extract regional features from images and generate layered feature representations by stacking multiple convolutional layers. In comparison, ViG is primarily used to process graph data, such as in social networks and knowledge graphs. Although ViG can accurately process graph data, it may not effectively extract and utilize regional features in image data or UPerNet.

5. Discussion

5.1. Comparing Datasets with Different Data Volumes

Data augmentation is a key technology in the field of deep learning. Through a series of transformations and expansions of the original dataset, this method can effectively enlarge the training sample size and improve the model’s generalization ability. In the training process of deep learning models, data augmentation can reduce the risk of overfitting and help the model learn more useful features to enhance the accuracy and stability of the prediction results. In this study, we expanded samples of the dataset to three times their original size. The data augmentation results show that UPerNet, ViG, and ViT all obviously improved in accuracy. This improvement benefited from data augmentation, which provided more diverse and abundant training samples and enabled the model to learn more useful features and patterns. Nevertheless, UPerNet presented the greatest improvements in accuracy.

Compared to models before data augmentation, UPerNet improved the PA, OA, Recall, and F1 score by 26%, 2%, 8%, and 18%, respectively; ViT improved them by 17%, 1%, 2%, and 11%; and ViG improved them by 18%, 1%, 5%, and 12%. Comparing the susceptibility prediction results of three models before and after data augmentation, ViG was found to provide more accurate susceptibility zoning than the other two models because ViG updates the representations of nodes by agglomerating information from the neighboring pixels, which enables the model to learn the dependency between nodes. For small samples, even with a limited sample size, ViG can still diversify representations of nodes through agglomeration.

After data augmentation, the susceptibility zoning predictions of UPerNet and ViG improved greatly.

Table 4. Comparison of evaluation metrics before and after data enhancement for different models.

Models	Data Volumes	PA	OA	Recall	F1-Score
UPerNet	original	0.52	0.96	0.69	0.59
	3× data enhancement	0.78	0.98	0.77	0.77
ViT	original	0.49	0.96	0.70	0.58
	3× data enhancement	0.66	0.97	0.72	0.69
ViG	original	0.54	0.96	0.69	0.61
	3× data enhancement	0.72	0.97	0.74	0.73

and Figure 13 suggest that UPerNet offered slightly better improvements than ViG because of its strength in processing image data. UPerNet’s convolution operations can effectively extract regional features from images and generate layered feature representations by stacking multiple convolutional layers. Data augmentation provides UPerNet with more image samples. The transformation and expansion of these samples additionally expose the model to different image patterns, further enhancing the feature extraction and prediction accuracy of UPerNet. ViG is excellent in processing graph data and can capture complex relationships and dependency patterns between nodes.

However, ViG might not be as effective as UPerNet when processing image data because the regional features and spatial structure of image data can be better extracted through convolution. ViG sometimes requires more complex transformations and representations when handling this type of data. Although data augmentation provides ViG with more graph data, ViG cannot be compared with UPerNet in terms of accuracy improvements because of its different processing mechanisms and data characteristics. Moreover, ViG might face challenges in computation efficiency and extensibility when processing large-sized graph data, which can also restrict its improvements in accuracy. While ViT has had success in the field of NLP, it may have limitations in processing image data.

Moreover, although data augmentation may help improve the accuracy of the ViT model, ViT may not be as efficient as UPerNet in extracting image features and processing spatial structures since it is not designed for image data. Furthermore, ViT has relatively high computation complexity and may require more computational resources and time when processing large image data. This factor may also affect the accuracy improvements of ViT.

5.2. Comparison of Models for Susceptibility Zones

In this study, areas within Wenchuan County were divided into five zones: relatively low landslide susceptibility, low landslide susceptibility, medium landslide susceptibility, high landslide susceptibility, and very high landslide susceptibility. The classified statistics for the susceptibility zoning results for landslide density are shown in

Table 5. Statistical zoning results of the different models.

Models	Zoning Levels	Percentage of Landslide Area Li (%)	Percentage of Susceptibility Zones Si (%)	Ri = Li/Si
UPerNet	Ⅰ	15.1	94.8	0.16
	Ⅱ	2.6	0.9	2.89
	Ⅲ	5.8	0.8	7.25
	Ⅳ	14.1	1.1	12.82
	Ⅴ	62.4	2.4	26
ViT	Ⅰ	7.8	90.3	0.09
	Ⅱ	7.4	3.6	2.06
	Ⅲ	10.9	2	5.45
	Ⅳ	29.7	2.3	12.91
	Ⅴ	44.2	1.8	24.56
ViG	Ⅰ	8.1	92.4	0.08
	Ⅱ	6.2	2.3	2.69
	Ⅲ	10.8	1.6	6.75
	Ⅳ	19.2	1.5	12.8
	Ⅴ	55.7	2.2	25.32

.

The evaluation criteria for this method were as follows: The very high landslide-susceptibility zone represented the smallest proportion in the entire study area with the maximum proportion (number) of landslide test set samples, while the low landslide-susceptibility zone represented the greatest proportion in the area with the minimum proportion (number) of landslide test-set samples. The proportion of each landslide susceptibility zone in Table 5 shows that the deep learning prediction results under the new strategy all satisfy the needs of susceptibility zoning. In other words, the low landslide-susceptibility zone occupies a higher proportion of the area with a lower rate of landslides, while the very high landslide-susceptibility zone occupies a smaller proportion with a higher rate of landslides. The zoning results of the three models all satisfy reasonable susceptibility requirements: $R_I<R_{II}<R_{III}<R_{IV}<R_V.$ The proportion of the very low landslide-susceptibility zone in all models was $S_{UPerNet}<S_{ViG}<S_{ViT}$; the zonal proportion of landslides was $L_{UPerNet}>L_{ViG}>L_{ViT}$; the proportion of the low landslide-susceptibility zone was $S_{UPerNet}>S_{ViG}>S_{ViT}$; and the zonal proportion of landslides was $L_{ViT}<L_{ViG}<L_{UPerNet}$.

With neighboring pixel collaboration, the UPerNet model offers the best results in zonal statistics. Under this model, the high landslide-susceptibility zone occupied only 2.4% of the entire area with 62.4% of landslide hazards, whereas the low landslide-susceptibility zone represented 94.8% of the entire area region, covering most of Wenchuan County, with only 15.1% of landslide hazards. The zoning and hazard prediction results both demonstrate that the UPerNet model can identify very high landslide-susceptibility zones.

The contributions of neighboring pixel collaboration to landslide susceptibility prediction are twofold. First, this method assists deep learning models in capturing the spatial patterns of landslide occurrence. Landslides often happen under certain geological environmental conditions that are distributed according to certain spatial laws. By involving information from neighboring pixels, the model can better recognize such patterns and evaluate the possibilities for landslide occurrence. Figure 14 suggests that flat areas such as rivers usually have low relevance in landslide occurrence, while steep hills represent more susceptible zones. The UPerNet, ViT, and DNN models have strong feature extraction and learning capabilities. These models can fully utilize the advantages of neighboring pixel collaboration and achieve excellent results in landslide susceptibility zoning by better distinguishing between flat and susceptible regions.

5.3. Comparison with the Traditional Method

Statistical techniques such as the informational method and the deterministic factor method [57,58,59] have been widely adopted in traditional studies of landslide susceptibility prediction to construct feature sample sets. These methods focus on calculating the specific quantitative relationships between landslides and various types of influencing factors to accurately assess the sensitivity of influencing factors to landslides. After combining sensitivity analysis with the spatial distribution of landslides, predictions of landslide susceptibility were derived via hierarchical superposition. This type of sample set is a two-dimensional text dataset that tends to ignore the precise boundary information of landslides. Even using the Transformer model [60], which is very advanced in the textual domain, could not achieve granular predictions of susceptibility.

To solve this problem, we proposed a novel strategy using neighboring pixel collaboration. Employing iterative pixel-by-pixel learning of 2D images with deep learning methods can utilize the maximum number of edge features and spatial shape features of landslides. A 2D image dataset was produced based on this strategy, which can retain the boundary information of landslides as completely as possible when predicting landslide susceptibility. We also used the Transformer model, which performs well in textual data modeling, to compute landslide susceptibility in Wenchuan County using the traditional strategy. Taking Miansi town as an example, we explored the landslide susceptibility prediction results using different strategies.

As shown in Figure 15, our new strategy combines different deep-learning models to label areas with flat terrain and a very low probability of landslide occurrence as very low susceptibility zones. At the same time, very high susceptibility zones accounted for a very low percentage of the study area. This outcome reflects the ultimate goal of landslide susceptibility prediction: to focus on areas that are extremely susceptible to landslides at the lowest possible monitoring cost. In contrast, the traditional strategy employed a continuous zoning approach to analyze the influencing factors, thus closely linking the river neighborhoods to the landslide susceptibility areas of the lower mountains and achieving the relevant data mining objectives. However, our new method offered significant advantages in preserving landslide boundary information, providing new perspectives and strategies for landslide susceptibility prediction.

6. Conclusions

This study proposed a new strategy using UPerNet, ViT, and ViG for neighboring pixel collaboration to effectively utilize landslide features such as margins and spatial shapes in landslide susceptibility prediction. This strategy efficiently combines the strengths of deep learning in image processing, sequence modeling, and data processing. We selected Wenchuan County as the study area, conducted various experiments, and obtained the following conclusions.

We designed an online data augmentation module that successfully expanded the sample size of the dataset to three times its original size. This expansion greatly improved the accuracy of the prediction models. Compared to the models before data augmentation, UperNet improved the PA, OA, Recall, and F1-score, respectively, by 26%, 2%, 8%, and 18%; ViT improved them by 17%, 1%, 2%, and 11%; and ViG improved them by 18%, 1%, 5%, and 12%.

The UPerNet model offered the best performance in susceptibility zoning prediction. The very high landslide-susceptibility zone predicted by UPerNet contained 62.4% of landslides, which was considerably better than the results from ViG (55.7%) and ViT (29.7%) because the regional features and spatial structures of the image data could be better extracted via convolution.

By using the latest machine learning and statistical models and incorporating neighboring pixel collaborative analysis, the proposed model could more accurately predict landslide susceptibility. These predictions identified flat areas such as rivers and distinguished between areas with high and very high landslide susceptibility.