Next Article in Journal
A Novel Method for Eliminating Glint in Water-Leaving Radiance from UAV Multispectral Imagery
Previous Article in Journal
Multi-Layer LEO Constellation Optimization Based on D-NSDE Algorithm
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 17
Issue 6
10.3390/rs17060995
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Landslide Identification from Post-Earthquake High-Resolution Remote Sensing Images Based on ResUNet–BFA

by
Zhenyu Zhao
1,2,
Shucheng Tan
2,3,*,
Yiquan Yang
4 and
Qinghua Zhang
1,2
1
Institute of International Rivers and Eco-Security, Yunnan University, Kunming 650091, China
2
Yunnan International Joint Laboratory of Critical Mineral Resources, Kunming 650500, China
3
Department of School of Earth Science, Yunnan University, Kunming 650091, China
4
School of Mathematics and Statistics, Yunnan University, Kunming 650091, China
*
Author to whom correspondence should be addressed.
Remote Sens. 2025, 17(6), 995; https://doi.org/10.3390/rs17060995
Submission received: 13 January 2025 / Revised: 1 March 2025 / Accepted: 7 March 2025 / Published: 12 March 2025
Browse Figures
Versions Notes

Abstract

:
The integration of deep learning and remote sensing for the rapid detection of landslides from high-resolution remote sensing imagery plays a crucial role in post-disaster emergency response. However, the availability of publicly accessible deep learning datasets specifically for landslide detection remains limited, posing challenges for researchers in meeting task requirements. To address this issue, this study develops and releases a deep learning landslide dataset using Google Earth imagery, focusing on the impact zones of the 2008 Wenchuan Ms8.0 earthquake, the 2014 Ludian Ms6.5 earthquake, and the 2017 Jiuzhaigou Ms7.0 earthquake as the research areas. The dataset contains 2727 samples with a spatial resolution of 1.06 m. To enhance landslide recognition, a lightweight boundary-focused attention (BFA) mechanism designed using the Canny operator is adopted. This mechanism improves the model’s ability to emphasize landslide edge features and is integrated with the ResUNet model, forming the ResUNet–BFA architecture for landslide identification. The experimental results indicate that the ResUNet–BFA model outperforms widely used algorithms in extracting landslide boundaries and details, resulting in fewer misclassifications and omissions. Additionally, compared with conventional attention mechanisms, the BFA achieves superior performance, producing recognition results that more closely align with actual labels.

1. Introduction

Seismic landslides represent the predominant form of seismic geohazard associated with large-scale seismic events in mountainous regions and often occur extensively across vast areas. These landslides are generally triggered when earthquake magnitudes exceed Ms4.0, with secondary landslide hazards becoming significantly severe during high-magnitude seismic events, particularly those exceeding Ms7.0 [1,2]. In the case of significant seismic events, thousands of landslides may be triggered, spanning hundreds of thousands of square kilometers [3]. In some major earthquakes, seismic landslides can contribute to more than 50% of the total loss of life and property [4]. Seismic landslides frequently cause additional disasters, such as dam failures, road damage, and bridge collapses [5], which complicate emergency response and field investigations, severely affecting post-disaster rescue operations and damage assessment. With the use of high-resolution remote sensing satellite images as data sources, the development of landslide identification models is crucial for accurately mapping the spatial distribution and components of seismic landslides. This approach is of significant practical value in guiding emergency rescue efforts, disaster assessment, and post-disaster reconstruction after earthquakes [6,7,8,9].
Traditional methods for landslide identification often depend on manual interpretation and expert judgment, which leads to issues such as low efficiency, high subjectivity, and challenges in processing large-scale remote sensing data. With the advent of the 21st century, machine learning techniques have become increasingly popular for large-scale remote sensing image landslide identification because of their high level of automation and adaptability. Commonly used machine learning classifiers in landslide detection include support vector machines [10,11,12,13,14], logistic regression [15,16,17,18], and random forests [18,19,20,21,22,23]. However, these models often exhibit poor generalizability because of the constraints of the models themselves. In 2012, the AlexNet neural network achieved remarkable success in the ImageNet competition, highlighting the benefits of deep learning models for image classification; since then, deep learning methods have been widely adopted across various fields [24,25,26,27,28]. For landslide recognition via remote sensing images, deep learning models learn features related to landslides from extensive datasets of annotated images. These models are then applied to unannotated remote sensing data, enabling efficient and precise pixel-level classification of landslides in remote sensing images. Scholars have explored deep learning model architectures [29,30]. Ghorbanzadeh et al. (2022) introduced an object-based image analysis (OBIA) method, integrating it with ResU-Net, leading to significant improvements in landslide detection accuracy. However, the OBIA method heavily relies on expert knowledge to design an effective set of classification rules; otherwise, its performance will be impacted [31]. Wu et al. (2022) incorporated a dual-attention mechanism into a convolutional neural network for landslide detection from remote sensing imagery and reported that the inclusion of the dual-attention mechanism resulted in higher identification accuracy and improved boundary segmentation compared to single neural networks. However, it excessively increases the model complexity [32]. WANG et al. (2022) proposed a remote sensing landslide target recognition algorithm that integrates improved self-attention and convolutional blocks, maintaining the model’s lightweight nature, but the accuracy in identifying landslide boundaries is less satisfactory [33]. Ji et al. (2020) used an attention-enhanced convolutional neural network for landslide identification from open satellite images and DEM datasets and concluded that the use of an attention mechanism and DEM data significantly increased landslide recognition accuracy, but they did not verify the model’s performance on relatively large-scale datasets [34].
In recent years, scholars in relevant fields have also conducted extensive research on landslide datasets [35]. Li et al. (2023) collected 127 landslide images and 767 non-landslide images from the Three Gorges Reservoir area and used these images along with transfer learning to train various deep learning models [36]. Lei et al. (2022) developed a loess landslide dataset using GF-1 imagery with a spatial resolution of 2 m, covering three counties in southeastern Gansu Province, China: Jingning, Zhuanglang, and Zhangjiachuan Hui Autonomous County. The dataset included 320 preexisting landslides, and a channel fusion convolutional neural network model was utilized for landslide classification [37]. Ullo et al. (2021) developed and created a publicly available target detection dataset via UAV close-up remote sensing imagery, which included 160 samples of both landslide and non-landslide images. They used this dataset to train a Mask R-CNN model for landslide detection through transfer learning [38]. Ji et al. (2020) created and shared a landslide identification dataset for Bijie City with a spatial resolution of 0.8 m, offering it to relevant researchers. They applied a spatial channel attention mechanism to a convolutional neural network for landslide identification, with favorable results [34]. Compared with large-scale deep learning datasets such as ImageNet, the datasets mentioned above have fewer data volumes and landslide samples, with the largest containing only 770 landslide image data. Xu et al. (2024) developed a large-scale multisensor dataset named the CAS Landslide Dataset, comprising 20,865 RGB images sourced from satellite and drone data across nine different regions, aiming to support deep-learning-based landslide detection. Despite the dataset’s outstanding performance in terms of quantity, quality, and diversity, it uniformly crops all original images to 512 × 512 pixels, which is relatively small and makes it difficult to encompass the overall characteristics of large landslides [39]. Specific information about these landslide datasets is summarized in Table 1.
In summary, existing research on deep learning models for landslide identification via high-resolution remote sensing imagery faces several limitations. First, despite advancements in datasets containing tens of thousands of samples, such as the CAS Landslide Dataset, most datasets remain relatively small, often comprising only a few hundred samples. High-quality publicly available remote sensing image datasets for deep learning in landslide identification are still scarce [40,41]. Second, most current deep learning models still exhibit inadequate accuracy in identifying landslide boundaries. While some models have improved precision, they have also significantly increased model complexity and computational demands. Thus, exploring ways to increase the accuracy of landslide boundary recognition while maintaining a lightweight model is worthwhile.
Therefore, this paper employs visual interpretation methods to annotate Google Earth images taken after an earthquake at the pixel level, supplemented by a 3D model created using digital elevation models and optical imagery data to ensure annotation accuracy [42,43,44]. Ultimately, a publicly accessible deep learning landslide dataset with a spatial resolution of 1.06 m was established. After image screening, the dataset contains 2727 high-quality images. Although it has not reached the ten-thousand level, as a specialized dataset for earthquake-triggered landslide scenarios, it can still provide important supplementary resources for related research. Additionally, this paper introduces a lightweight boundary-focused attention (BFA) mechanism based on ResUNet. The BFA focuses on enhancing feature representation of landslide boundaries, enabling the model to more accurately capture transition zones between landslide areas and their surrounding environments. This allows the model to dynamically adjust the importance weights of different locations across the entire image, thereby effectively enhancing long-range dependency modeling capabilities while keeping the model lightweight. This is particularly important for landslides covering large areas. Figure 1 illustrates the research flowchart of this paper.

2. Remote Sensing Imagery Landslide Dataset

2.1. Data Description

Southwest China, located on the eastern margin of the Tibetan Plateau, has complex topography shaped by the continuous uplift of the plateau over the past million years. The region is characterized by a fragile geological environment and frequent high-magnitude seismic activity, leading to a high occurrence of earthquakes. Consequently, earthquake-induced landslides in this area are particularly severe. In recent years, several major earthquakes, including the 2008 Wenchuan Ms8.0 earthquake, the 2013 Lushan Ms7.0 earthquake, the 2014 Ludian Ms6.5 earthquake, the 2017 Jiuzhaigou Ms7.0 earthquake, and the 2022 Luding Ms6.8 earthquake, have struck southwest China, triggering numerous seismic landslides. These landslides are notable for their large scale, high frequency, and extensive impact, causing substantial environmental damage and significant disruptions to human life and property [2,45,46,47,48,49].
This study examines landslides caused by earthquakes in Wenchuan County, Sichuan Province, in 2008; Ludian County, Yunnan Province, in 2014; and Jiuzhaigou County, Sichuan Province, in 2017. Remote sensing images collected after these earthquakes are preprocessed and manually interpreted to construct a sample dataset of landslides for deep learning purposes. Wenchuan County was struck by a major 8.0-magnitude earthquake on 12 May 2008, which triggered about 15,000 landslides, avalanches, and mudslides [50]. Ludian County experienced a 6.5-magnitude earthquake on 3 August 2014, leading to at least 1024 landslides with areas exceeding 100 m2 [51]. Jiuzhaigou County was affected by a 7.0-magnitude earthquake on 8 August 2017, resulting in at least 4800 landslides [52]. As shown in Table 2, these three earthquakes caused extensive casualties and significant economic losses in the affected regions.
To ensure the reliability of the landslide dataset, the study area is selected on the basis of a relatively high density of landslides and minimal cloud cover in the remote sensing images. This selection is based on publicly available coseismic landslide vector data from the aforementioned earthquakes [53,54,55], as well as the quality of the remote sensing images. The distribution of optical remote sensing images is presented in Figure 2, encompassing seven image scenes with a total coverage area of 458 km2. The optical image data utilized in this study were obtained from Google Earth and feature a spatial resolution of 1.06 m. Because multiple sensors are used in Google Earth imagery, the acquisition dates are not uniform. The DEM data employed for mapping the landslide extent vector data were sourced from ALOS (Advanced Land Observation Satellite), with a pixel resolution of 12.5 m × 12.5 m.

2.2. Data Processing

Image data processing involves tasks such as image cropping, landslide labeling, and data augmentation. Coseismic landslides vary in size, and to ensure that the training images effectively cover the typical range of landslide events while retaining sufficient background information, the remotely sensed images are cropped to 1536 × 1536. Because the currently publicly available landslide vector annotations related to the study area are relatively coarse and inconsistent across different datasets, this study uses them as a reference and employs ArcGIS Pro to construct a 3D model via optical remote sensing imagery and DEM data as auxiliary references (e.g., Figure 3).
All the landslide boundaries in the remote sensing images are delineated more precisely via Labelme software version 5.5.0. This includes more complete annotations of the landslide source area, debris flow area, and deposition area (e.g., Figure 4), allowing the deep learning model to learn the overall characteristics of landslides.
The dataset includes challenging samples with intricate landslide details and complex backgrounds (Figure 5a), simpler samples with fewer landslide details and plain backgrounds (Figure 5b), samples obstructed by dense clouds and fog (Figure 5c), samples exhibiting significant image distortion (Figure 5d), and samples with excessive brightness or darkness (Figure 5e,f), among others. Rapid response and rescue are required during earthquake-induced landslides, making it impractical to screen and process remote sensing images comprehensively and meticulously. Therefore, our model needs to be capable of handling images of varying quality. In the data annotation phase, we adopted a targeted image screening strategy to ensure the robustness of the model. Samples with severe image distortions that could significantly mislead model training were excluded, whereas images affected by cloud cover or poor lighting conditions but still retaining usable features for identification were retained. This is because, in real disaster scenarios, such affected images are common. Including them in the training data allows the model to encounter a broader range of image variations, thereby bringing it closer to real-world application environments.
After screening, a total of 2727 images were selected, with a ratio of about 3:1 of positive images (containing landslides) to negative images (not containing landslides). Using Labelme software, a total of 8786 landslides were annotated, covering an area of about 47.33 km2. Size distribution analysis reveals three distinct classes: (1) megascale landslides (>0.1 km2) represent 1% of the inventory, (2) medium-scale landslides (0.01–0.1 km2) constitute 10%, and (3) small-scale landslides (<0.01 km2) account for the predominant 89% of the inventory. The data are shuffled and split randomly, with 80% used for training, 10% for validation, and 10% for testing. The dataset distribution is presented in Table 3.
To enhance the model’s generalization ability, data augmentation methods [56,57,58,59,60], such as random rotation, flipping, and adjustments to brightness or contrast, are employed to expand the training set. The ratio of original to augmented data is 1:3, totaling 8724.

3. Research Methodology

3.1. ResUNet Network

The deep learning model ResUNet merges the strengths of the U-Net architecture and the residual network (ResNet). U-Net was initially developed for medical image segmentation but has also been widely used in remote sensing image processing [61,62,63,64]. The network is composed of an encoder (downsampling path) and a decoder (upsampling path). The encoder reduces the feature map size through convolutional and pooling layers while increasing the number of channels to capture both detailed and global information in the image. The decoder uses inverse convolutional layers and skip connections to restore the resolution of the feature map, combining high-resolution features from the encoder with low-resolution features from the decoder to produce accurate segmentation results [65]. The overall structure of the U-Net network is shown in Figure 6.
However, when developing deeper neural networks, traditional architectures often face the issue of vanishing or exploding gradients, which limits training effectiveness and model performance. To resolve this, residual networks (ResNets) introduce “residual blocks” or “skip connections”, which facilitate more efficient gradient propagation, addressing the challenge of training deeper networks [66,67,68]. These residual connections allow information to flow directly from one layer to the next, bypassing intermediate layers, thus stabilizing gradient flow and making the training of very deep networks feasible [69]. A comparison between the standard network structure and the residual network structure is shown in Figure 7.
By integrating the benefits of the two networks described above, ResUNet incorporates residual connections into each convolutional layer of U-Net, forming the ResBlock [70]. This modification not only maintains U-Net’s strong feature extraction capabilities but also utilizes ResNet’s ability to address gradient issues, thereby improving the model’s learning efficiency and generalization performance [71,72,73]. For landslide detection in remote sensing images, ResUNet offers greater accuracy, enhanced robustness, faster convergence, and better performance in detecting small targets. Consequently, it provides more consistent and accurate results when handling complex and variable datasets, significantly improving the localization and classification of detailed features such as landslides [74,75].

3.2. Lightweight Edge-Focused Attention Mechanism BFA

To overcome the challenges in seismic landslide identification in remote sensing images, particularly in areas with sparse vegetation where landslides and bare ground surfaces share similar spectral characteristics, and because the irregularity of landslide edges often leads to misclassification or omission, we introduce a lightweight edge-enhanced attention mechanism called boundary-focused attention (BFA) [76]. This mechanism effectively integrates the traditional Canny edge detection algorithm [77,78,79] with deep learning models, aiming to increase the network’s sensitivity to landslide boundaries and improve the accuracy of boundary detection [80].
The Canny edge detection algorithm identifies edges by evaluating the local maxima of the image gradient, which is smoothed through the first-order differentiation of a Gaussian function to minimize noise and accurately locate true edges. The key aspect of the algorithm is the use of a quasi-Gaussian function for the smoothing operation:
G ( x , y ) = 1 2 π σ 2 e − x 2 + y 2 2 σ 2
Specifically, a quasi-Gaussian filter is applied for preprocessing, and edges are categorized into strong and weak edges by setting high and low thresholds, retaining weak edges only when they are connected to strong edges. This method effectively reduces noise while preserving edge continuity, making it especially effective for extracting the subtle edges of landslides in remotely sensed images [81].
A compact boundary-focused attention mechanism, BFA, was designed to enhance edge features by preprocessing the input image via the Canny operator. The edge information extracted through the Canny operator is integrated with the original image via channel cascading, resulting in an enhanced input representation. A 1 × 1 convolutional layer dynamically adjusts the feature weights, allowing the model to emphasize relevant edge features. During training, the model learns to identify the most critical edge features for the recognition task, thereby improving the precision of landslide boundary detection and significantly enhancing the accuracy and reliability of landslide identification. The detailed structure of the BFA is illustrated in Figure 8.

3.3. General Structure of ResUNet–BFA

ResUNet–BFA is a deep learning model specifically designed for landslide recognition. Its unique architecture and feature extraction strategy can effectively address the complex challenges in landslide recognition tasks. This model employs ResNet50 as the encoder and utilizes the ResNet50 network pretrained on the ImageNet dataset to gradually extract multiscale features. Owing to its deep convolutional structure and residual connections, ResNet50 has a powerful hierarchical feature extraction ability [82,83,84]. In the lower convolutional layers, the network focuses on capturing low-level features of the image, such as edges and textures. These low-level features can be used to identify the basic contours and surface characteristics of landslide areas, providing fundamental clues for subsequent analysis. As the network deepens, the high-level convolutional layers gradually learn more abstract and semantically rich features, including the overall structure of landslide areas and their relationships with the surrounding terrain. These high-level features are crucial for accurately identifying landslide areas. The model collects feature maps from different levels through the IntermediateLayerGetter, providing rich information for the subsequent feature fusion and decoding processes. At the output of the last residual block of ResNet50, the boundary-focused attention (BFA) mechanism is introduced. This mechanism aims to enhance the ability to capture landslide boundary features and improve the model’s understanding of complex terrain changes. The BFA module first converts the input feature map into a grayscale image, applies the Canny operator for edge detection, and concatenates the extracted edge image with the original feature map at the channel level. A 1 × 1 convolutional layer, followed by a ReLU activation function, dynamically adjusts the feature weights. The decoder utilizes the U-Net architecture, progressively restoring spatial resolution through stepwise upsampling. Additionally, it integrates multiscale features from the encoder, allowing the model to maintain a global perspective while emphasizing local details during landslide recognition. This approach enables precise semantic segmentation of the input image. The overall structure of ResUNet–BFA is depicted in Figure 9.
The BFA module is placed after the deepest feature map of the encoder, specifically at the output of the final residual block of ResNet50, to optimize the combination of semantic information and boundary details. This arrangement helps recover spatial information lost due to the increased depth of the network, improving the quality of the decoder input while minimizing computational demands. By introducing the BFA module at this point, the model enhances its ability to capture boundary features while preserving high-level semantic data, which is crucial for tasks such as landslide detection that require precise boundary delineation. This approach not only takes full advantage of the feature extraction power of the pretrained model, reducing training time, but also improves the accuracy of landslide area localization by focusing on boundary features. The multiscale feature fusion strategy, along with an efficient spatial resolution recovery mechanism, ensures that the model performs high-precision and robust landslide detection in complex scenarios.

4. Experiment and Analysis

4.1. Model Training Setup

The experiments are carried out via the PyTorch 2.3.0 open-source deep learning framework in the AutoDL cloud environment (https://www.autodl.com/home, accessed on 15 January 2025), with the detailed experimental environment settings provided in Table 4.
In the training process, all the models are trained with the same set of hyperparameters to ensure a fair comparison. The image size refers to the initial input image dimensions. Although the dataset produced and made publicly available in this paper has an image size of 1536 × 1536, owing to hardware constraints, the input images are uniformly resized to 224 × 224 during training.
The research uses the sum of the focal loss and Dice loss as the loss function and selects the optimal model weights on the validation set on the basis of the loss. Among them, focal loss is an improvement over cross-entropy loss. When there is class imbalance, the ordinary cross-entropy loss is dominated by many easily classifiable samples, resulting in insufficient learning of the model for minority class samples. The focal loss adjusts the weights of different samples by introducing a modulation factor 1 − p t γ (where p t is the probability that the model predicts a sample to be of the correct class, and where γ is a hyperparameter). It reduces the weight of easily classifiable samples and increases the weight of difficult-to-classify samples. In landslide detection, if landslide samples are the minority class, the focal loss allows the model to pay more attention to their features, alleviating the problem of class imbalance. The Dice loss is calculated on the basis of the Dice coefficient, which measures the similarity between the predicted result and the true label by considering true positives ( T P ), false positives ( F P ), and false negatives ( F N ). Its formula is as follows:
D i c e _ l o s s = 1 − ( 1 + β 2 ) * t p + s m o o t h ( 1 + β 2 ) * t p + β 2 * f n + f p + s m o o t h
where β and s m o o t h are hyperparameters. The Dice loss is more sensitive to prediction errors in the boundary regions. Since the boundary regions often have significant differences between the predicted results and the true labels, this leads to a decrease in true positives ( T P ) and an increase in false negatives ( F N ) and false positives ( F P ), thereby increasing the value of Dice loss. To reduce the loss, the model improves the prediction accuracy in the boundary regions. In landslide detection, Dice loss can prompt the model to more accurately delineate the boundaries of landslide areas.
The initial learning rate is set to 0.0001, and an adaptive learning rate adjustment strategy is adopted. It combines the Warmup phase and linear learning rate decay (with a decay exponent of 0.9) to dynamically adjust the learning rate, which helps the model escape from local optimal solutions. The Adam adaptive learning rate optimization algorithm is used to avoid premature convergence. An early stopping mechanism is applied, with patience set to 10 to prevent overfitting. The specific settings of the hyperparameters are shown in Table 5.

4.2. Evaluation Metrics

The confusion matrix is a widely used tool for evaluating the performance of classification models, particularly in machine learning and statistical analysis. It is a two-dimensional table that compares predicted labels with actual labels. The structure of the confusion matrix is presented in Table 6. In a binary classification problem, category A represents the positive class, whereas category B represents the negative class. Correct predictions are categorized as true, whereas incorrect predictions are classified as false, resulting in four fundamental components: true positive (TP), where the model correctly identifies positive samples as class A; true negative (TN), where the model correctly identifies negative samples as class B; false positive (FP), where the model incorrectly classifies negative samples as class A; and false negative (FN), where the model incorrectly classifies positive samples as class B.
The confusion matrix can be utilized to compute various secondary classification metrics, including the recall, precision, F1 score, and mean intersection over union (MIoU). Recall indicates the model’s ability to correctly identify positive samples, whereas precision measures the accuracy of the model in predicting positive samples. The F1 score combines both precision and recall into a single metric, offering a comprehensive evaluation of the model’s performance. MIoU is used to quantify the extent of overlap between the predicted regions and the true regions, assessing the accuracy of the model’s predictions in terms of spatial alignment. The formula is shown below:
P r e c i s i o n = T P T P + F P
R e c a l l = T P T P + F N
F 1 = 2 × P r e c i s i o n × R e c a l l P r e c i s i o n + R e c a l l
M I O U = 1 | C | ∑ c ∈ C T P ( c ) T P ( c ) + F P ( c ) + F N ( c )
where C represents the set of categories and TP(c) refers to the true positives for category c, which is the count of pixels correctly predicted as category c. FP(c) represents the false positives for category c, which is the number of pixels incorrectly classified as category c. FN(c) indicates the false negatives for category c, referring to the number of pixels that should have been predicted as category c but were not correctly identified.
Moreover, the mean pixel accuracy (MPA) is the average of the pixel accuracies for all categories. For each category, the pixel accuracy is calculated as the ratio of correctly predicted pixels to the total number of pixels for that category. The formula for this calculation is as follows:
A c c u r a c y = T P + T N T P + T N + F P + F N

4.3. Analysis of Results

4.3.1. Comparison of the Recognition Results Between the Proposed Model and Other Models

In this study, the commonly used semantic segmentation networks Unet [85], MultiResUNet [86], DeepLabV3+ [87], TransUNet [88], and ResUNet [70] serve as control models, with hyperparameter settings during training matching those used for the ResUNet–BFA model proposed here. The experimental results show that the training and testing times of the proposed ResUNet–BFA model are moderate, and the training time is reduced compared with that of ResUNet, demonstrating that the improvements made to the model not only increase its computational complexity but also optimize the feature extraction and processing workflow, enabling the model to learn and infer more efficiently. The specific durations are shown in Table 7.
The experimental results indicate that the ResUNet–BFA model outperforms the other models in terms of precision, F1 score, MPA, and MIoU, with improvements of 2.34%, 2.06%, 1.28%, and 1.44%, respectively, over those of the second-place model. The performance metrics of each model for landslide detection on the test set are provided in Table 8.
When deep-learning-based semantic segmentation models are applied for landslide detection in remote sensing images, the substantial size of individual images necessitates significant computational resources for direct processing. Consequently, image chunking is frequently employed for recognition. During this process, the inherent randomness and complexity of the landslide background contribute to variations in both the shape and background of the landslide within the segmented images. Certain image segments may contain limited information, where only small portions of landslides are visible, and the background remains relatively simple. Conversely, other segments may provide more comprehensive details, including complex landslide structures or backgrounds with multiple feature types. As demonstrated in Figure 10, the enhanced ResUNet–BFA model proposed in this study effectively identifies landslide boundaries in both simple and complex cases. Compared with conventional semantic segmentation models, it results in a reduced recognition error region and demonstrates an improved ability to detect finer landslide details.

4.3.2. Comparison of the Recognition Results Between the Attention Mechanism in This Paper and Other Attention Mechanisms

Figure 11 presents the results of various attention mechanisms, including the classical squeeze-and-excitation (SE) channel attention mechanism [89], spatial attention mechanism (SAM) [90], convolutional block attention module (CBAM) [90], and our proposed lightweight boundary-focused attention (BFA) mechanism, which are applied to train and recognize simple landslide samples in the ResUNet model. The figure also shows the loss curves for both training and validation of the BFA method. The results clearly show that the training and validation losses for the BFA approach exhibit high stability with minimal fluctuations throughout the training process, highlighting its superior generalizability and model stability.
Figure 12 displays the heatmaps generated from the predictions of different attention mechanisms via the Grad-CAM visualization technique. Grad-CAM helps identify the most influential regions of the feature map by tracking gradients during the network’s backpropagation process, which are then used to create the heatmap [91,92]. This visualization not only highlights the role of the attention mechanism in the model’s decision-making but also offers a clear view of how much focus each attention module places on the landslide regions in the feature map, with red indicating areas of high attention.
The different attention mechanisms impact landslide detection in different ways, particularly in simple samples where focus areas tend to be diverse and disorganized. SE demonstrates improved accuracy in targeting landslide regions; however, it is affected by focusing bias and lacks precision in defining boundary details. In contrast, the BFA not only integrates multiscale information but also focuses on the fine details of the landslide boundary, leading to more accurate identification of the landslide area. In the heatmap, the BFA demonstrates a superior ability to capture the contours of the landslide area compared with the other attention mechanisms, further confirming its outstanding performance in landslide identification.
According to Table 9, the BFA outperforms SE, CBAM, and SAM in the key metrics of the F1 score, MPA, and MIoU, with values of 78.18, 88.03, and 80.09, respectively. These results represent minimum improvements of 1.91%, 0.17%, and 1.37%, respectively, over those of the other models.

4.3.3. Validation of Cross-Regional Seismic Landslide Identification

To assess the effectiveness of the seismic landslide dataset and the ResUNet–BFA model, external validation was performed using a landslide event triggered by a magnitude-6.8 earthquake in Luding County, Sichuan Province, in 2022. The Luding region, located in southwest China, is characterized by complex terrain and dense vegetation, exhibiting geographical and environmental similarities to the training dataset while also presenting certain differences. This scenario provides a suitable testing ground to evaluate the model’s performance on previously unseen data.
For validation, high-resolution remote sensing imagery from Google Earth was utilized to analyze an area significantly impacted by landslides following the Luding earthquake (Figure 13a). The landslide detection results (Figure 13b) were compared with manually annotated landslide masks derived from vector annotations provided by Zhao et al. [53], yielding a Dice coefficient of 0.55, with an MIoU of 0.64. These findings indicate that the ResUNet–BFA model has strong adaptability and generalization capabilities in regions with complex topography and high vegetation coverage, such as southwest China. This adaptability enhances the model’s potential for future applications in landslide monitoring and emergency response efforts in earthquake-affected areas.
In addition, to further verify the quality and robustness of our dataset, we compared it with previously published datasets. We took the selected remote sensing images of the Luding earthquake as the validation set and used our created dataset, the Bijie Landslide Dataset [34], the Sichuan and Surrounding Areas Landslide Dataset [93], and the Landslide4Sense Dataset [94] as the training sets. To ensure consistency during training, we standardized the images to a resolution of 224 × 224 pixels. Table 10 shows our experimental results. Compared with the other three publicly available datasets, our dataset performs better in terms of recall, F1 score, and MIoU. These results highlight its outstanding ability to accurately identify landslides within the designated task area.

5. Discussion

This study proposes the ResUNet–BFA model for identifying earthquake-induced landslides and trains and validates the performance of the ResUNet–BFA model in a custom-made landslide dataset based on three regions: Wenchuan, Jiuzhaigou, and Ludian. The model proposed in this paper is directly applied to the 2022 Luding earthquake, demonstrating the feasibility of this research in wide-area earthquake-induced landslide applications. On the basis of this study, several aspects can be improved and enhanced:
  • The study area selected for this research comprises mostly mountainous regions with high vegetation coverage. Before the earthquake, most of the landslides were covered by vegetation, but after the earthquake, the vegetation on the slope surface was damaged, and many slopes changed from forested areas to bare land, which can be relatively easily identified through RGB remote sensing images. However, for those areas that were originally bare land and had no vegetation coverage before the earthquake, the surface coverage changes were minimal after the earthquake, and the distinction between landslides and the surrounding environment was not obvious, making it difficult to distinguish them via optical images. Moreover, this study has high requirements for the quality of optical images. When post-earthquake images are covered by thick clouds or snow, this method is not effective. This is a relatively serious problem for emergency response to landslide disasters, as a certain revisit period is needed to obtain high-quality images. In future research, efforts will be made to study areas with less vegetation coverage and integrate synthetic aperture radar data.
  • The annotation of landslide samples is highly important. In the absence of onsite investigations, even in high-resolution remote sensing images, interpretation experts may misidentify or overlook mixed landslide pixels and small landslide areas. Therefore, both the training set and the test set may include FPs and FNs, that is, labels that incorrectly represent non-landslides as landslides and features that are missed and not marked as landslides. This uncertainty affects all the experiments but does not systematically alter the findings of this paper.
  • This study conducts analyses on satellite data with a unified pixel size. Although this simplifies processing and ensures data consistency, application limitations exist when integrating multisource resolution data (such as Landsat and MODIS) and overlaying them with heterogeneous raster/vector data. To address this issue, resampling techniques (such as bilinear interpolation cubic convolution) combined with a maximum likelihood classifier can be used to unify the data resolution and ensure the compatibility of multisource data [95]. In addition, for landslide identification, integrating multidimensional influencing factors (such as DEMs, stratigraphic attributes, distance from rivers, distribution of fold/fault zones, lithology, slope, rainfall, vegetation indices, etc.) and dynamic characteristic data (such as synthetic aperture radar interferometry data) can overcome the accuracy bottleneck of RGB optical images. Therefore, advanced image fusion methods (such as band synthesis and multisource heterogeneous data fusion strategies) and semantic feature extraction models can be adopted to construct a comprehensive analysis framework that considers geographical environmental features and kinematic laws, thereby improving detection reliability [96].
  • The model proposed in this article can be executed more intensively via other advanced block-based or modular strategies. This includes searching for predictions via a block-combined network structure [97], mapping via a hybridized modular structure [98], denoising via 3D filtering and block matching [99], and employing block-based CNN models for image detection [100].
  • This article focuses on the emergency identification of post-occurrence landslides, aiming to assist rescue operations through the application of deep learning technologies. However, adhering to the principle that “prevention is better than cure”, future explorations could involve leveraging advanced deep learning models such as the transformer architecture [101] and Bayesian neural networks (BNNs) [102], along with ensemble learning strategies [103], to achieve landslide prediction and uncertainty analysis. This approach would more effectively contribute to disaster prevention and reduce reliance on emergency response.

6. Conclusions

This paper presents and publicly releases a high-quality landslide dataset comprising 2727 post-earthquake high-resolution remote sensing images. This dataset’s large-scale spatial distribution characteristics and annotation standards can provide foundational data support for subsequent research. The dataset is designed to drive innovation in practical applications and improve data quality, providing researchers with a valuable resource for advancing landslide identification studies. Addressing the challenges posed by earthquake-induced landslides contributes to safeguarding human lives and property, thereby strengthening disaster mitigation efforts.
The ResUNet–BFA model is proposed for landslide detection; it uses ResNet50 as the encoder and incorporates a self-designed boundary-focused attention (BFA) mechanism at the output of the final residual block. The BFA mechanism integrates Canny edge detection results with the original feature map and dynamically adjusts feature weights, improving the model’s ability to capture complex terrain variations. The decoder follows the U-Net architecture, progressively restoring spatial resolution through upsampling while merging multiscale features from the encoder to preserve both the global context and local details.
The ResUNet–BFA model maintains a lightweight design that ensures computational efficiency while significantly enhancing recognition accuracy. The results closely align with landslide boundaries identified by human experts, making the model highly applicable for practical landslide identification. It can thus serve as a powerful tool for detailed hazard assessments and support decision-making in post-disaster emergency response. Future research will focus on expanding the dataset scale, improving the data quality, and further enhancing the model’s capabilities.

Author Contributions

Conceptualization, Z.Z. and S.T.; methodology, Z.Z., Y.Y. and Q.Z.; investigation, Z.Z.; writing—original draft preparation, Z.Z.; writing—review and editing, S.T., Y.Y. and Q.Z.; visualization, Z.Z.; supervision, S.T., Y.Y. and Q.Z.; project administration, S.T.; funding acquisition, S.T. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Yunnan Fundamental Research Projects under Grant No. 202301BF070001-020; the Yunnan Key Research and Development Plan Program under Grant No. 202303AP140020; and the Science and Technology Innovation Team Program of Yunnan Province Education Department under Grant No. CY22624109.

Data Availability Statement

The related code and data are available at https://github.com/zhao952711/landslides_identification.git (accessed on 1 March 2025).

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Harp, E.L.; Keefer, D.K.; Sato, H.P.; Yagi, H. Landslide inventories: The essential part of seismic landslide hazard analyses. Eng. Geol. 2011, 122, 9–21. [Google Scholar] [CrossRef]
  2. Dai, L.; Fan, X.; Wang, X.; Fang, C.; Zou, C.; Tang, X.; Wei, Z.; Xia, M.; Wang, D.; Xu, Q. Coseismic landslides triggered by the 2022 Luding Ms6. 8 earthquake, China. Landslides 2023, 20, 1277–1292. [Google Scholar] [CrossRef]
  3. Jibson, R.W.; Harp, E.L.; Michael, J.A. A method for producing digital probabilistic seismic landslide hazard maps. Eng. Geol. 2000, 58, 271–289. [Google Scholar] [CrossRef]
  4. Wartman, J.; Dunham, L.; Tiwari, B.; Pradel, D. Landslides in eastern Honshu induced by the 2011 Tohoku earthquake. Bull. Seismol. Soc. Am. 2013, 103, 1503–1521. [Google Scholar] [CrossRef]
  5. Lin, C.H.; Kumagai, H.; Ando, M.; Shin, T.C. Detection of landslides and submarine slumps using broadband seismic networks. Geophys. Res. Lett. 2010, 37. [Google Scholar] [CrossRef]
  6. Fu, R.; He, J.; Liu, G.; Li, W.; Mao, J.; He, M.; Lin, Y. Fast seismic landslide detection based on improved mask R-CNN. Rem. Sen. 2022, 14, 3928. [Google Scholar] [CrossRef]
  7. Liu, R.; Li, L.; Pirasteh, S.; Lai, Z.; Yang, X.; Shahabi, H. The performance quality of LR, SVM, and RF for earthquake-induced landslides susceptibility mapping incorporating remote sensing imagery. Arab. J. Geosci. 2021, 14, 259. [Google Scholar] [CrossRef]
  8. Guo, X.; Fu, B.; Du, J.; Shi, P.; Li, J.; Li, Z.; Du, J.; Chen, Q.; Fu, H. Monitoring and assessment for the susceptibility of landslide changes after the 2017 ms 7.0 Jiuzhaigou earthquake using the remote sensing technology. Front. Earth Sc. 2021, 9, 633117. [Google Scholar] [CrossRef]
  9. Casagli, N.; Intrieri, E.; Tofani, V.; Gigli, G.; Raspini, F. Landslide detection, monitoring and prediction with remote-sensing techniques. Nat. Rev. Earth Env. 2023, 4, 51–64. [Google Scholar] [CrossRef]
  10. Suthaharan, S. Support vector machine. In Machine Learning Models and Algorithms for Big Data Classification. Integrated Series in Information Systems; Springer: Boston, MA, USA, 2016; Volume 36, pp. 207–235. [Google Scholar]
  11. Kamran, K.V.; Feizizadeh, B.; Khorrami, B.; Ebadi, Y. A comparative approach of support vector machine kernel functions for GIS-based landslide susceptibility mapping. Appl. Geomat. 2021, 13, 837–851. [Google Scholar] [CrossRef]
  12. Pham, B.T.; Prakash, I.; Khosravi, K.; Chapi, K.; Trinh, P.T.; Ngo, T.Q.; Hosseini, S.V.; Bui, D.T. A comparison of Support Vector Machines and Bayesian algorithms for landslide susceptibility modelling. Geocarto Int. 2019, 34, 1385–1407. [Google Scholar] [CrossRef]
  13. Yao, X.; Tham, L.G.; Dai, F.C. Landslide susceptibility mapping based on support vector machine: A case study on natural slopes of Hong Kong, China. Geomorphology 2008, 101, 572–582. [Google Scholar] [CrossRef]
  14. Pourghasemi, H.R.; Jirandeh, A.G.; Pradhan, B.; Xu, C.; Gokceoglu, C. Landslide susceptibility mapping using support vector machine and GIS at the Golestan Province, Iran. J. Earth Syst. Sci. 2013, 122, 349–369. [Google Scholar] [CrossRef]
  15. Budimir, M.E.A.; Atkinson, P.M.; Lewis, H.G. A systematic review of landslide probability mapping using logistic regression. Landslides 2015, 12, 419–436. [Google Scholar] [CrossRef]
  16. Sujatha, E.R.; Sridhar, V. Landslide susceptibility analysis: A logistic regression model case study in Coonoor, India. Hydrology 2021, 8, 41. [Google Scholar] [CrossRef]
  17. Sun, D.; Xu, J.; Wen, H.; Wang, D. Assessment of landslide susceptibility mapping based on Bayesian hyperparameter optimization: A comparison between logistic regression and random forest. Eng. Geol. 2021, 281, 105972. [Google Scholar] [CrossRef]
  18. Chowdhury, M.S.; Rahman, M.N.; Sheikh, M.S.; Sayeid, M.A.; Mahmud, K.H.; Hafsa, B. GIS-based landslide susceptibility mapping using logistic regression, random forest and decision and regression tree models in Chattogram District, Bangladesh. Heliyon 2024, 10, e23424. [Google Scholar] [CrossRef]
  19. Krkač, M.; Špoljarić, D.; Bernat, S.; Arbanas, S.M. Method for prediction of landslide movements based on random forests. Landslides 2017, 14, 947–960. [Google Scholar] [CrossRef]
  20. Zhang, W.; He, Y.; Wang, L.; Liu, S.; Meng, X. Landslide Susceptibility mapping using random forest and extreme gradient boosting: A case study of Fengjie, Chongqing. Geol. J. 2023, 58, 2372–2387. [Google Scholar] [CrossRef]
  21. Wang, S.; Zhuang, J.; Zheng, J.; Fan, H.; Kong, J.; Zhan, J. Application of Bayesian hyperparameter optimized random forest and XGBoost model for landslide susceptibility mapping. Front. Earth Sc. 2021, 9, 712240. [Google Scholar] [CrossRef]
  22. Taalab, K.; Cheng, T.; Zhang, Y. Mapping landslide susceptibility and types using Random Forest. Big Earth Data 2018, 2, 159–178. [Google Scholar] [CrossRef]
  23. Chen, F.; Yu, B.; Li, B. A practical trial of landslide detection from single-temporal Landsat8 images using contour-based proposals and random forest: A case study of national Nepal. Landslides 2018, 15, 453–464. [Google Scholar] [CrossRef]
  24. Mohan, A.; Singh, A.K.; Kumar, B.; Dwivedi, R. Review on remote sensing methods for landslide detection using machine and deep learning. Trans. Emerg. Telecommun. Technol. 2021, 32, e3998. [Google Scholar] [CrossRef]
  25. Schoenfeldt, E.; Winocur, D.; Pánek, T.; Korup, O. Deep learning reveals one of Earth’s largest landslide terrain in Patagonia. Earth Planet. Sc. Lett. 2022, 593, 117642. [Google Scholar] [CrossRef]
  26. Ge, Y.; Liu, G.; Tang, H.; Zhao, B.; Xiong, C. Comparative analysis of five convolutional neural networks for landslide susceptibility assessment. Bull. Eng. Geol. Environ. 2023, 82, 377. [Google Scholar] [CrossRef]
  27. Aslam, B.; Zafar, A.; Khalil, U. Comparative analysis of multiple conventional neural networks for landslide susceptibility mapping. Nat. Hazards 2023, 115, 673–707. [Google Scholar] [CrossRef]
  28. Li, C.; Yi, B.; Gao, P.; Li, H.; Sun, J.; Chen, X.; Zhong, C. Valuable clues for DCNN-based landslide detection from a comparative assessment in the Wenchuan earthquake area. Sensors 2021, 21, 5191. [Google Scholar] [CrossRef]
  29. Chen, H.; He, Y.; Zhang, L.; Yao, S.; Yang, W.; Fang, Y.; Liu, Y.; Gao, B. A landslide extraction method of channel attention mechanism U-Net network based on Sentinel-2A remote sensing images. Int. J. Digit. Earth 2023, 16, 552–577. [Google Scholar] [CrossRef]
  30. Meena, S.R.; Soares, L.P.; Grohmann, C.H.; Van Westen, C.; Bhuyan, K.; Singh, R.P.; Floris, M.; Catani, F. Landslide detection in the Himalayas using machine learning algorithms and U-Net. Landslides 2022, 19, 1209–1229. [Google Scholar] [CrossRef]
  31. Ghorbanzadeh, O.; Shahabi, H.; Crivellari, A.; Homayouni, S.; Blaschke, T.; Ghamisi, P. Landslide detection using deep learning and object-based image analysis. Landslides 2022, 19, 929–939. [Google Scholar] [CrossRef]
  32. Wu, Q.; Zhou, C.; Huang, F.; Yao, C. Optimization of the landslide identification method based on a dual attention mechanism. Bull. Geol. Sci. Technol. 2022, 41, 246–253. [Google Scholar]
  33. Wang, Y.; Zhang, P.; Sun, K.; Sun, X.; Liu, L. Remote sensing landslide target recognition based on attention fusion. Chin. J. Liq. Cryst. Disp. 2022, 37, 1498–1506. [Google Scholar] [CrossRef]
  34. Ji, S.; Yu, D.; Shen, C.; Li, W.; Xu, Q. Landslide detection from an open satellite imagery and digital elevation model dataset using attention boosted convolutional neural networks. Landslides 2020, 17, 1337–1352. [Google Scholar] [CrossRef]
  35. Sreelakshmi, S.; Chandra, S.S.V. Visual saliency-based landslide identification using super-resolution remote sensing data. Results Eng. 2024, 21, 101656. [Google Scholar] [CrossRef]
  36. Li, Y.; Wang, P.; Feng, Q.; Ji, X.; Jin, D.; Gong, J. Landslide detection based on shipborne images and deep learning models: A case study in the Three Gorges Reservoir Area in China. Landslides 2023, 20, 547–558. [Google Scholar] [CrossRef]
  37. Lei, F.; Wang, Y.; Longpeng, I. Landslide identification using remote sensing images and DEM based on convolutional neural network: A case study of loess landslide. Rem. Sens. Nat. Resour. 2022, 34, 224–230. [Google Scholar]
  38. Ullo, S.L.; Mohan, A.; Sebastianelli, A.; Ahamed, S.E.; Kumar, B.; Dwivedi, R.; Sinha, G.R. A new mask R-CNN-based method for improved landslide detection. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 14, 3799–3810. [Google Scholar] [CrossRef]
  39. Xu, Y.; Ouyang, C.; Xu, Q.; Wang, D.; Zhao, B.; Luo, Y. CAS landslide dataset: A large-scale and multisensor dataset for deep learning-based landslide detection. Sci. Data 2024, 11, 12. [Google Scholar] [CrossRef]
  40. Zhang, X.; Yu, W.; Pun, M.O.; Shi, W. Cross-domain landslide mapping from large-scale remote sensing images using prototype-guided domain-aware progressive representation learning. ISPRS J. Photogramm. 2023, 197, 1–17. [Google Scholar] [CrossRef]
  41. Liu, J.; Du, Y.; Han, C.; Xie, F. Analysis of magnitude–frequency distribution of earthquakes in the Sichuan basin, southwest China. Seismol. Res. Lett. 2024, 95, 3482–3493. [Google Scholar] [CrossRef]
  42. Shahri, A.A.; Shan, C.; Zäll, E.; Larsson, S. Spatial distribution modeling of subsurface bedrock using a developed automated intelligence deep learning procedure: A case study in Sweden. J. Rock Mech. Geotech. Eng. 2021, 13, 1300–1310. [Google Scholar] [CrossRef]
  43. Li, S.; Xiong, L.; Tang, G.; Strobl, J. Deep learning-based approach for landform classification from integrated data sources of digital elevation model and imagery. Geomorphology 2020, 354, 107045. [Google Scholar] [CrossRef]
  44. Zandsalimi, Z.; Barbosa, S.A.; Alemazkoor, N.; Goodall, J.L.; Shafiee-Jood, M. Deep learning-based downscaling of global digital elevation models for enhanced urban flood modeling. J. Hydrol. 2025, 653, 132687. [Google Scholar] [CrossRef]
  45. Chen, M.; Tang, C.; Li, M.; Xiong, J.; Luo, Y.; Shi, Q.; Zhang, X.; Tie, Y.; Feng, Q. Changes of surface recovery at coseismic landslides and their driving factors in the Wenchuan earthquake-affected area. Catena 2022, 210, 105871. [Google Scholar] [CrossRef]
  46. Li, G.; Zang, M.; Qi, S.; Bo, J.; Yang, G.; Liu, T. An infinite slope model considering unloading joints for spatial evaluation of coseismic landslide hazards triggered by a reverse seismogenic fault: A case study of the 2013 lushan earthquake. Sustainability 2023, 16, 138. [Google Scholar] [CrossRef]
  47. He, X.; Xu, C.; Qi, W.; Huang, Y. Contrasting landslides distribution patterns and seismic rupture processes of 2014 Jinggu and Ludian earthquakes, China. Sci. Rep. 2024, 14, 28470. [Google Scholar] [CrossRef]
  48. Chang, M.; Cui, P.; Xu, L.; Zhou, Y. The spatial distribution characteristics of coseismic landslides triggered by the Ms7. 0 Lushan earthquake and Ms7. 0 Jiuzhaigou earthquake in southwest China. Environ. Sci. Pollut. Res. 2021, 28, 20549–20569. [Google Scholar] [CrossRef]
  49. Su, F.; Liu, H.; Han, Y. The extraction of mountain hazard induced by Wenchuan earthquake and analysis of its distributing characteristic. J. Remote Sens. 2008, 12, 956–963. [Google Scholar]
  50. Fu, X.; Sheng, Q.; Li, G.; Zhang, Z.; Zhou, Y.; Du, Y. Analysis of landslide stability under seismic action and subsequent rainfall: A case study on the Ganjiazhai giant landslide along the Zhaotong-Qiaojia road during the 2014 Ludian earthquake, Yunnan, China. Bull. Eng. Geol. Environ. 2020, 79, 5229–5248. [Google Scholar] [CrossRef]
  51. Li, Q.; Zhang, J.F.; Luo, Y.; Jiao, Q.S. Recognition of earthquake-induced landslide and spatial distribution patterns triggered by the Jiuzhaigou earthquake in August 8, 2017. J. Remote Sens. 2019, 23, 785–795. [Google Scholar] [CrossRef]
  52. Nava, L.; Monserrat, O.; Catani, F. Improving landslide detection on SAR data through deep learning. IEEE Geosci. Remote Sens. Lett. 2021, 19, 4020405. [Google Scholar] [CrossRef]
  53. Zhao, T.; Zhang, S.C.; He, X.N.; Xue, B.W.; Zha, F.K. Improved DeepLabV3+ model for landslide identification in high-resolution remote sensing images after earthquakes. Natl. Remote Sens. Bull. 2024, 28, 2293–2305. [Google Scholar] [CrossRef]
  54. Xu, C.; Xu, X.W.; Shen, L.L.; Dou, S.; Wu, S.E.; Tian, Y.Y.; Li, X. Cataloging landslides triggered by the 2014 Ludian MS6.5 earthquake and their implications for some seismic parameters. Seismol. Geol. 2014, 36, 1186–1203. [Google Scholar] [CrossRef]
  55. Xu, C. Distribution Map of Landslide Data from the Ms7.0 Earthquake in Jiuzhaigou County, Sichuan Province on August 8, 2017; National Cryosphere Desert Data Center: Beijing, China, 2022; Available online: http://www.ncdc.ac.cn (accessed on 1 March 2025). [CrossRef]
  56. Wei, R.; Ye, C.; Sui, T.; Zhang, H.; Ge, Y.; Li, Y. A feature enhancement framework for landslide detection. Int. J. Appl. Earth Obs. 2023, 124, 103521. [Google Scholar] [CrossRef]
  57. Li, P.; Wang, Y.; Xu, G.; Wang, L. LandslideCL: Towards robust landslide analysis guided by contrastive learning. Landslides 2023, 20, 461–474. [Google Scholar] [CrossRef]
  58. Lv, N.; Ma, H.; Chen, C.; Pei, Q.; Zhou, Y.; Xiao, F.; Li, J. Remote sensing data augmentation through adversarial training. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 14, 9318–9333. [Google Scholar] [CrossRef]
  59. Yu, X.; Wu, X.; Luo, C.; Ren, P. Deep learning in remote sensing scene classification: A data augmentation enhanced convolutional neural network framework. GIScience Remote Sens. 2017, 54, 741–758. [Google Scholar] [CrossRef]
  60. Kaushal, A.; Gupta, A.K.; Sehgal, V.K. A semantic segmentation framework with UNet-pyramid for landslide prediction using remote sensing data. Sci. Rep. 2024, 14, 30071. [Google Scholar] [CrossRef]
  61. Tan, C.; Chen, T.; Liu, J.; Deng, X.; Wang, H.; Ma, J. Building extraction from Unmanned Aerial Vehicle (UAV) data in a landslide-affected scattered mountainous area based on res-unet. Sustainability 2024, 16, 9791. [Google Scholar] [CrossRef]
  62. Niu, C.; Gao, O.; Lu, W.; Liu, W.; Lai, T. Reg-SA–UNet++: A lightweight landslide detection network based on single-temporal images captured postlandslide. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2022, 15, 9746–9759. [Google Scholar] [CrossRef]
  63. Li, H.; He, Y.; Xu, Q.; Deng, J.; Li, W.; Wei, Y. Detection and segmentation of loess landslides via satellite images: A two-phase framework. Landslides 2022, 19, 673–686. [Google Scholar] [CrossRef]
  64. Cao, H.; Wang, Y.; Chen, J.; Jiang, D.; Zhang, X.; Tian, Q.; Wang, M. Swin-unet: Unet-like pure transformer for medical image segmentation. In Proceedings of the European Conference on Computer Vision, Tel Aviv, Israel, 23–27 October 2022; Karlinsky, L., Michaeli, T., Nishino, K., Eds.; Springer: Cham, Switzerland, 2022; pp. 205–218. [Google Scholar]
  65. Targ, S.; Almeida, D.; Lyman, K. Resnet in resnet: Generalizing residual architectures. arXiv 2016, arXiv:1603.08029. [Google Scholar] [CrossRef]
  66. Hacıefendioğlu, K.; Demir, G.; Başağa, H.B. Landslide detection using visualization techniques for deep convolutional neural network models. Nat. Hazards 2021, 109, 329–350. [Google Scholar] [CrossRef]
  67. Wu, Z.; Shen, C.; Van Den Hengel, A. Wider or deeper: Revisiting the resnet model for visual recognition. Pattern Recogn. 2019, 90, 119–133. [Google Scholar] [CrossRef]
  68. Liu, T.; Chen, T.; Niu, R.; Plaza, A. Landslide detection mapping employing CNN, ResNet, and DenseNet in the three gorges reservoir, China. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 14, 11417–11428. [Google Scholar] [CrossRef]
  69. Song, Y.; Zou, Y.; Li, Y.; He, Y.; Wu, W.; Niu, R.; Xu, S. Enhancing landslide detection with SBConv-Optimized U-Net architecture based on multisource remote sensing data. Land 2024, 13, 835. [Google Scholar] [CrossRef]
  70. Diakogiannis, F.I.; Waldner, F.; Caccetta, P.; Wu, C. ResUNet-a: A deep learning framework for semantic segmentation of remotely sensed data. ISPRS J. Photogramm. Remote Sens. 2020, 162, 94–114. [Google Scholar] [CrossRef]
  71. Yao, G.; Zhou, W.; Liu, M.; Xu, Q.; Wang, H.; Li, J.; Ju, Y. An empirical study of the convolution neural networks based detection on object with ambiguous boundary in remote sensing imagery—A case of potential loess landslide. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 15, 323–338. [Google Scholar] [CrossRef]
  72. Wen, Y.; Ma, X.; Zhang, X.; Pun, M.O. GCD-DDPM: A generative change detection model based on difference-feature guided DDPM. IEEE T. Geosci. Remote 2024, 62, 5404416. [Google Scholar] [CrossRef]
  73. Chen, L.; Zhang, W.; Yi, Y.; Zhang, Z.; Chao, S. Long time-series glacier outlines in the three-rivers headwater region from 1986 to 2021 based on deep learning. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2022, 15, 5734–5752. [Google Scholar] [CrossRef]
  74. Chen, J.; Wu, M.; Yan, H.; Xie, B.; Zhang, C. Change-aware network for damaged roads recognition and assessment based on multi-temporal remote sensing imageries. In Proceedings of the Chinese Conference on Pattern Recognition and Computer Vision (PRCV), Xiamen, China, 13–15 October 2024; Liu, Q., Ma, Z., Zheng, W., Zha, H., Chen, X., Wang, L., Ji, R., Eds.; Springer: Singapore, 2024; pp. 255–266. [Google Scholar]
  75. Zhou, M.; Zhou, Y.; Yang, D.; Song, K. Remote sensing image classification based on canny operator enhanced edge features. Sensors 2024, 24, 3912. [Google Scholar] [CrossRef] [PubMed]
  76. Zhao, Z.; Tan, S.; Zhang, Q.; Chen, H. Automatic identification model for landslide disaster using remote sensing images based on improved multiresunet. IEEE Access 2025, 13, 10653–10662. [Google Scholar] [CrossRef]
  77. Le, C.; Pham, L.; Lampert, J.; Schlögl, M.; Schindler, A. Landslide detection and segmentation using remote sensing images and deep neural networks. In Proceedings of the IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium, Athens, Greece, 7–12 July 2024; IEEE: Piscataway, NJ, USA, 2024; pp. 9582–9586. [Google Scholar]
  78. Ciccone, F.; Ceruti, A.; Bacciaglia, A.; Meisina, C. Automating landslips segmentation for damage assessment: A comparison between deep learning and classical models. In Proceedings of the International Conference of the Italian Association of Design Methods and Tools for Industrial Engineering, Florence, Italy, 6–8 September 2023; Springer: Cham, Switzerland, 2023; pp. 91–99. [Google Scholar]
  79. Ding, L.; Goshtasby, A. On the Canny edge detector. Pattern Recogn. 2001, 34, 721–725. [Google Scholar] [CrossRef]
  80. Ren, Z.; Ma, J.; Liu, J.; Deng, X.; Zhang, G.; Guo, H. Enhancing deep learning-based landslide detection from open satellite imagery via multisource data fusion of spectral, textural, and topographical features: A case study of old landslide detection in the Three Gorges Reservoir Area (TGRA). Geocarto Int. 2024, 39, 2421224. [Google Scholar] [CrossRef]
  81. Zhang, R.; Lv, J.; Yang, Y.; Wang, T.; Liu, G. Analysis of the impact of terrain factors and data fusion methods on uncertainty in intelligent landslide detection. Landslides 2024, 21, 1849–1864. [Google Scholar] [CrossRef]
  82. Wang, H.; Nie, D.; Tuo, X.; Zhong, Y. Research on crack monitoring at the trailing edge of landslides based on image processing. Landslides 2020, 17, 985–1007. [Google Scholar] [CrossRef]
  83. Theckedath, D.; Sedamkar, R.R. Detecting affect states using VGG16, ResNet50 and SE-ResNet50 networks. SN Comput. Sci. 2020, 1, 79. [Google Scholar] [CrossRef]
  84. Qin, H.; Wang, J.; Mao, X.; Zhao, Z.; Gao, X.; Lu, W. An improved faster R-CNN method for landslide detection in remote sensing images. J. Geovis. Spat. Anal. 2024, 8, 2. [Google Scholar] [CrossRef]
  85. Ronneberger, O.; Fischer, P.; Brox, T. U-net: Convolutional networks for biomedical image segmentation. In Proceedings of the Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015. MICCAI 2015. Lecture Notes in Computer Science, Munich, Germany, 5–9 October 2015; Navab, N., Hornegger, J., Wells, W., Frangi, A., Eds.; Springer: Cham, Switzerland, 2015; pp. 234–241. [Google Scholar]
  86. Ibtehaz, N.; Rahman, M.S. MultiResUNet: Rethinking the U-Net architecture for multimodal biomedical image segmentation. Neural Netw. 2020, 121, 74–87. [Google Scholar] [CrossRef]
  87. Chen, L.C.; Zhu, Y.; Papandreou, G.; Schroff, F.; Adam, H. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European Conference On Computer Vision (ECCV), Munich, Germany, 8–14 September 2018; Springer: Cham, Switzerland, 2018; pp. 833–851. [Google Scholar]
  88. Chen, J.; Lu, Y.; Yu, Q.; Luo, X.; Adeli, E.; Wang, Y.; Lu, L.; Yuille, A.L.; Zhou, Y. Transunet: Transformers make strong encoders for medical image segmentation. arXiv 2021, arXiv:2102.04306. [Google Scholar] [CrossRef]
  89. Hu, J.; Shen, L.; Sun, G. Squeeze-and-excitation networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Salt Lake City, UT, USA, 18–22 June 2018; pp. 7132–7141. [Google Scholar]
  90. Woo, S.; Park, J.; Lee, J.Y.; Kweon, I.S. Cbam: Convolutional block attention module. In Proceedings of the European Conference on Computer Vision (ECCV), Munich, Germany, 8–14 September 2018; pp. 3–19. [Google Scholar]
  91. Hacıefendioğlu, K.; Adanur, S.; Demir, G. Automatic landslide segmentation using a combination of grad-CAM visualization and K-means clustering techniques. Iran. J. Sci. Technol. Trans. Civ. Eng. 2024, 48, 943–959. [Google Scholar] [CrossRef]
  92. Huang, Y.; Xie, C.; Li, T.; Xu, C.; He, X.; Shao, X.; Xu, X.; Zhan, T.; Chen, Z. An open-accessed inventory of landslides triggered by the MS 6.8 Luding earthquake, China on September 5, 2022. Earthq. Res. Adv. 2023, 3, 100181. [Google Scholar] [CrossRef]
  93. Zeng, C.; Cao, Z.; Su, F.; Zeng, Z.; Yu, C. A dataset of high-precision aerial imagery and interpretation of landslide and debris flow disaster in Sichuan and surrounding areas between 2008 and 2020. China Sci. Data 2022, 7, 191–201. [Google Scholar] [CrossRef]
  94. Ghorbanzadeh, O.; Xu, Y.; Ghamisi, P.; Kopp, M.; Kreil, D. Landslide4sense: Reference benchmark data and deep learning models for landslide detection. IEEE Trans. Geosci. Remote Sens. 2022, 60, 5633017. [Google Scholar] [CrossRef]
  95. Shahri, A.A.; Spross, J.; Johansson, F.; Larsson, S. Landslide susceptibility hazard map in southwest Sweden using artificial neural network. Catena 2019, 183, 104225. [Google Scholar] [CrossRef]
  96. Liu, S.; Zhou, J.; Qiu, Y.; Chen, J.; Zhu, X.; Chen, H. The FIRST model: Spatiotemporal fusion incorrporting spectral autocorrelation. Remote Sens. Environ. Interdiscip. J. 2022, 279, 113111. [Google Scholar] [CrossRef]
  97. Edrich, A.K.; Yildiz, A.; Roscher, R.; Bast, A.; Graf, F.; Kowalski, J. A modular framework for FAIR shallow landslide susceptibility mapping based on machine learning. Nat. Hazards 2024, 120, 8953–8982. [Google Scholar] [CrossRef]
  98. Wang, J.; Li, S.; Zhang, R. A modular deep learning framework for landslide susceptibility mapping using multi-source geospatial data. Bull. Eng. Geol. Environ. 2020, 79, 3425–3441. [Google Scholar] [CrossRef]
  99. Dabov, K.; Foi, A.; Katkovnik, V.; Egiazarian, K. Image denoising with block-matching and 3D filtering. In Proceedings of the Image Processing: Algorithms and Systems, Neural Networks, and Machine Learning, San Jose, CA, USA, 16–18 January 2006; SPIE: Bellingham, WA, USA, 2006; pp. 354–365. [Google Scholar]
  100. Wang, H.; Zhang, L.; Yin, K.; Luo, H.; Li, J. Landslide identification using machine learning. Geosci. Front. 2021, 12, 351–364. [Google Scholar] [CrossRef]
  101. Lv, P.; Ma, L.; Li, Q.; Du, F. ShapeFormer: A shape-enhanced vision transformer model for optical remote sensing image landslide detection. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2023, 16, 2681–2689. [Google Scholar] [CrossRef]
  102. Saha, S.; Roy, J.; Pradhan, B.; Hembram, T.K. Hybrid ensemble machine learning approaches for landslide susceptibility mapping using different sampling ratios at East Sikkim Himalayan, India. Adv. Space Res. 2021, 68, 2819–2840. [Google Scholar] [CrossRef]
  103. Abbaszadeh Shahri, A.; Chunling, S.; Larsson, S. A hybrid ensemble-based automated deep learning approach to generate 3D geo-models and uncertainty analysis. Eng. Comput. 2024, 40, 1501–1516. [Google Scholar] [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions, and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions, or products referred to in the content.
Remotesensing 17 00995 g001
Figure 1. Flowchart of the study.
Figure 1. Flowchart of the study.
Remotesensing 17 00995 g001
Remotesensing 17 00995 g002
Figure 2. Remote sensing image location distribution.
Figure 2. Remote sensing image location distribution.
Remotesensing 17 00995 g002
Remotesensing 17 00995 g003
Figure 3. Three-dimensional modeling assistance.
Figure 3. Three-dimensional modeling assistance.
Remotesensing 17 00995 g003
Remotesensing 17 00995 g004
Figure 4. Comparison of landslide annotations before and after modification. (a) Before modification. (b) After modification.
Figure 4. Comparison of landslide annotations before and after modification. (a) Before modification. (b) After modification.
Remotesensing 17 00995 g004
Remotesensing 17 00995 g005
Figure 5. Sample examples. (a) Difficult samples; (b) simple samples; (c) cloud and fog obstructions; (d) image distortion; (e) over-illumination; (f) under illumination.
Figure 5. Sample examples. (a) Difficult samples; (b) simple samples; (c) cloud and fog obstructions; (d) image distortion; (e) over-illumination; (f) under illumination.
Remotesensing 17 00995 g005
Remotesensing 17 00995 g006
Figure 6. U-Net architecture.
Figure 6. U-Net architecture.
Remotesensing 17 00995 g006
Remotesensing 17 00995 g007
Figure 7. Schematic diagram of the residual network structure.
Figure 7. Schematic diagram of the residual network structure.
Remotesensing 17 00995 g007
Remotesensing 17 00995 g008
Figure 8. BFA structure diagram.
Figure 8. BFA structure diagram.
Remotesensing 17 00995 g008
Remotesensing 17 00995 g009
Figure 9. General structure of ResUNet–BFA.
Figure 9. General structure of ResUNet–BFA.
Remotesensing 17 00995 g009
Remotesensing 17 00995 g010
Figure 10. Comparison of simple sample prediction results of landslides with different semantic segmentation models. (a) Simple samples; (b) difficult samples.
Figure 10. Comparison of simple sample prediction results of landslides with different semantic segmentation models. (a) Simple samples; (b) difficult samples.
Remotesensing 17 00995 g010
Remotesensing 17 00995 g011
Figure 11. Loss curves for different attentional mechanisms. (a) SE; (b) SAM; (c) CBAM; (d) BFA.
Figure 11. Loss curves for different attentional mechanisms. (a) SE; (b) SAM; (c) CBAM; (d) BFA.
Remotesensing 17 00995 g011
Remotesensing 17 00995 g012
Figure 12. Heatmap of prediction by different attention mechanisms. (a) Simple samples; (b) difficult samples.
Figure 12. Heatmap of prediction by different attention mechanisms. (a) Simple samples; (b) difficult samples.
Remotesensing 17 00995 g012
Remotesensing 17 00995 g013
Figure 13. Recognition results for the Luding earthquake landslide. (a) Remote sensing image; (b) label; (c) landslide detection results.
Figure 13. Recognition results for the Luding earthquake landslide. (a) Remote sensing image; (b) label; (c) landslide detection results.
Remotesensing 17 00995 g013
Table 1. Specific information of some typical landslide datasets.
Table 1. Specific information of some typical landslide datasets.
Landslide DatasetNumber of ImagesOriginal Resolution (m)Cropped Size (pixels)Study AreaData Source
Three Gorges Reservoir Area Landslide Dataset894Unknown224 × 224Three Gorges Reservoir Area in ChinaShipborne Images
Loess Landslide Dataset21542 × 2512 × 512Southeastern Gansu Province in ChinaGF-1 Satellite Images
UAV Landslide Dataset160Unknown512 × 512Multiple Regions GloballyUnmanned Aerial Vehicle (UAV) imagery
Bijie Landslide Dataset27730.8 × 0.8Varies by landslide sizeBijie in ChinaTripleSat Satellite Images
CAS Landslide Dataset20,8650.2 × 0.2–5 × 5512 × 512Nine Different Regions GloballyMultiple Satellite and UAV Images
Table 2. Overview of seismic hazards in the study area.
Table 2. Overview of seismic hazards in the study area.
Earthquake NameMagnitudeDateNumber of Aftershocks with Magnitude 4.0 or AboveDeath TollDirect Economic Loss/Billion USD
Wenchuan Earthquake8.012 May 200831169,2271198.72
Ludian Earthquake6.53 August 2014461728.09
Jiuzhaigou Earthquake7.08 August 201732511.41
Table 3. Dataset image distribution.
Table 3. Dataset image distribution.
CategoryLandslideNon-LandslideTotal
Training set16425392181
Validation set20469273
Test set19776273
Total20436842727
Table 4. Experimental environment.
Table 4. Experimental environment.
Hardware environmentCPU12 vCPU Intel(R) Xeon(R) Platinum 8352 V CPU @ 2.10GHZ
GPUNVIDIA GeForce RTX 4090(24GB), 1
Operating systemUbuntu22.04
Software environmentTorch2.3.0 (Cuda 12.1)
Python3.12
Table 5. Model training hyperparameter settings.
Table 5. Model training hyperparameter settings.
Parameter NameParameter Value
Image Size224 × 224
Loss functionFocal Loss + Dice Loss
Batch Size16
Learning Rate0.0001
Max Epoch50
OptimizerAdam
Decay Exponent0.9
Early stopping patience10
Table 6. Confusion matrix.
Table 6. Confusion matrix.
Actual ClassPredicted Class
PositiveNegative
PositiveTrue Positive (TP)False Negative (FN)
NegativeFalse Positive (FP)True Negative (TN)
Table 7. Comparison of training and testing times for different models. The italicized numbers indicate the best results for each indicator.
Table 7. Comparison of training and testing times for different models. The italicized numbers indicate the best results for each indicator.
ModelTraining Time (Minutes)Testing Time (Seconds)
UNet39.2296.33
MultiResUNet57.9054.53
DeepLabv3+77.8249.19
TransUNet29.9769.14
ResUNet67.8833.41
ResUNet–BFA31.2744.77
Table 8. Landslide detection results of common network structures. The italicized numbers indicate the best results for each indicator.
Table 8. Landslide detection results of common network structures. The italicized numbers indicate the best results for each indicator.
ModelTPFPFNTNPrecisionRecallF1 ScoreMPAMIoU
UNet816,541355,479280,15512,245,87369.6774.4571.9883.7275.65
MultiResUNet636,150535,870138,15012,387,87854.2882.1665.3776.5971.70
DeepLabv3+837,791334,229236,53612,289,49271.4877.9874.5984.8077.52
TransUNet887,852284,168282,89012,243,13875.7575.8475.8086.7578.30
ResUNet846,307325,713205,41312,320,61572.2180.4776.1285.2978.65
ResUNet–BFA915,268256,752254,17312,271,85578.0978.2778.1888.0380.09
Table 9. Landslide detection results of different attention modules on ResUNet. The italicized numbers indicate the best results for each indicator.
Table 9. Landslide detection results of different attention modules on ResUNet. The italicized numbers indicate the best results for each indicator.
DescriptionTPFPFNTNPrecisionRecallF1 ScoreMPAMIoU
SE919,423252,597342,74912,183,27978.4572.8475.5487.8678.02
SAM792,349379,671181,08512,344,94367.6181.4073.8683.0877.11
CBAM867,388304,632235,14212,290,88674.0178.6776.2786.0778.72
BFA915,268256,752254,17312,271,85578.0978.2778.1888.0380.09
Table 10. Results of dataset comparison. The italicized numbers indicate the best results for each indicator.
Table 10. Results of dataset comparison. The italicized numbers indicate the best results for each indicator.
DatasetNumberPrecisionRecallF1 ScoreMPAMIoU
Bijie77031.8219.0923.8660.6949.31
Sichuan and Surrounding Areas Landslide Dataset5971.1824.7636.7480.8955.99
Landslide4Sense379958.860.4820.95673.4744.25
Ours272757.8452.4054.9875.7263.48
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Zhao, Z.; Tan, S.; Yang, Y.; Zhang, Q. Landslide Identification from Post-Earthquake High-Resolution Remote Sensing Images Based on ResUNet–BFA. Remote Sens. 2025, 17, 995. https://doi.org/10.3390/rs17060995

AMA Style

Zhao Z, Tan S, Yang Y, Zhang Q. Landslide Identification from Post-Earthquake High-Resolution Remote Sensing Images Based on ResUNet–BFA. Remote Sensing. 2025; 17(6):995. https://doi.org/10.3390/rs17060995

Chicago/Turabian Style

Zhao, Zhenyu, Shucheng Tan, Yiquan Yang, and Qinghua Zhang. 2025. "Landslide Identification from Post-Earthquake High-Resolution Remote Sensing Images Based on ResUNet–BFA" Remote Sensing 17, no. 6: 995. https://doi.org/10.3390/rs17060995

APA Style

Zhao, Z., Tan, S., Yang, Y., & Zhang, Q. (2025). Landslide Identification from Post-Earthquake High-Resolution Remote Sensing Images Based on ResUNet–BFA. Remote Sensing, 17(6), 995. https://doi.org/10.3390/rs17060995

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop