Identification of Landslides in the Hilly Areas of Eastern China Using High-Resolution GF-2 Images and Deep Learning Models

Abstract

Small, dispersed, and vegetated creeping landslides in hilly areas of eastern China hinder traditional remote sensing and detection. To address this, this study takes Yixian County (Anhui Province) as a representative area. Based on high-resolution GF-2 satellite images, three deep learning models embedded with the Squeeze-and-Excitation (SE) attention mechanism (ResNet18-SE, VGG13-SE, UNet-SE) were developed and compared with the original UNet model. Combined with field investigation, landslide mapping and accuracy assessment were conducted to evaluate the feature extraction capabilities of four models. The results indicate that the UNet-SE model achieved optimal performance (Precision: 0.911, Recall: 0.685, F1-score: 0.782, Kappa: 0.730, IoU: 0.643). Its F1-score exceeds ResNet18-SE, VGG13-SE, and the original UNet by 8%, 3%, and 5%, respectively, proving superior regional adaptability and generalization performance. Additional verification on creeping landslides in Kecun Town (Yixian County) and post-earthquake landslides in Lushan County (Sichuan Province) further confirms the reliability of the UNet-SE model. Furthermore, Frequency Ratio (FR), Random Forest (RF), and SHapley Additive exPlanations (SHAP) were adopted to reveal the correlation between landslide occurrence and seven geological-environmental factors. Landslides are most susceptible to develop at elevations of 400–500 m, NDVI values of 0.4–0.5, slopes below 10°, east and northeast aspects, 300–500 m away from rivers, 500–1000 m away from faults, and areas covered by soft sedimentary lithology. Distance from faults, distance from rivers, and elevation are identified as the three favorable conditional factors. In conclusion, the proposed landslide detection framework can provide reliable spatial data and technical references for regional geological hazard prevention, ecological conservation and sustainable development in hilly areas.

1. Introduction

Landslides are among the most destructive natural disasters globally, second only to earthquakes in terms of economic losses and casualties, and they pose a serious threat to human life and property [1]. They are typically triggered by earthquakes, intense rainfall, or inappropriate engineering activities, and their characteristics and impacts vary significantly depending on the triggering factors [2,3]. In contrast to the sudden, large-scale, and spatially concentrated landslides induced by earthquakes, creeping landslides in hilly regions develop gradually, are spatially dispersed, and are often concealed. These landslides typically appear as multiple localized, small-scale failures that are difficult to detect and, if left unaddressed, may result in severe consequences [4,5].

To mitigate the damage caused by landslides, accurate identification and real-time monitoring are essential for effective disaster prevention and risk reduction. Currently, both airborne and spaceborne remote sensing technologies have become powerful tools for landslide detection, owing to their wide area coverage and multi-temporal observation capabilities. Various types of remote sensing data have been successfully applied to landslide identification, including optical imagery [6,7], unmanned aerial vehicle (UAV) data [8,9], LiDAR scans [10,11,12], and synthetic aperture radar (SAR) data [13,14,15]. However, each data source has inherent limitations: optical imagery often struggles to detect subtle terrain changes; the high cost of UAVs and LiDAR restricts their use in high-frequency monitoring; and the relatively low spatial resolution of SAR makes it difficult to identify small-scale, shallow landslides [16]. These limitations highlight the challenges of relying on a single remote sensing source. To address these shortcomings, recent research has increasingly focused on the synergistic integration of multi-source remote sensing data. By combining complementary data types, researchers aim to enhance the accuracy and reliability of landslide detection and monitoring [17,18,19]. For example, Xun et al. integrated optical imagery with deformation maps derived from InSAR for an object-based analysis of potential landslides [20], while Cai et al. fused multi-source data from InSAR, LiDAR, and optical remote sensing, in addition to conducting field investigations, to achieve the accurate identification and dynamic monitoring of post-earthquake landslides [21].

With the advancement of artificial intelligence, deep learning models have demonstrated superior performance in image recognition tasks compared to traditional machine learning approaches such as Random Forest (RF) [22,23] and Support Vector Machines (SVMs) [24,25], owing to their deeper network architectures, stronger generalization capabilities, and greater robustness. These advantages have enabled deep learning models to more effectively extract and recognize landslide features [26,27,28]. For instance, Wei et al. proposed a feature-enhancement framework named AMU-Net, which integrates attention mechanisms and multi-scale features into the U-Net architecture to improve landslide boundary delineation [29]. Ghorbanzadeh et al. conducted a comparative analysis of SVMs, RF, artificial neural networks, and Convolutional Neural Networks (CNNs) in the context of landslide detection [30]. Wu et al. combined the strengths of CNNs and transformers to develop the SCDUNet model, which enables a more efficient extraction of spectral and topographic information [31]. The integration of deep learning techniques with multi-source remote sensing data offers a promising technical pathway for the accurate identification of creeping landslides in complex hilly terrains.

2. Data and Methods

2.1. Study Area

Anhui Province is located in eastern China and exhibits significant geomorphological variation between its northern and southern regions. The northern part features low-lying terrain and forms part of the Huaihe River Plain, while the southern region is characterized by elevated and hilly landscapes typical of eastern China. Influenced by the subtropical monsoon climate, the region experiences concentrated rainfall during the wet season, which frequently triggers landslide disasters [32,33,34]. The hilly terrain of southern Anhui, particularly Yixian County, represents a hotspot for landslides and other geological hazards within the province. Geomorphologically, the area is marked by rugged mountainous terrain, with slopes often exceeding 30°. Climatically, it falls within a humid subtropical monsoon zone, receiving abundant but unevenly distributed rainfall both intra- and inter-annually, with annual precipitation ranging from 1400 to 1900 mm and frequent heavy rainfall events during the rainy season [35,36]. These characteristics make the region highly susceptible to landslides. Moreover, the dense vegetation cover contributes to the concealment of landslides, and the complex geological environment presents considerable challenges for hazard detection and mitigation. Therefore, conducting research on the remote sensing-based identification of landslides in such areas holds substantial significance. In this study, Yixian County in southern Anhui Province is selected as the research area, and Hongtan Township was designated as the test area (Figure 1). Leveraging multispectral remote sensing imagery from the Gaofen-2 (GF-2) satellite, we apply machine learning and deep learning algorithms to identify and extract spatial information related to landslide occurrences. Additionally, we perform a correlation analysis between landslides and a set of geological and environmental conditional factors, aiming to reveal the critical conditions that contribute to the formation and occurrence of landslides in this region.

2.2. Data Sources and Technical Strategy

In this study, we integrated multiple datasets, including GF-2 satellite imagery, administrative boundary data, digital elevation data, and geological information. Detailed dataset descriptions are provided in

Table 1. Data presentation.

Data	Source	Resolution	Date
Panchromatic images	https://data.cresda.cn, accessed on 17 April 2026	0.8 m	12 March 2024
Multispectral images	https://data.cresda.cn, accessed on 17 April 2026	3.2 m	12 March 2024
Basin boundaries	https://www.tianditu.gov.cn, accessed on 17 April 2026	-	-
Digital elevation model	https://search.asf.alaska.edu/, accessed on 17 April 2026	12.5 m	2006–2011
Land use data	http://www.tianditu.gov.cn, accessed on 17 April 2026	30 m	2020
Geological data	https://www.ngcc.cn, accessed on 17 April 2026	-	-

.

2.2.1. Remote Sensing Images

GF-2 multispectral remote sensing imagery was employed in this study to identify the spatial distribution of landslides. The data were obtained from the China Centre for Resources Satellite Data and Application, with imagery acquired on 12 March 2024. The GF-2 satellite provides a panchromatic spatial resolution of 0.8 m and a multispectral resolution of 3.2 m. The multispectral data consist of four bands, namely, blue, green, red, and near-infrared bands, with wavelength ranges of 0.45–0.52 μm, 0.52–0.59 μm, 0.63–0.69 μm, and 0.77–0.89 μm, respectively [37]. To further validate the accuracy of the remote sensing-based extraction results, a field aerial survey of the landslide-affected area was conducted using a DJI Mini 2 UAV (manufactured by DJI, Shenzhen, China). The UAV is equipped with a 12-megapixel visible light sensor and can capture high-resolution images with up to 4000 × 3000 pixels. Additionally, the cross-regional validation experiment employed the Globally Distributed Coseismic Landslide Dataset (GDCLD; https://doi.org/10.5281/zenodo.13612636), which provides GF-6 true-color remote sensing imagery of earthquake-induced landslides in Lushan County, Sichuan Province, along with corresponding ground-truth labels [38]. Given that the model construction and training in this study were based on GF-2 satellite imagery, a high-quality cloud-free GF-2 image covering the same area (acquired on 11 January 2024) was selected as input data to ensure consistency in data sources.

2.2.2. Conditional Factor Data

In this study, we utilized a digital elevation model (DEM) with a spatial resolution of 12.5 m, derived from imagery provided by the Japanese Earth observation satellite ALOS and obtained through the Alaska Satellite Facility Vertex Data Portal. Based on this DEM, topographic conditional factors, including elevation, slope, aspect, and other surface morphology indicators, were extracted to characterize the geomorphological conditions of the study area. Land use data were obtained from the National Platform for Common Geospatial Information Services of China, with a resolution of 30 m and based on the 2020 land use classification standard. These data include categories such as forest land, cultivated land, and water bodies, and they were used to extract environmental variables related to the vegetation distribution and surface hydrology. Geological data were derived from field surveys conducted during regional geological mapping projects led by the Anhui Institute of Geological Survey between 2011 and 2014. These datasets include geological conditional factors such as stratigraphy and lithology, bedding orientation and dip angle, and fault structures.

2.2.3. Technical Strategy

This study focused on Yixian County in Anhui Province as the research area, with the objective of constructing a training dataset for landslide identification based on GF-2 satellite imagery, supplementary non-remote sensing data related to geological and environmental conditions, and UAV-based field observations. The dataset was developed through the preprocessing of GF-2 multispectral imagery, ALOS-derived DEMs, and geological map data. By integrating attention mechanisms—specifically a Squeeze-and-Excitation (SE) block—with various CNN architectures, together with the conventional UNet model, landslide spatial information was identified and extracted. To further verify the accuracy and applicability of the remote sensing-based recognition results, creeping landslide imagery from Kecun Town in Yixian County and post-earthquake landslide imagery from Lushan County, Sichuan Province, derived from the Globally Distributed Coseismic Landslide Dataset (GDCLD), were adopted for comparative experiments under the same methodological framework. In addition, field surveys were conducted to validate landslide locations and verify the accuracy of remote sensing-based detection results. Finally, a spatial correlation analysis was performed by combining the identified landslides with key geological and environmental conditional factors, in order to investigate the formation mechanisms of typical landslides in the hilly regions of southern Anhui Province (as shown in Figure 2).

2.3. Data Processing

2.3.1. Image Preprocessing

Using ENVI 5.6 and ArcGIS 10.2 software, radiometric calibration, atmospheric correction, and image fusion were performed on the panchromatic and multispectral bands of the GF-2 satellite. As a result, a four-band remote sensing image with a spatial resolution of 0.8 m was generated. Subsequently, training labels were created based on prior knowledge, and both the RGB true-color imagery of the study area and the corresponding labels were normalized in preparation for input into the deep learning network. To ensure the reliability of the spatial correlation analysis of landslides, the DEM was resampled to match the spatial resolution of the imagery. GIS spatial analysis tools were then used to derive raster layers of the slope, aspect, and elevation. Finally, all raster layers were projected to a unified coordinate system and clipped to the extent of the study area.

In the cross-regional validation experiments, landslide sample annotations were reconstructed at the corresponding locations on GF-2 imagery using the landslide ground-truth labels from Lushan County provided by the GDCLD, followed by the same data preprocessing workflow as in the original study area. Furthermore, to eliminate spectral discrepancies between images from different regions, histogram matching and Pseudo-Invariant Feature (PIF)-based relative atmospheric correction were applied to the Lushan County imagery, aligning its spectral characteristics with those of the study area [39]. Finally, each model was trained and validated using identical model parameter settings.

2.3.2. Sample Dataset Construction

To avoid data leakage during model training, each landslide polygon within the study area was treated as an independent sampling unit. All polygons were randomly divided into spatially independent training, validation and test datasets following a standard ratio of 7:2:1 according to sample quantities. This strategy guarantees that pixels from an identical landslide body will not be distributed across multiple datasets, which effectively eliminates overestimated model accuracy induced by spatial autocorrelation effects. Affected by dense vegetation coverage in the study area, most landslide surfaces are covered by vegetation, leading to highly similar spectral features between landslide bodies and the surrounding background land covers. To mitigate this interference, a field masking method was adopted for normalized remote sensing images to improve landslide detection accuracy [40]. Restricted by the small area and insufficient landslide samples, image patches with a low proportion of landslide pixels were discarded to build a high-quality dataset. Meanwhile, data augmentation strategies were adopted to enrich the diversity of training samples [41]. A sliding window of 128 × 128 pixels and a moving stride of 40 pixels was utilized for image patch cropping. Finally, 3683 image patches were obtained, with a landslide versus non-landslide sample ratio of roughly 1:5.7. Afterwards, two common augmentation operations, horizontal flipping and vertical flipping, were applied to the image patches to further expand the training dataset.

2.4. CNN Model and Validation

2.4.1. Network Models

(1)

SE

The attention mechanism was originally proposed in the field of natural language processing and has since been rapidly extended to various domains, including image recognition, image segmentation, and speech recognition, where it has achieved significant advancements [42]. The mechanism operates by computing the correlations between input features to generate a set of weight coefficients. These weights are then applied to the original features through weighted summation, allowing the model to emphasize important components while suppressing less relevant information. This process enhances the model’s ability to capture and represent critical features [43]. The SE block is a lightweight and efficient channel attention module, which consists of the following three main steps:

Squeeze Operation: Global average pooling is applied to each feature channel to compress the two-dimensional spatial information into a one-dimensional scalar (as shown in Equation (1)), thereby capturing the global contextual information of each channel. This operation reduces the spatial dimensions and enables the network to learn channel-wise global features more effectively.

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

(1)

where H and W represent the height and width of the feature map, respectively; u_c(i,j) denotes the pixel value in the i-th row and j-th column of the c-th channel; and z_c represents the global average value of the c-th channel.

Excitation Operation: The compressed channel descriptors are passed through a nonlinear modeling process to learn inter-channel dependencies and assign weight coefficients to each channel. This operation primarily consists of two fully connected (FC) layers and two activation functions. The first FC layer employs the ReLU activation function and reduces the channel dimensionality to decrease the computational complexity. The second FC layer restores the dimensionality to its original size, followed by a Sigmoid activation function that maps the importance of each channel to a weight between 0 and 1 (as shown in Equation (2)).

$$s=F_{ex}(z,W)=\sigma \left(g\right(z,W\left)\right)=\sigma \left(W_2\delta \right(W_{1}z\left)\right)$$

(2)

where z is the input vector; σ and δ denote the Sigmoid and ReLU activation functions, respectively; and W₁ and W₂ represent the weight matrices of the first and second FC layers, respectively.

Reweight operation: The original feature map u_c of the c-th channel is multiplied by the corresponding channel attention weight s_c generated during the Excitation operation to perform reweighting (as shown in Equation (3)). This process enhances salient features, suppresses redundant information, and improves the model’s ability to extract meaningful representations.

$$\widetilde{x_c}=F_{scale}(u_c,s_c)=s_c·u_c$$

(3)

(2)

SE-CNNs

UNet is a CNN-based image segmentation model that effectively preserves boundary details and spatial information while maintaining a deep semantic understanding [44]. Its core architecture follows a symmetric encoder–decoder structure: The encoder progressively downsamples feature maps through a series of convolution and pooling operations, extracting high-level semantic features while reducing the spatial resolution and increasing the channel depth. The decoder restores the spatial resolution through upsampling operations and incorporates corresponding encoder features via skip connections. This architecture enables the effective fusion of low-level spatial details (e.g., boundary contours) with high-level semantic representations, thereby significantly improving localization accuracy and boundary recognition in pixel-wise classification tasks. As a result, UNet has been widely adopted in various fine-grained segmentation applications [45,46,47]. However, its performance remains limited in complex scenes characterized by ambiguous boundaries (e.g., landslide zones) or severe background interference [48]. To enhance the model’s ability to capture features in critical regions, in this study, we integrate an attention mechanism into the UNet architecture. As shown in Figure 3, the SE module is incorporated into both the downsampling path and the skip connections. This design strengthens the encoder’s channel-wise feature representation and enhances the discriminative channel responses before concatenation with the decoder features, thereby suppressing redundant information and improving the overall quality of feature fusion in the decoding stage.

VGG13 is a CNN that employs sequentially stacked 3 × 3 convolutional kernels combined with 2 × 2 max pooling layers to perform spatial downsampling and progressively extract higher-level semantic features [49]. In this study, while preserving its original architecture, consisting of 10 convolutional layers and 3 FC layers, the SE module is integrated after each convolutional block to enhance feature representation via channel attention. This modification significantly improves the model’s discriminative power and generalization ability, particularly in scenarios involving small-target detection and class imbalance, while maintaining the simplicity and efficiency of the original network design.

ResNet18 is a relatively shallow residual network whose core component—the residual block—consists of two 3 × 3 convolutional layers and employs a shortcut connection to achieve identity mapping, effectively addressing the problem of gradient vanishing [50]. The full network comprises 17 convolutional layers, with channel depth progressively increasing through a sequence of four residual stages designed to capture multi-scale features. These features are ultimately aggregated using a global average pooling layer and passed through a FC layer for classification [51]. In this study, the SE module is embedded within each residual block, positioned after the convolution operations and before the residual summation. This integration enhances the model’s ability to selectively emphasize informative channel features, thereby improving segmentation performance while preserving the lightweight nature of the original architecture.

2.4.2. Hyperparameter Tuning

To improve the model’s segmentation performance under conditions of limited training samples and class imbalance, in this study, we employ a composite loss function that combines dice loss and focal loss, each weighted equally at 50%.

The dice loss function mitigates the adverse effects of class imbalance between positive and negative samples and is particularly effective in scenarios where the target region is small and the background dominates [52]. It emphasizes the overlap between model predictions and ground-truth labels by computing the ratio of their intersection to the sum of their sizes, thereby directly measuring the degree of spatial agreement [53].

$$L_{\mathrm{D}\mathrm{i}\mathrm{c}\mathrm{e}}=1-{\displaystyle \frac{2{\sum }_j^h{\sum }_j^w{p}_{i,j}{g}_{i,j}+ϵ}{{\sum }_i^h{\sum }_j^w{p}_{i,j}+g_{i,j}+ϵ}}$$

(4)

where h and w represent the height and width of the image, respectively, resulting in a total of h × w pixels. p_i,j denotes the predicted value at pixel position (i,j), and g_i,j denotes the corresponding ground-truth label (either 0 or 1). ϵ is a small constant added to prevent division by zero, typically set to 1 × 10⁻⁵.

The focal loss function is specifically designed to address the problem of class imbalance. It retains the advantages of cross-entropy loss while improving the model’s focus on hard-to-classify samples by assigning greater weights to more difficult examples. This enhances the model’s sensitivity to minority classes and improves the overall classification performance [54,55].

$$\mathrm{F}\mathrm{L}\left(p_t\right)={-{\alpha }_t(1-p_t)}^y\mathrm{l}\mathrm{o}\mathrm{g}\left(p_t\right)$$

(5)

where p represents the predicted probability; α is a class-balancing coefficient ranging from 0 to 1 that adjusts the relative weights of positive and negative samples; and γ is a focusing parameter that controls the degree to which the model emphasizes hard-to-classify samples. In this study, α and γ were set to 0.6 and 2, respectively.

In addition to the loss function configuration, the other training hyperparameters were set as follows. The model was trained using the AdamW optimizer [47], with an initial learning rate of 1 × 10⁻⁴. A stepwise learning rate decay strategy was adopted, reducing the learning rate by a factor of 0.1 every 10 epochs. The batch size was set to 32, and the total number of training epochs was 50. To prevent overfitting, early stopping was employed with a patience of 10 epochs. The average training duration was approximately 42 min. All models were trained on a workstation configured with an NVIDIA GeForce RTX 3060 GPU (NVIDIA, Santa Clara, CA, USA) and an AMD Ryzen 5 5600G processor (3.90 GHz) (ADM, Santa Clara, CA, USA).

2.4.3. Accuracy Evaluation

To comprehensively evaluate the performance of the binary classification models in the landslide detection task, the predicted results are compared against the ground-truth labels to construct a confusion matrix. This yields four pixel-level classification outcomes: true positive (TP), false positive (FP), false negative (FN), and true negative (TN). Based on these basic metrics, this study employs Precision, Recall, and F1-score for model evaluation [56]. Furthermore, to conduct a more comprehensive evaluation of model performance, two additional mainstream evaluation metrics are adopted: the Kappa coefficient and Intersection over Union (IoU). The Kappa coefficient quantifies the agreement between model classification outputs and ground-truth labels while eliminating the interference of random consistency caused by accidental matching. By comparison, IoU directly calculates the overlap between predicted landslide regions and ground-truth regions, which acts as a core geometric indicator for pixel-level segmentation accuracy assessment [57,58], as defined in Equations (6)–(10).

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(6)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(7)

$$\mathrm{F}1-\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}={\displaystyle \frac{2\times \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\times \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}}$$

(8)

$$\begin{array}{c}Kappa={\displaystyle \frac{P_{\mathrm{o}}-P_{\mathrm{e}}}{1-P_{\mathrm{e}}}} P_{\mathrm{o}}={\displaystyle \frac{\mathrm{T}\mathrm{P}+\mathrm{T}\mathrm{N}}{N}} \\ P_{\mathrm{e}}={\displaystyle \frac{(\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N})\times (\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P})+(\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{P})\times (\mathrm{T}\mathrm{N}+\mathrm{F}\mathrm{N})}{N^2}}\end{array}$$

(9)

$$\mathrm{I}\mathrm{o}\mathrm{U}={\displaystyle \frac{O}{U}}$$

(10)

where O denotes the Area of Overlap between actual and predicted landslide samples, and U denotes the Area of Union between actual and predicted landslide samples.

3. Results and Evaluation

3.1. UNet-SE Results

According to the landslide identification results of the UNet-SE model in the test area (Figure 4a), landslides are primarily concentrated along steep slopes near ridgelines and valley sides, as well as in areas with abrupt topographic changes, zigzag road edges, and locations adjacent to rivers. The overall distribution is spatially uneven and exhibits localized clustering. To further analyze their spatial characteristics, four representative subregions with a high density of landslides were selected and magnified for detailed visualization (Figure 4b–e). A total of 11 landslides were identified within the test area, which could be categorized into three types based on their spatial distribution patterns: roadside slope zones, natural mountainous zones, and areas adjacent to rivers or residential settlements. This classification primarily considers the geomorphic settings of the landslides and their proximity to human activities or drainage systems, aiming to reflect the distributional tendencies of landslides across different spatial environments rather than strictly relying on their triggering mechanisms or kinematic types.

Roadside Slope Zones: Landslides in roadside slope zones are primarily concentrated along the steep slopes on both sides of mountain roads, including Landslide Points Y1, Y4, Y6, Y8, and Y10. Landslide Point Y1 exhibits a tongue-shaped extension toward the southeast (SE), with an aspect ranging from the northeast (NE) to the southeast (SE). It covers a relatively large area, is located adjacent to the road and near residential buildings, and shows obvious surface exposure. In the true-color imagery, it appears as a bare yellow tone with a smooth texture and light hue, contrasting sharply with the coarse and irregular texture of the surrounding natural vegetation (Figure 4b). Landslide Point Y4 is a small, elongated strip located at a sharp bend in the road, sliding toward the road edge. It has a clearly defined boundary, a bright tone in the imagery, and a coarse texture (Figure 4b). Landslide Point Y6 also presents as a small tongue-shaped feature, sliding toward the southeast. Although partially covered by vegetation, its fine and smooth texture makes the landslide boundary distinguishable in the imagery (Figure 4c). Landslide Point Y8 is located at the junction of the inner slope of a mountain road and a gully, presenting a small-scale, short tongue-shaped morphology with a dominant sliding direction of south to southeast (S-SE). This landslide is limited in size, and locally exposed soil appears light yellow to light brown in color. The internal surface shows slight undulations with a relatively smooth texture (Figure 4d). Landslide Point Y10 is situated on the steep outer slope of an S-shaped curve along the mountain road, showing a short tongue-like morphology inclined toward the south or southeast. In the image, the exposed central soil mass appears yellow-white, surrounded by light green vegetation, with a clearly defined boundary (Figure 4e).

Natural Mountainous Zones: Landslides in natural mountainous zones are spatially distant from roads and residential areas, and they are typically found in regions with significant topographic relief. This category includes Landslide Points Y3, Y5, and Y9. Landslide Point Y3 is located in an elevated position near the central part of the mountain. It covers a relatively small area and slides toward the southeast (SE). In the imagery, it appears as a yellowish-white, elliptical-shaped feature that contrasts clearly with the surrounding vegetation (Figure 4b). Landslide Point Y5 has a complete and irregular outline, covers a larger area, and appears bright white-yellow in the image. The landslide originates from the ridge crest and extends downslope along both flanks, with a sliding direction consistent with the natural terrain gradient (Figure 4c). Landslide Point Y9 is situated on a steep southwestern slope and presents a narrow, strip-like morphology sliding toward the southwest (SW). In the imagery, it is characterized by a grayish-yellow tone (Figure 4e).

River-Adjacent or Residential Zones: Landslides in river-adjacent or residential zones are located near rivers, villages, and farmland, where secondary sliding events tend to pose the greatest hazards and economic losses. This category includes Landslide Points Y2, Y7, and Y11. Landslide Point Y2 is situated on a steep slope above a settlement and exhibits a tongue-shaped slide pattern oriented toward the southeast (SE). The imagery reveals a rough texture surrounded by a large area of bare, light-yellow ground, which may include both the landslide body and adjacent agricultural land (Figure 4e). Landslide Point Y7 appears as a narrow strip with a bright yellow tone in the imagery. It is located along the riverbank and is inclined in a southwest (SW) direction (Figure 4d). Landslide Point Y11 is the largest among all identified landslides. It features a rough and structurally fragmented texture with clearly defined boundaries. The landslide slides southeastward toward the river and is composed of an upper wide-body section and a lower narrow-body section along a continuous path. This morphology suggests the possibility of multiple overlapping secondary sliding events (Figure 4e).

The UNet-SE model demonstrates strong performance in accurately identifying major landslide locations in complex hilly terrain. It effectively captures landslides across varying spatial scales, preserving their continuity and complete morphology. These results indicate the model’s robust capability in extracting the overall spatial characteristics of landslides.

3.2. Comparative Validation of Methods on the Study Area

To evaluate the landslide detection performance of four deep learning models (UNet, UNet-SE, VGG13-SE, and ResNet18-SE) under different conditions, three representative landslide points (Points Y1, Y2, and Y5) within the test area were selected as evaluation samples. The recognition results of the four models were then compared and analyzed based on these samples.

Roadside Slope Zones: Landslide Point Y1—The results obtained from ResNet18-SE show a fragmented structure with blurry and discontinuous boundaries. Notably, some omission errors are also observed at the slope toe in the northeastern direction. In contrast, VGG13-SE roughly reconstructs the continuous sliding surface and performs better than ResNet18-SE in maintaining spatial coherence. However, both models show misclassification along the road edge, mistakenly identifying bare ground as landslides. The UNet and UNet-SE both delineate the landslide head scarp with higher accuracy, whereas UNet-SE yields the fewest false positives overall (Figure 5d–g, Row 1, red circles).

Natural Mountainous Zones: Landslide Points Y5—In areas with complex textures along landslide boundaries or shaded hillsides, ResNet18-SE, VGG13-SE, and UNet all exhibit varying degrees of misclassification. In contrast, the extraction results obtained from UNet-SE align closely with the actual landslide contours, providing more precise and continuous delineation of landslide margins and effectively minimizing both false positives and false negatives (Figure 5d–g, Rows 3, red circles).

River-Adjacent or Residential Zones: Landslide Point Y2—Both UNet and ResNet18-SE incorrectly identify the shaded, back-slope side of the hillside as a landslide. VGG13-SE includes a small, vegetated patch located in the transition zone between bare ground and the landslide area, leading to significant boundary overestimation. In contrast, UNet-SE produces more accurate results and effectively suppresses background interference (Figure 5d–g, Row 2, red circles). It is worth noting, however, that the landslide boundary at this location exhibits low visual contrast in the imagery, and the background contains extensive bare-colored areas. All four models show some degree of over-segmentation, and the results at this location require further validation through field investigation.

3.3. Transfer Regional Comparison

To evaluate the cross-regional generalization capability of the models, two validation areas were selected: Kecun Town in Yixian County and Lushan County in Sichuan Province. Kecun shares similar geomorphological characteristics with the study area; both are creeping landslide zones in the hilly region of southern Anhui. In contrast, Lushan represents an earthquake-induced landslide area in the central mountainous region of western Sichuan, exhibiting notable differences in geomorphology and landslide-triggering mechanisms. The experimental results for each validation area are presented and analyzed below.

3.3.1. Cross-Regional Validation in Kecun

A total of nine landslides (K1–K9) are distributed within the Kecun Town cross-regional validation area. These landslides are generally small in scale and spatially dispersed, mostly occurring as scattered patches on steep slopes, along both sides of valleys, and at road cut slopes. The results in Figure 6 show that the model clearly delineates the boundaries of each landslide, with recognition results closely matching the actual spatial locations. However, misclassifications occur in some areas with severe background interference and similar textures. Overall, the UNet-SE model accurately captures the information of small, dispersed landslides in the Kecun Town area. The model’s satisfactory recognition performance in this validation area further demonstrates its transferability for identifying small-scale, scattered landslides under similar geological environments.

3.3.2. Cross-Regional Validation in Lushan

A total of seven landslides (L1–L7) were identified in the Lushan cross-regional validation area, all of which are characterized by relatively large scales and spatially concentrated distribution. Figure 7 presents the overall recognition results of the UNet-SE model in this area. The results show that UNet-SE successfully identified the main extents of all seven landslides with considerable accuracy, and the recognition results closely match the ground-truth labels in spatial location. For landslides with larger areas and relatively clear boundaries (e.g., L3 and L6), the model produced continuous and complete recognition results with well-defined main sliding body contours. For landslides with blurred boundaries or complex sliding surface shapes (e.g., L5 and L7), the model still effectively captured the main landslide extent, albeit with some overestimation. Overall, UNet-SE demonstrates good identification performance in the earthquake-induced landslide area of Lushan, further validating its applicability and generalization capability in large-scale, earthquake-triggered landslide identification tasks.

3.3.3. Comparative Validation and Analysis

To further evaluate the performance differences in the models under different geomorphological types and landslide-triggering conditions, three representative landslide points were selected for comparative analysis: Landslide point K8 in the cross-regional validation area of Kecun, Landslide point L6 in the cross-regional validation area of Lushan, and Landslide point Y11 in Hongtan. Figure 8 presents the recognition results of the three landslide points obtained by the UNet, ResNet18-SE, VGG13-SE, and UNet-SE models. A detailed comparative analysis of model performance is conducted for each landslide point, focusing on boundary integrity, spatial continuity, misclassification patterns, and accuracy evaluation.

Landslide Point K8: The middle and rear parts of this landslide are characterized by dense vegetation cover, which significantly increases the difficulty of identification. Overall, ResNet18-SE produces fragmented and discontinuous landslide boundaries, VGG13-SE exhibits obvious omission errors, and UNet fails to effectively extract landslide features in areas with high vegetation cover. In contrast, UNet-SE delineates a more complete landslide outline that closely matches the ground-truth labels, maintaining good identification performance even in densely vegetated regions (Figure 8d–g, Row 1, red circles).

Landslide Point L6: This landslide presents an irregular fan-shaped morphology, characterized by a large area of bright white exposed surface and a sliding direction toward the SE. Its lower portion is adjacent to roads and residential areas, making it prone to misclassification. In this region, all models except UNet-SE misidentify linear road features as landslides to varying degrees, highlighting UNet-SE’s superior boundary delineation capability and its effectiveness in suppressing background interference (Figure 8d–g, Row 2, red circles).

Landslide Point Y11: This landslide is structurally complex, with a curved and blurred boundary, and it is classified as a multi-stage sliding type, making it particularly difficult to detect. Among the four models, UNet-SE achieves the best performance, not only successfully identifying the main sliding body but also capturing a suspected secondary landslide area. Its output most closely matches the actual label boundary, reflecting its superior generalization capability in complex terrain (Figure 8f, Row 3, red circle). In comparison, for the slope toe oriented toward the southeast, UNet exhibits extensive omission errors. Similarly, ResNet18-SE identifies the main landslide body as several disconnected segments, failing to reconstruct a continuous slip surface, while VGG13-SE approximates the overall extent of the main sliding area; however, both fail to detect the secondary landslide surface covered by dense vegetation (Figure 8d,e,g, Row 3, red circles).

Table 2. Accuracy metrics for the study area and comparison areas.

Area	Models	Precision	Recall	F1-Score	Kappa	IoU
Kecun	ResNet18-SE	0.745	0.751	0.748	0.704	0.597
	VGG13-SE	0.709	0.756	0.731	0.684	0.577
	UNet	0.761	0.689	0.723	0.678	0.566
	UNet-SE	0.784	0.755	0.769	0.730	0.625
Lushan	ResNet18-SE	0.674	0.850	0.752	0.676	0.602
	VGG13-SE	0.641	0.881	0.742	0.659	0.590
	UNet	0.679	0.840	0.751	0.675	0.601
	UNet-SE	0.693	0.854	0.765	0.694	0.620
Hongtan	ResNet18-SE	0.847	0.604	0.705	0.637	0.545
	VGG13-SE	0.890	0.651	0.752	0.694	0.603
	UNet	0.861	0.634	0.731	0.667	0.575
	UNet-SE	0.911	0.685	0.782	0.730	0.642

presents the landslide identification accuracy metrics of the four models across three regions: the Hongtan study area test zone, the Kecun cross-regional generalization validation area, and the Lushan cross-regional generalization validation area.

In the Hongtan study area test zone, UNet-SE performs best in the landslide recognition task. According to the confusion matrix, UNet-SE achieved a TP of 800,082, a TN of 3,947,553, an FN of 367,921, and an FP of 78,172. It achieves a high Precision of 0.911, indicating that the vast majority of areas predicted as landslides correspond to actual landslides, with the fewest FPs and a stronger suppression of background interference. Additionally, the model achieves a Recall of 0.685, surpassing ResNet18-SE (0.604), VGG13-SE (0.651), and UNet (0.634). This suggests that UNet-SE results in the fewest missed detections (FNs), enabling a more complete reconstruction of landslide structures and improved recognition of small-scale landslides. As the harmonic mean of Precision and Recall, the F1-score further reflects the model’s balance between accuracy and completeness. UNet-SE achieves the highest F1-score at 0.782, confirming its superior overall performance. Meanwhile, this model yields a Kappa coefficient of 0.730 and an IoU of 0.642, both superior to those of other comparative models. This demonstrates that it possesses the optimal chance-corrected classification consistency and the highest spatial overlap consistency between predicted landslide areas and ground-truth areas.

In the cross-regional generalization validation experiments carried out in Kecun and Lushan, the UNet-SE model obtained the optimal F1-score across both study sites. In Kecun, the UNet-SE model delivered competitive detection performance (F1 = 0.769, Precision = 0.784, Recall = 0.755, Kappa = 0.730, IoU = 0.625) with well-balanced evaluation metrics. According to the confusion matrix, it yielded TP = 313,576, TN = 2,316,323, FN = 101,756, and FP = 86,393. Kecun has consistent geomorphological and geological conditions with the primary study area, given that both sites belong to the hilly zone of southern Anhui Province. Accordingly, the model suffered minor performance degradation in this region: its F1-score dropped by roughly 1.3% relative to the result obtained in Hongtan, while Kappa and IoU remained steady. This outcome verifies the reliable transferability of the UNet-SE model under similar topographic backgrounds. In Lushan, the UNet-SE model yielded an F1-score of 0.765, a Precision of 0.693, a Recall of 0.854, a Kappa of 0.694, and an IoU of 0.620. According to the confusion matrix, it achieved TP = 521,131, TN = 2,058,883, FN = 89,093, and FP = 230,861. Compared with Kecun, Lushan presented lower Precision but higher Recall, accompanied by slight declines in Kappa and IoU values. Such performance differences mainly stem from different landslide types: landslides in Lushan are mostly co-seismic landslides featured by larger spatial scales, wide exposed sections and distinct boundary contours. Nevertheless, these landslides share highly similar spectral signatures with surrounding bare land and broken slopes, which brings more commission errors. In contrast, the large landslide scale and prominent surface features help reduce omission errors. Importantly, the F1-score of the UNet-SE model in Lushan only declined by about 1.7% compared with Hongtan, demonstrating a limited overall performance drop across different regional scenarios.

In summary, across the two cross-regional generalization experiments, the UNet-SE model exhibited stable generalization performance under both analogous and heterogeneous geomorphological conditions. Its F1-score degradation was kept within 2% for both validation sites, and both Kappa and IoU remained at relatively high levels. These findings confirm the strong generalization robustness of the UNet-SE model under varying geomorphological backgrounds and different landslide-triggering mechanisms.

3.4. Field Validation

To illustrate the representative characteristics of landslide samples from different datasets in the study area and to verify the reliability of the model identification results, representative landslide points from the training, validation, and test datasets were selected for comparative analysis, as shown in Figure 9. Specifically, Figure 9e,f present the field-validated UAV images of test-set landslide points Y1 and Y2 (Figure 5), respectively. The yellow dashed lines indicate the actual landslide boundaries delineated through manual interpretation. The main characteristic parameters of the landslides, including area, slope, aspect, and lithology, are summarized in

Table 3. Field data statistics.

No.	Area (m2)	Slope (°)	Aspect	Lithology	Dip Direction	Rock Sample	XPL (×20)
Y1	19,924	30–45	NE	Silty mudstone	+
Y2	13,546	25–30	SW, SE	Calcareous mudstone	+
Y4	1191	40	SE	Silty mudstone	+
Y11	21,627	35	SE	Calcareous shale	+

Note: “+” denotes compliant landslides, where the sliding direction aligns with the dip of the rock strata. XPL (×20): cross-polarized light microscopy at 20× magnification.

.

Landslide Point Y1 (Figure 9e): This landslide exhibits a northeast (NE) orientation and represents a typical mountainous slope failure. It occurred on a steep hillslope, with an estimated gradient of approximately 45° near the crown, gradually decreasing to around 30° at the toe adjacent to the road. The landslide boundaries are well-defined, and its overall morphology displays a tongue-shaped downslope movement. The surface has been significantly disturbed, and the landslide is relatively extensive in scale, covering an area of approximately 19,924 m²—nearly the entire hillslope. The central portion of the slide is largely devoid of vegetation, characterized by exposed bare ground, while the marginal zones are sparsely colonized by naturally regenerated grasses and shrubs. These features suggest that the landslide has entered a phase of relative stabilization. The lithology of the affected zone is silty mudstone, with well-developed joint structures observed in rock samples. Under 20× cross-polarized light microscopy, the rock shows densely packed fine-grained particles. This lithological composition is typically associated with low mechanical strength, high weathering susceptibility, and a pronounced tendency toward softening and instability.

Landslide Point Y2 (Figure 9f): This landslide is situated at the base of the hillside and consists of a composite failure formed by multiple secondary landslide events. The total affected area is approximately 13,546 m², and the terrain exhibits relatively gentle relief, with slopes ranging from 25° to 30°. Agricultural land surrounds the landslide body, and portions of the area have been artificially reclaimed into terraced fields following the stabilization of the slide. This land use condition results in minimal visual distinction between the landslide body and the surrounding background, making the landslide highly concealed and easily confused with non-landslide features. The lithology of the landslide zone is calcareous mudstone, which displays a dense internal structure and well-developed lamination. This type of rock typically has a low shear strength along the bedding planes and is prone to softening upon water infiltration, often leading to sliding along stratification or joint surfaces. These characteristics make it a classic example of a weak, failure-prone rock layer.

Table 3 presents characteristic data of four representative landslides within the test area. The results indicate that the slope gradients of these landslides are generally greater than 30°, reflecting pronounced topographic relief and substantial variation in landslide scale. These landslides are predominantly along-layer types, meaning that their sliding directions are largely consistent with the dip direction of the underlying rock strata. This pattern suggests strong structural control by the overlying sedimentary formations. Moreover, the lithological composition of the area is dominated by softer rock types such as silty mudstone, calcareous mudstone, and calcareous shale, as evident in the hand specimens. These rocks exhibit significant internal fracturing, as observed under cross-polarized light microscopy (XPL), and they are prone to weakening upon water infiltration, which facilitates the development of slip surfaces and increases the likelihood of slope failure.

4. Correlation Analysis of Conditional Factors

4.1. Extraction of Conditional Factors

Landslides typically result from a combination of internal and external factors, with causative mechanisms that are both complex and regionally variable. To identify the key conditions contributing to landslide occurrences in the study area, seven representative conditional factors encompassing both environmental and geological domains were selected: elevation, the NDVI, slope, aspect, distance from rivers, distance from faults, and rock hardness. Among these, the classification standard for rock hardness is primarily based on the Mohs hardness measurement results of rock samples, dividing rock hardness into four grades.

To ensure data uniformity and support subsequent model input, a standardized preprocessing workflow was conducted in ArcGIS. Firstly, all datasets were unified into an identical projected coordinate system. The 12.5 m ALOS DEM was resampled to align with the 0.8 m spatial resolution of GF-2 imagery, and cubic convolution resampling was adopted to preserve topographic details and reduce geometric distortion to the maximum extent [59]. Based on the preprocessed geospatial data, seven conditioning factor layers were extracted: elevation, slope and aspect derived from the resampled DEM; NDVI calculated from GF-2 remote sensing images; river distance and fault distance generated from vector data via the Euclidean distance tool; and rock hardness classified from geological vector data in accordance with the Mohs hardness standard. Afterwards, all factor layers were uniformly clipped to the study area boundary to guarantee consistent spatial coverage and accurate pixel registration, followed by unified reclassification and normalization (Figure 10). Ultimately, the processed raster and vector datasets were stored in the Hadoop Distributed File System (HDFS). The GeoWave 1.2.0 tool was further used to build spatial indices and establish a complete multi-factor spatial database.

4.2. Correlation Analysis

4.2.1. Probability Analysis Using FR

To elucidate the spatial distribution patterns of landslides in relation to environmental factors, the FR method was employed to quantitatively assess seven key hazard-inducing variables: elevation, the NDVI, slope, aspect, distance from rivers, distance from faults, and rock hardness. The FR is defined as the ratio between the proportion of landslide pixels and the proportion of the total area within a given factor class (FR = ratio of landslides/ratio of domain). An FR value greater than 1 indicates an increased relative probability of landslide occurrence under that condition, whereas a value less than 1 indicates a decreased probability (

Table 4. Spatial relationship between conditional factors and landslide occurrence based on FR analysis.

Factor	Level	Ratio of Landslides	Ratio of Domain	FR
Elevation	<200	0.156	0.059	2.652
	200–300	0.395	0.443	0.891
	300–400	0.065	0.320	0.204
	400–500	0.384	0.146	2.629
	>500	0.000	0.032	0.000
NDVI	0–0.2	0.007	0.033	0.223
	0.2–0.4	0.425	0.516	0.824
	0.4–0.5	0.372	0.281	1.324
	0.5–0.6	0.181	0.150	1.208
	>0.6	0.014	0.021	0.679
Slope/°	<10	0.298	0.221	1.344
	10–20	0.122	0.098	1.247
	20–30	0.122	0.117	1.039
	30–35	0.358	0.428	0.836
	>35	0.100	0.135	0.741
Aspect	North	0.020	0.067	0.296
	Northeast	0.256	0.120	2.133
	East	0.068	0.036	1.890
	Southeast	0.123	0.145	0.848
	South	0.064	0.057	1.131
	Southwest	0.159	0.155	1.025
	West	0.004	0.046	0.080
	Northwest	0.009	0.152	0.056
	Plane	0.297	0.221	1.343
Distance from rivers/m	<50	0.046	0.110	0.416
	50–150	0.231	0.196	1.180
	150–300	0.120	0.235	0.508
	300–500	0.360	0.227	1.584
	500–1000	0.244	0.209	1.165
	>1000	0.000	0.023	0.000
Distance from faults/m	<100	0.031	0.059	0.528
	100–200	0.003	0.059	0.049
	200–500	0.028	0.174	0.162
	500–1000	0.484	0.257	1.885
	1000–3000	0.344	0.285	1.209
	>3000	0.110	0.166	0.658
Rock hardness	Hard rock	0.000	0.000	0.000
	Moderately hard rock	0.000	0.027	0.000
	Moderately soft rock	0.238	0.487	0.488
	Soft rock	0.762	0.487	1.567

) [60]. The analysis revealed that landslide occurrences exhibited distinct clustering trends across various classifications of each factor.

(1)

Environmental factors

Elevation: Approximately 38.4% of the identified landslides were concentrated within the 400–500 m elevation range, with a FR of 2.629 (Table 4), indicating this zone as highly susceptible to landslide occurrence. In contrast, no landslides were observed in areas above a 500 m elevation, suggesting that higher-altitude zones exhibit greater terrain stability.

NDVI: The 0.4–0.5 NDVI interval exhibited a FR of 1.324 (Table 4), indicating a higher frequency of landslide occurrence in areas with moderate vegetation cover. This pattern may be attributed to post-landslide ecological recovery, where the regrowth of weeds and shrubs increases the NDVI values in previously disturbed zones.

Slope: Although the 30–35° slope interval accounted for the highest proportion of landslides (35.8%), its FR was only 0.836—lower than that of slopes less than 10° (FR = 1.344, Table 4). This indicates a higher landslide density per unit area in gently sloping terrain, suggesting that such areas may be more susceptible to slope failure than usually expected. This statistical result does not contradict the field observation that representative landslides predominantly occur on steeper slopes, as the two reflect different scales of landslide distribution. The FR method measures landslide area per unit area, rather than the typicality or individual scale of landslide occurrence.

Aspect: A significant concentration of landslides was observed on northeast-facing (FR = 2.133) and east-facing (FR = 1.89) slopes (Table 4), which may be attributed to differential erosion resistance influenced by variations in solar radiation and rainfall exposure.

Distance from rivers: In the 300–500 m buffer zone from rivers, landslides accounted for 36.0% of the total, with a FR of 1.584 (Table 4), indicating that fluvial disturbance and toe erosion had a significant impact on slope stability. In contrast, landslides were rare within 50 m of rivers (FR = 0.416), which may be attributed to the presence of engineered stabilization measures such as retaining walls or embankments.

(2)

Geological factors

Distance from faults: A total of 48.4% of landslides occurred within the 500–1000 m buffer zone from fault lines, with a FR of 1.885 (Table 4), indicating heightened landslide susceptibility in this range. This distribution pattern may be attributed to structural variations in the surrounding rock mass: while the core zone of the fault is typically characterized by dense, compacted rock due to tectonic compression, the adjacent fracture zones tend to contain more loosely structured, fragmented rock, which is more vulnerable to failure under rainfall infiltration and surface erosion.

Rock hardness: A total of 76.2% of landslides occurred in very soft rock areas, corresponding to an FR of 1.567 (Table 4). The loose structure and low shear strength of soft rock formations make them highly vulnerable to destabilization when exposed to surface water infiltration. In contrast, landslide susceptibility was significantly lower in areas with moderately soft rock (FR = 0.488) and virtually absent in hard rock zones (FR = 0), underscoring the critical role of lithologic strength in slope stability.

The characteristics of conditions highly associated with landslide occurrence can be summarized as follows: elevations ranging from 400 to 500 m, NDVI values between 0.4 and 0.5, slopes less than 10°, northeast- or east-facing aspects, distance of 300–500 m from rivers and 500–1000 m from faults, and lithologies consisting of soft sedimentary rocks. This combination of conditional factors indicates that landslides are more likely to occur in areas characterized by pronounced topographic relief, weak and fractured rock structures, and intense hydrological disturbance.

4.2.2. Importance Analysis Using RF

Random Forest (RF) is well-suited for handling high-dimensional, nonlinear relationships, and it is effective in identifying the relative contributions of input variables [61]. To quantify the dominant influence of each landslide-inducing factor, the RF model was applied to evaluate the importance of the seven selected variables (Figure 11).

Aspect exhibits the highest importance score (0.235), indicating that variations in slope aspect significantly influence rock weathering and erosion intensity by altering solar radiation and rainfall distribution patterns. These processes, in turn, govern landslide susceptibility. This finding is consistent with the FR analysis, where NE- and E-facing slopes showed elevated FR values exceeding 1.89. Distance from rivers (0.2) and distance from faults (0.171) represent the effects of hydraulic erosion and tectonic fracture zones, respectively, highlighting their combined influence on slope stability. In terms of elevation (0.135) and the NDVI (0.115), elevation variability is closely associated with climate-driven weathering processes, while the NDVI reflects the capacity of vegetation to stabilize slopes. Areas with sparse vegetation cover (NDVI < 0.4) are more prone to soil erosion and, consequently, increased landslide susceptibility. Rock hardness (0.077) and slope (0.067) exhibit relatively low importance: the study area is dominated by soft rock lithologies (76.2% of the total; Table 4) with minimal lithological variability. In addition, slope gradients are largely concentrated within a narrow range (20–35° accounting for 56.3% of the total), resulting in limited contrast in slope conditions.

The RF analysis results indicate that landslide triggering is governed by the coupling of multiple factors. Although rock hardness and slope individually exhibit low contributions in single-factor analyses, their interactions with other factors (e.g., soft rock combined with heavy rainfall) can still directly trigger landslides. Future studies should aim to quantitatively evaluate these interaction mechanisms in order to enhance the accuracy of early warning systems.

4.2.3. Sensitivity Analysis Using SHAP

Currently, some studies have focused on neural network updating or feature selection from the perspective of various sensitivity analysis techniques [62,63,64]. However, the FR and RF methods can separately output frequency ratio values for each factor category and factor importance rankings, but fail to capture the nonlinear interactions and bidirectional driving effects of landslide conditioning factors, thus limiting the interpretation of landslide formation mechanisms. In this study, SHapley Additive exPlanations (SHAP) was employed for pixel-level nonlinear sensitivity analysis [65,66], which can quantify the directional contribution of each geo-environmental factor to landslide initiation.

As illustrated in the SHAP distribution plot (Figure 12), positive SHAP values on the right side facilitate landslide occurrence, whereas negative SHAP values on the left side restrain landslide development. Elevation exhibits an obvious bimodal threshold effect: both low and high elevation zones correspond to positive SHAP values that raise landslide susceptibility, while medium elevation zones fall within negative SHAP intervals and suppress landslide activities. In terms of fault distance and river distance, short distances (areas close to river channels and fault zones) generate negative SHAP values with inhibitory impacts on landslides; by contrast, moderate and relatively long distances produce positive SHAP values and greatly elevate landslide susceptibility. Rock hardness and slope gradient present stable positive driving effects, suggesting that loose rock masses and gentle slopes are more susceptible to landslide hazards. Slope aspect displays scattered data points across both positive and negative intervals without obvious monotonic trends, demonstrating its complex nonlinear influence on shallow creeping landslides. Meanwhile, NDVI yields SHAP values concentrated around zero, revealing that vegetation coverage exerts a negligible independent impact on landslide prediction in the study area.

Comparative analysis indicates consistent results as well as complementary discrepancies among FR, RF and SHAP outputs. Slope ranks last in the RF importance evaluation with a contribution score of 0.067, yet low slope values maintain a stable positive driving effect in SHAP analysis. FR results further verify that gentle slopes (<20°) are high-risk zones for landslide development, which cross-validates the critical role of gentle slopes in local landslide formation. Elevation ranks fourth in the RF importance ranking, but its distinctive bimodal threshold characteristic can only be detected via SHAP, given that its driving effect takes effect only when exceeding two critical thresholds. Moreover, FR recognizes fault distances of 500–1000 m and river distances of 300–500 m as high-susceptibility ranges, while SHAP further quantifies the directional contribution of the two distance factors (Figure 12) and clarifies their threshold pattern: short distances inhibit landslide occurrence, while relatively long distances promote landslide development. The integrated combination of RF and SHAP, supplemented with statistical verification from FR, enhances both the prediction accuracy and physical interpretability of landslide susceptibility mapping. Meanwhile, it also proves the core strengths of SHAP in extracting nonlinear thresholds, directional factor contributions and interactive relationships among conditioning factors.

5. Discussion

5.1. CNNs + SE Model Performance of Landslide Identification

The hilly regions of eastern China are characterized by small, scattered, and highly concealed landslides under dense vegetation, which severely restricts the accuracy of remote sensing identification. In this study, the adopted 0.8 m high-resolution GF-2 satellite imagery delivers refined micro-geomorphic features and subtle spectral information, which builds a solid data basis for capturing weak characteristic signals of concealed landslides [67]. On this basis, the integration of a CNNs + SE model effectively enhances the feature extraction ability for complex landslide targets. In particular, the UNet-SE model embeds SE modules into the encoding and skip-connection stages of the original UNet, which realizes adaptive weight assignment of key feature channels. This design strengthens the perception of fuzzy boundaries and small-scale landslides, suppresses interference from vegetation, bare land, and farmland, and reduces boundary localization errors by 19%. Quantitative results verify that the UNet-SE model obtains the optimal overall performance, with a Precision of 0.911, a Recall of 0.685, an F1-score of 0.782, a Kappa coefficient of 0.730 and an IoU of 0.642. It outperforms the original UNet, ResNet18-SE and VGG13-SE models in all evaluation indicators. Nevertheless, although GF-2 high-resolution imagery contributes to accurate landslide detection, its high data acquisition cost restricts its popularization and practical application in large-scale regional landslide monitoring. Future work will concentrate on exploring more cost-effective remote sensing data sources to promote the engineering application of relevant landslide detection models.

Although the Recall value (0.685) of UNet-SE is relatively modest due to the high concealment and limited sample size of landslides, the optimized hybrid loss function and data augmentation strategy have effectively reduced missed detections. Further analysis indicates that the high omission rate (approximately 31.5%) primarily occurs in three specific types of landslides: those with spectral characteristics similar to farmland, extremely small-scale landslides, and those with blurred boundaries caused by dense vegetation cover. In the future, detection accuracy for such difficult-to-identify landslide samples can be further enhanced by fusing multi-source geospatial data, including DEM and SAR imagery, and by adopting multi-scale attention mechanisms and semi-supervised learning strategies. Furthermore, building multi-temporal remote sensing sequence datasets to support dynamic landslide monitoring will be another vital research direction. By comparison, ResNet18-SE and VGG13-SE exhibit unsatisfactory detection performance. Specifically, ResNet18-SE has a relatively shallow network structure and a limited receptive field, which restricts its ability to extract global contextual information. VGG13-SE merely stacks convolutional layers sequentially and lacks advanced spatial feature fusion modules, so it fails to distinguish landslide pixels from complex background pixels in scenarios with variable landslide sizes and vague boundaries. This defect eventually causes severe over-segmentation and blurry landslide contours. In comparison, the UNet-SE model remedies spatial feature loss during downsampling via its classic encoder–decoder structure and skip connections. Meanwhile, embedded SE modules strengthen adaptive channel feature screening. Such structural advantages allow the model to precisely locate small-scale, irregular and blurred concealed landslides, reduce false detection errors, and guarantee intact and accurate landslide segmentation results. It is worth mentioning that ResNet18, VGG13 and vanilla UNet were selected as baseline models mainly due to limited training samples in this study. Deeper and more advanced network architectures are highly susceptible to overfitting, given the insufficient dataset size.

Compared with published studies, the proposed UNet-SE model yields competitive performance under more challenging identification scenarios. Hegde et al. embedded attention mechanisms into the U-Net architecture and tested the model on a landslide dataset collected in Bijie, Guizhou Province. Their study area featured obvious landslide characteristics and diverse landslide types, with the model obtaining an F1-score of 0.74 [68]. Hu et al. developed the CALandDet model based on a large-scale global multi-regional benchmark dataset with abundant landslide samples and categories, achieving an F1-score of 0.8164 [69]. Zhang et al. constructed MB-Net for creeping landslide detection using Sentinel-1 InSAR data, which reached an F1-score of 80.91% and an IoU of 67.94% on the public ISSLIDE dataset [70]. In comparison, landslides in our study area pose far greater detection difficulties. These landslides are dominated by ultra-small-scale bodies with fuzzy boundaries; dense vegetation coverage leads to weak spectral features, and the available training samples are also limited. Despite such harsh conditions, our UNet-SE model still achieves an F1-score of 0.782 and an IoU of 0.642. Specifically, compared with the model proposed by Hegde et al., our method achieves a higher F1-score in a more complex study area, verifying its superior feature extraction capacity. Although our F1-score is 3.4 percentage points lower than that reported by Hu et al., our study area has substantially higher detection difficulty than their global multi-regional dataset, which demonstrates the promising robustness of our model. As for MB-Net, the two models only have an IoU gap of 3.7 percentage points. Nevertheless, MB-Net relies on explicit InSAR deformation signals as direct physical evidence for landslide detection, while our model only uses indirect, concealed spectral and texture information derived from optical remote sensing images. This further proves the powerful deep feature mining ability of UNet-SE under insufficient effective feature information. In conclusion, despite the significantly higher landslide detection difficulty in our study area relative to the above comparative studies, the UNet-SE model still delivers competitive detection performance. The results validate the reliability and applicability of the proposed model for concealed small-scale landslide recognition in complex hilly environments.

5.2. Cross-Regional Generalization Ability

In this study, cross-regional generalization tests were implemented to evaluate the spatial adaptability of the UNet-SE model, covering intra-regional independent validation in Hongtan (Yixian County), inter-regional transfer validation in Kecun (Yixian County), and cross-regional extrapolation in Lushan County (Sichuan Province). Four models, namely UNet-SE, vanilla UNet, ResNet18-SE and VGG13-SE, were adopted for transfer learning. All models were initialized with pre-trained weights obtained from the source domain (Hongtan, Yixian County) to further explore the generalization ability driven by network architectures. The cross-regional experimental results presented a consistent variation trend: both Kecun and Lushan yielded obviously higher Recall values than Hongtan, accompanied by slight decreases in F1-score, Kappa coefficient and IoU (Table 2). The improved Recall is mainly attributed to more distinguishable landslide features in the two validation areas. Specifically, Lushan is dominated by large-scale exposed co-seismic landslides with clear boundaries. Although Kecun belongs to the same southern Anhui hilly area as Hongtan, local landslides there are easier to identify than the small-sized, highly concealed and vegetation-shielded landslides in Hongtan. By contrast, the slight drop of F1-score, Kappa and IoU stems from reduced Precision, which is caused by limited false positive samples induced by domain shifts. Such domain discrepancies include differences in topographic conditions, land cover complexity, surface spectral characteristics and background noise interference.

Notably, ablation experiments further validated that the SE channel attention module constitutes the core contributor to the model’s generalization capacity, Recall, and F1-score stability. The baseline UNet model without the SE module displayed lower cross-regional Recall and a more evident F1-score decline (Table 2), proving that adaptive feature weighting strengthens the extraction of intrinsic landslide features and suppresses interference from complex backgrounds. Overall, the UNet-SE model possesses favorable regional generalization and feasible method-based transferability; it achieves high Recall to ensure disaster identification completeness in cross-regional scenarios, with a slight F1-score reduction caused by controllable false positives, and can be steadily applied to both creeping and earthquake-induced landslide scenarios.

5.3. Landslide Controlling Conditions and Sustainability Implications

Yixian County in southern Anhui is a typical hilly area in eastern China, featuring large topographic relief, active faults, and complex lithology. The regional bedrock is mainly composed of soft sedimentary rocks such as silty mudstone and calcareous mudstone, which feature low shear strength and weak weathering resistance. Weathering of these bedrocks produces silty clay with low shear strength and high water sensitivity. This clayey soil is susceptible to strain softening under rainfall infiltration, and acts as a critical internal controlling factor governing landslide occurrence [71,72]. Under the humid subtropical monsoon climate, the study area has concentrated rainfall and abundant annual precipitation (average annual precipitation > 2000 mm, according to the Anhui Statistical Yearbook, https://www.cnki.net/), providing the critical triggering condition for frequent landslides. Meanwhile, dense vegetation and fragmented micro-topography make most landslides highly concealed, with obscure boundaries and weak spectral differences from the background, posing significant challenges for traditional remote sensing identification.

Integrating the analytical outputs of FR, RF and SHAP methods, this study concludes that the three approaches present both consistent results and distinctive discrepancies in factor importance rankings and landslide sensitivity patterns. The FR method relies on single-factor statistical analysis, which can characterize the macroscopic correlation between conditioning factors and landslide occurrence yet fails to quantify multi-factor interactive effects. As a nonlinear ensemble learning algorithm, RF performs well in determining the importance ranking of dominant influencing factors, whereas it cannot intuitively interpret factor threshold effects and directional driving contributions. In comparison, SHAP is competent in revealing nonlinear threshold characteristics, directional factor contributions and internal interactions among multiple conditioning factors. Accordingly, the divergences across the three methods are not mutually contradictory, but jointly reflect the complex formation mechanisms of landslides from diverse analytical perspectives.

As a technical limitation, it should be noted that due to the steep terrain and dense vegetation in the mountainous areas of Yixian County, field surveys were confined solely to accessible regions, while remote slopes where landslides may develop could not be validated in situ. Consequently, the ground-truth data used in this study are inherently incomplete. In addition, the distribution of clayey materials was not taken into account, despite it being a critical factor governing landslide occurrence. Previous studies have addressed related aspects, including soil classification analysis in quick-clay landslides [73], the influence of clay on landslide initiation and development [74], and the characteristics and occurrence patterns of clay landslides [75]. This issue will be further investigated in future research.

6. Conclusions

(1) To address the challenges of dense vegetation cover and strong landslide concealment in the hilly regions of eastern China, in this study, we utilize GF-2 high-resolution fused imagery to construct a UNet-SE deep learning model incorporating a channel attention mechanism (an SE block), which effectively enhances the extraction of detailed landslide boundary features and improves recognition accuracy.

(2) Comparative recognition results of the four models reveal that UNet-SE achieves optimal performance across all evaluation metrics, with a Precision of 0.911, a Recall of 0.685, an F1-score of 0.782, a Kappa coefficient of 0.730, and an IoU of 0.642. Specifically, its F1-score is 8%, 3% and 5% higher than those of ResNet18-SE, VGG13-SE and the original UNet, respectively. The UNet architecture employs skip connections to fuse deep semantic features and shallow detailed features. Combined with embedded SE blocks, the model delivers enhanced channel feature extraction and landslide boundary localization capacity, further boosting its stability and generalization performance. Cross-regional generalization experiments on creeping landslides in Kecun Town of Yixian County and post-earthquake landslides in Lushan County of Sichuan Province further validate the reliability of the proposed UNet-SE model.

(3) Correlation analysis conducted via FR, RF and SHAP methods indicates the typical predisposing conditions for landslide occurrence in the study area: elevation of 400–500 m, NDVI between 0.4 and 0.5, slope gradient less than 10°, east and northeast slope aspects, river distance of 300–500 m, fault distance of 500–1000 m, and dominant soft sedimentary lithology. Comprehensive analysis confirms that distance from faults, distance from rivers, and elevation were identified as the three dominant driving factors.

In conclusion, the technical method and research results of this study are conducive to coordinating the relationship between geological disaster prevention, ecological environment protection, and socio-economic development. The experience of refined landslide identification based on GF-2 high-resolution images and deep learning can be extended to other hilly regions in eastern China with dense vegetation and frequent concealed landslides, providing a technical reference for disaster prevention and mitigation and sustainable development in similar hilly areas.