Landslide identification method and system fused with deep learning

Abstract

The invention discloses a landslide identification method and system fused with deep learning, and belongs to the technical field of geological disaster detection, and the method comprises the steps: employing a pre-trained landslide semantic segmentation model to carry out the recognition of obtained landslide remote sensing image data, and outputting a landslide region binary segmentation result; wherein the landslide semantic segmentation model takes a TransUNet model as a skeleton and comprises an encoder, a coordinate attention module is introduced into a jump connection layer, and an edge guidance attention module is introduced behind a decoder; the encoder extracts high-resolution characteristics of the landslide remote sensing image data; the coordinate attention module performs spatial screening on the high-resolution features and then performs weighted fusion with the high-resolution features; and the edge guiding attention module obtains the edge contour feature of the high-resolution feature, and fuses the semantic feature of the weighted fusion feature predicted and generated by the decoder, the edge contour feature and the weighted fusion feature to obtain a binary segmentation result. The model can strengthen edge features and improve identification precision.

Description

Technical Field

[0001] This invention relates to the field of geological disaster detection technology, and more specifically to a landslide identification method and system that integrates deep learning. Background Technology

[0002] Landslides, as one of the most widespread and destructive types of geological hazards globally, are often closely related to factors such as seismic activity, heavy rainfall events, dynamic changes in groundwater, extreme weather events, and river erosion. These sudden geological processes not only cause drastic changes in the Earth's surface morphology but can also trigger secondary disasters such as disruption of transportation networks, damage to buildings and facilities, and casualties, posing a dual threat to human society and the natural ecosystem.

[0003] Against this backdrop, systematic landslide identification is crucial. Accurately recording landslide spatial distribution, temporal characteristics, and morphological information is an indispensable core data foundation for constructing disaster risk assessment models and formulating disaster prevention and mitigation strategies. However, due to the spectral similarity between landslides and surrounding bare land, and their inherent complexity, intelligent landslide identification remains a challenge. Current landslide detection methods based on remote sensing imagery are mainly divided into five categories: visual interpretation methods, pixel classification methods, object-oriented classification methods, machine learning methods, and deep learning methods. These methods differ significantly in their technical principles and application effectiveness.

[0004] Visual interpretation relies on the prior knowledge and experience of professionals, identifying landslides by visually interpreting their morphological features and spectral anomalies. Pixel classification uses individual pixels as the basic unit, utilizing spectral response values ​​or temporal differences to partition landslides. Object-oriented classification overcomes pixel limitations by aggregating homogeneous pixels into objects through multi-scale segmentation, constructing classification rules based on the objects' geometric shape, texture, and topological relationships. Machine learning, through adaptive feature learning mechanisms, demonstrates significant advantages in processing high-dimensional remote sensing data and mapping complex nonlinear relationships of landslide features. Its mainstream algorithms include Support Vector Machines (SVM), Decision Trees (DT), Random Forests (RF), and Artificial Neural Networks (ANN), effectively mining the feature interactions of multi-source data such as spectral, topographic, and texture data. Deep learning-based landslide detection technologies mainly include two categories: target detection and semantic segmentation. Target detection methods locate the spatial position of landslides by generating bounding boxes, enabling rapid identification and localization of multi-target landslides. Semantic segmentation methods extract pixel-level landslide boundaries through pixel-by-pixel classification, and are particularly adept at handling fine-grained segmentation of fragmented landslides. Extensive research has shown that the TranUNet model has a semantic gap between the encoder and decoder, which introduces a lot of noise into the model and interferes with the decoder's accurate segmentation of the target, resulting in weak boundary problems.

[0005] In recent years, deep learning technology has made breakthrough progress, especially the hierarchical feature extraction mechanism and end-to-end learning paradigm of convolutional neural networks (CNNs), which have provided a new path to overcome the bottlenecks of traditional landslide identification methods, such as reliance on human experience and low efficiency. However, there are still problems such as low accuracy of extraction results and blurred boundaries, which make it impossible to accurately identify landslide problems in complex environments. Summary of the Invention

[0006] To address the problems existing in the above-mentioned fields, this invention proposes a landslide recognition method and system that integrates deep learning. The constructed landslide semantic segmentation model can effectively shorten the semantic gap by introducing a coordinate attention module in the skip connection layer, thereby suppressing semantic noise. After the encoder, an edge-guided attention module can be introduced to fuse multiple features, thereby suppressing noise interference and alleviating the weak boundary problem.

[0007] To address the aforementioned technical problems, this invention discloses a landslide identification method integrating deep learning, comprising the following steps: Acquire landslide remote sensing image data; A pre-trained landslide semantic segmentation model is used to identify landslide remote sensing image data and output binary segmentation results of landslide areas and non-landslide areas. The landslide semantic segmentation model uses the TransUNet model as its skeleton, including an encoder, a skip connection layer, and a decoder. A coordinate attention module is introduced in the skip connection layer, and an edge-guided attention module is introduced after the decoder. The encoder extracts high-resolution features from landslide remote sensing image data; a coordinate attention module is introduced to spatially filter the high-resolution features, and the filtered features are weighted and fused with the high-resolution features; an edge-guided attention module is used to obtain the edge contour features of the high-resolution features, and the semantic features, edge contour features, and weighted fusion features predicted by the decoder are fused together to output the landslide target range features and perform binarization mapping to obtain the binary segmentation result.

[0008] Preferably, the step of using a pre-trained landslide semantic segmentation model to identify landslide remote sensing image data and outputting binary segmentation results for landslide areas and non-landslide areas specifically includes: A regular grid partitioning strategy is adopted. By determining the coverage grid coordinates of the original image of the landslide sample in the landslide remote sensing image of the study area, the original image is divided into sub-patterns of fixed size consistent with the input of the landslide semantic segmentation model. When the edge part of the landslide remote sensing image of the study area is insufficient to be divided into a complete sub-pattern, the boundary area of ​​the sub-pattern is filled with black to fill it into a complete sub-pattern, and the row and column numbers of each sub-pattern are determined. After normalizing the multiple sub-maps, they are input into the trained landslide semantic segmentation model, which outputs a binary map corresponding to each sub-map, where 1 represents a landslide area and 0 represents a non-landslide area. Based on the row and column numbers of the sub-plots, a regular block recombination strategy is adopted to recombine the binary images corresponding to all sub-plots, remove the edge fill parts, and finally obtain a complete binary image with the same size as the landslide sample in the original study area.

[0009] Preferably, the extraction of high-resolution features from landslide remote sensing image data via the encoder specifically includes: The encoder includes a ResNet50 convolutional network and a Transformer. During the encoding process, multiple convolutional layers of the ResNet50 convolutional network are used to extract feature maps of landslide remote sensing image data from shallow to deep layers through stepwise downsampling. The deepest feature map is used as input to the Transformer. The Transformer establishes a global context relationship between the feature maps, thereby extracting high-resolution features from the landslide remote sensing image data.

[0010] Preferably, the step of spatially filtering high-resolution features by introducing a coordinate attention module and then weightedly fusing the filtered features with the high-resolution features specifically includes: By introducing a coordinate attention module, the attention weights of high-resolution features in the horizontal and vertical directions are calculated separately, and then weighted and fused to obtain the coordinate attention-weighted features, i.e. the selected features. The coordinate attention-weighted features are combined with the high-resolution features extracted by the encoder by element-wise addition or concatenation to obtain weighted fused features.

[0011] Preferably, the high-resolution edge contour features obtained through the edge-guided attention module are fused together with the semantic features, edge contour features, and weighted fusion features generated by the decoder to output the landslide target range features and perform binarization mapping to obtain a binary segmentation result, specifically including: The weighted fused features are input into the decoder, and the semantic features of the generated weighted fused features are predicted through upsampling and convolution operations. By introducing an edge-guided attention module, the edge contour features of the high-resolution feature map are obtained after convolution with the Laplacian operator. The semantic features, edge contour features and weighted fusion features are then weighted and fused through an attention mechanism. Finally, the fused features that combine boundary information, spatial positioning and semantic information are output, namely the target range features. The target range features are mapped to binary segmentation results by thresholding or classifying, where 0 represents non-landslide areas and 1 represents landslide areas.

[0012] Preferably, acquiring landslide remote sensing image data further includes data augmentation of the landslide remote sensing image dataset, specifically including: The acquired landslide remote sensing image data was divided into training set, validation set and test set. Each dataset was then randomly rotated at any angle, Gaussian noise was added, horizontal / vertical flipping was performed and brightness and contrast were adjusted, and then normalized to obtain the normalized training set, validation set and test set. The normalized training set is input into the landslide semantic segmentation model for model training.

[0013] Preferably, the step of inputting the normalized training set into the landslide semantic segmentation model for model training also includes inputting the test set into the trained landslide semantic segmentation model and outputting the binary segmentation results of the landslide area and non-landslide area through verification.

[0014] Preferably, it also includes a landslide identification system incorporating deep learning, comprising: The data acquisition module is used to acquire landslide remote sensing image data; The landslide identification module is used to identify landslide remote sensing image data using a pre-trained landslide semantic segmentation model, and outputs binary segmentation results of landslide areas and non-landslide areas. The landslide semantic segmentation model uses the TransUNet model as its backbone, including an encoder, a skip connection layer, and a decoder. A coordinate attention module is introduced in the skip connection layer, and an edge-guided attention module is introduced after the decoder. The encoder extracts high-resolution features from the landslide remote sensing image data; the coordinate attention module performs spatial filtering on the high-resolution features, and then weights and fuses the filtered features with the high-resolution features; the edge-guided attention module obtains the edge contour features of the high-resolution features, and then fuses the semantic features, edge contour features, and weighted fusion features predicted by the decoder to output the landslide target range features, which are then binarized to obtain the binary segmentation result.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The proposed landslide recognition method, which integrates deep learning, constructs a landslide semantic segmentation model. This model introduces a coordinate attention module in the skip connection layer. By extracting high-resolution features from the encoder and performing spatial filtering, the filtered features are weighted and fused with the high-resolution features. This reduces the semantic gap between the encoder and decoder, thus suppressing semantic noise. Simultaneously, by introducing an edge-guided attention module after the decoder, edge contour features of the high-resolution features are obtained. The semantic features, edge contour features, and weighted fused features generated by the decoder are fused together. The edge contour features serve as a guiding signal, enhancing the weight of the target boundary region. The weighted fused features supplement spatial localization, and the decoder prediction provides semantic guidance. The final output is a target range feature that combines accurate boundary information, clear spatial localization, and strong semantic consistency. By strengthening edge feature information, noise interference is suppressed, mitigating the weak boundary problem and further improving the model's prediction accuracy. Attached Figure Description

[0016] Figure 1 This is a flowchart of the landslide identification method integrating deep learning proposed in this invention; Figure 2 The resulting image of data enhancement provided for an embodiment of the present invention; Figure 3 This is the network architecture of the landslide semantic segmentation model provided in the embodiments of the present invention; Figure 4 Comparative experimental results provided for embodiments of the present invention; Figure 5 The ablation experiment results provided in the embodiments of the present invention; Figure 6 The present invention provides the recognition results of applying the trained CETransUNet model to the actual research area. Detailed Implementation

[0017] The following will refer to the appendices in the embodiments of the present invention. Figures 1-6 The technical solutions in the embodiments of the present invention will be clearly and completely described. It should be understood that the terminology used in the present invention is only for describing particular implementation methods and is not intended to limit the present invention.

[0018] This invention significantly improves the accuracy of landslide remote sensing image interpretation by constructing a landslide semantic segmentation model optimized by Transformer and dual attention, namely the CETransUNet model. It demonstrates high positioning accuracy and boundary clarity in regional landslide identification tasks, providing reliable technical support for rapid response and risk assessment in the field of disaster prevention and mitigation.

[0019] Example like Figure 1As shown, this invention proposes a landslide identification method that integrates deep learning, comprising the following steps: S1: Obtain the landslide remote sensing image dataset; S2: A pre-trained landslide semantic segmentation model is used to identify landslide remote sensing image data and output binary segmentation results of landslide areas and non-landslide areas; The constructed landslide semantic segmentation model uses the TransUNet model as its backbone, including an encoder, a skip connection layer, and a decoder. A coordinate attention module is introduced in the skip connection layer, and an edge-guided attention module is introduced after the decoder. The encoder extracts high-resolution features from landslide remote sensing image data; a coordinate attention module is introduced to spatially filter the high-resolution features, and the filtered features are weighted and fused with the high-resolution features; an edge-guided attention module is used to obtain the edge contour features of the high-resolution features, and the decoder predicts and generates the semantic features, edge contour features, and weighted fusion features of the weighted fusion features. The three features are then fused to output the landslide target range features and perform binarization mapping to obtain the binary segmentation result.

[0020] Specifically, in step S1, landslide remote sensing image data is acquired. In this embodiment, the Iburi-Tobu dataset from Japan is acquired, and 80% is randomly selected as the training set, 10% as the validation set, and 10% as the test set. Then, data augmentation is performed using four methods through Python code: random rotation at arbitrary angles, addition of Gaussian noise, horizontal / vertical flipping, and brightness / contrast adjustment. Figure 2 As shown, the process also includes normalizing the augmented datasets, thereby expanding the Iburi-Tobu dataset to 7420 landslide images of 256×256 pixels each, resulting in a normalized training set (5936 images), a validation set (742 images), and a test set (742 images). This normalized training set is then input into the landslide semantic segmentation model for training.

[0021] In step S2, the construction process of the landslide semantic segmentation model (CETransUNet model) specifically includes: The encoder of the CETransUNet model includes a ResNet50 convolutional network and a Transformer. During the encoding process, multiple convolutional layers of the ResNet50 convolutional network are used to extract feature maps from landslide remote sensing image data from shallow to deep layers through progressive downsampling. The feature map of the deepest layer is used as the input of the Transformer, and the global context relationship of the feature maps is established through the Transformer to extract high-resolution features from the landslide remote sensing image data.

[0022] A coordinate attention module is introduced into the skip connection layer to perform spatial filtering of high-resolution features. The attention weights of high-resolution features in the horizontal and vertical directions are calculated independently and then weighted and fused to obtain the coordinate attention-weighted features, i.e. the filtered features. The coordinate attention-weighted features are combined with the high-resolution features extracted by the encoder by element-wise addition or concatenation to obtain weighted fused features.

[0023] The weighted fused features are input into the decoder, and the semantic features of the generated weighted fused features are predicted through upsampling and convolution operations. Following the decoder, an edge-guided attention module is introduced. This module convolves the high-resolution feature map with a Laplacian operator, then calculates the edge contour features using the Laplacian function. An attention mechanism is then used to weight and fuse the semantic features, edge contour features, and weighted fusion features, ultimately outputting a fused feature that combines boundary information, spatial localization, and semantic information—the target range feature. This target range feature is then mapped to a binary segmentation result using thresholding or a classifier, where 0 represents a non-slope area and 1 represents a slope area. The network architecture of the constructed CETransUNet model is as follows: Figure 3 As shown.

[0024] The edge-guided attention module performs weighted fusion of the above three types of features through an attention mechanism (for example, using edge information as a "guiding signal" to enhance the weight of the target boundary region; using encoder detail features to supplement spatial positioning; and using decoder preliminary features to provide semantic guidance), and finally outputs a fused feature that combines accurate boundary information, clear spatial positioning and strong semantic consistency, namely the target range feature.

[0025] The loss function uses a combination of binary cross-entropy and Dice Loss: In the formula, and These represent the binary cross-entropy loss function and the Dice Loss function, respectively. The former ensures pixel-level probability likelihood optimization, while the latter addresses the extreme class imbalance between foreground and background in landslide images through overlap metric enhancement.

[0026] In step S3, the constructed landslide semantic segmentation model is trained using the training set and validation set, and the predictive performance of the model is verified using the test set.

[0027] The landslide semantic segmentation model constructed in step S2 was trained using 5936 training images and 742 validation images. The Adam optimizer was used for gradient updates, with the core parameter set to an initial learning rate of 1×10. The learning rate scheduling strategy employed a cosine annealing global optimization mechanism, and the total number of training epochs was strictly limited to 60. The batch size was 16. After training, the 742 test images were input into the trained landslide semantic segmentation model for testing.

[0028] To evaluate the landslide hazard identification capability of the CETransUNet model constructed in this invention, this embodiment selected seven representative benchmark models, including UNet, UNet++, and SegUNet, for comparative experiments. Specific experimental results are shown in Table 1 and... Figure 4 As shown in Table 1, the accuracy statistics of the quantitative indicators for each model comparison experiment are presented. Figure 4 These are the prediction results for each model.

[0029] Table 1. Precision statistics of quantitative indicators for each model in comparative experiments. Among them, IoU, MIoU, Precision, Recall, and F1-Score are the intersection-over-union ratio, average intersection-over-union ratio, precision, recall, and F1 score, respectively.

[0030] Analysis of the data in Table 1 shows that the CETransUNet model constructed in this invention has achieved significant improvements, with the highest values ​​in all five metrics: IoU, MIoU, Precision, Recall, and F1-Score, at 74.77%, 86.00%, 85.61%, 85.48%, and 85.53%, respectively. Compared with the UNet++ model, these figures represent improvements of 5.96%, 3.35%, 4.03%, 4.08%, and 4.06%, respectively.

[0031] This invention uses ablation experiments to verify the effectiveness of each improved module of the CETransUNet model in extracting landslide hazard areas. Four sets of ablation experiments were set up: the original TransUNet model, a TransUNet model structure with added coordinate attention module, a TransUNet model structure with added edge-guided attention module, and a TransUNet model structure with both coordinate attention and edge-guided attention modules. Based on the same experimental setup, all four models were trained for 60 rounds, and ablation experiments were conducted using a test set created with the same dataset. The experimental results are shown in Table 2 and... Figure 5 As shown in Table 2, the accuracy statistics of quantitative indicators in the ablation experimental model are presented. Figure 5 This is the predicted result of the ablation experiment.

[0032] Table 2. Accuracy Statistics of Quantitative Indicators in Ablation Experimental Models Analysis of the data in Table 2 shows that the CETransUNet model constructed in this invention achieves significant improvements, reaching the highest values ​​in all five metrics: IoU, MIoU, Precision, Recall, and F1-Score. Specifically, the TransUNet model structure with the added coordinate attention module reduces the semantic gap and suppresses noise errors in the skip connection layer compared to the original TransUNet model; the TransUNet model structure with the added edge guidance attention module strengthens edge features after the decoder, improving recognition accuracy compared to the original TransUNet model; and the TransUNet model structure with both the coordinate attention module and the edge guidance module not only reduces the semantic gap in the skip connection layer but also strengthens edge features, thus further improving the model's prediction accuracy.

[0033] The process also includes step S4: After the landslide remote sensing images in the landslide remote sensing image dataset of the study area are standardized by regular block division and edge filling, they are input into the trained CETransUNet model to extract the landslide range and stitch the binary images together.

[0034] Specifically, landslide samples from a watershed in the study area were obtained through visual interpretation combined with UAV imagery and Google Earth imagery.

[0035] A regular grid partitioning strategy is adopted. By calculating the coverage grid coordinates of the original image in the landslide samples of the landslide remote sensing image of the study area, the original image is divided into multiple small image sub-patterns of fixed size that are consistent with the input of the constructed CETransUNet model. This ensures that each small image sub-pattern can adapt to the input requirements of the CETransUNet model and avoids memory overflow or the inability of the CETransUNet model to calculate due to excessive size. At the same time, the row and column numbers of each small image sub-pattern are determined.

[0036] When the edge of the landslide remote sensing image in the study area is insufficient to be divided into a complete small image sub-pattern, the boundary area of ​​the sub-pattern is filled with black to complete the small image sub-pattern, ensuring that all areas can be processed by the CETransUNet model and avoiding the loss of edge information.

[0037] After normalizing the segmented small image sub-patterns, they are input into the trained CETransUNet model. The encoder extracts the local features and global context of the small image sub-patterns, and combined with the edge-guided attention module introduced by the decoder, the landslide area within the small image sub-pattern is accurately identified. The binary map corresponding to each small image sub-pattern is output (1 for landslide area and 0 for non-landslide area), realizing the "pixel-level" marking of the landslide range of a single image. This step overcomes the technical defects of traditional manual interpretation, which is time-consuming and highly subjective, thereby realizing the automated and efficient extraction of the landslide range.

[0038] The binary image corresponding to a single small image sub-plot can only reflect the landslide situation in a local area. It is necessary to stitch together the whole to restore the whole. Therefore, according to the row and column numbers recorded when the regular grid division strategy was adopted, the binary images of all small image sub-plots were recombined, and the edge filling parts were removed. Finally, a complete binary image with the same size as the landslide sample of a certain watershed in the original study area was obtained.

[0039] This binary map can visually present the spatial distribution, shape, and extent of all landslides in the study area, and can be directly used for subsequent analyses (such as landslide area statistics, hazard assessment, and overlay analysis with topographic / geological data).

[0040] In short, step S4 is a complete process of "model adaptation to large images → automated extraction of local landslides → stitching and restoration of global distribution". The core objective is to apply the trained model to the actual research area and output a complete range of results that can be directly used for landslide disaster analysis.

[0041] like Figure 6 The figure shows the recognition results of applying the trained CETransUNet model to a real research area according to an embodiment of the present invention. Experimental results show that even in complex interference environments such as low light and water reflection and cloud cover, the CETransUNet model still successfully identified most of the landslides in the area. The specific performance indicators are as follows: IoU 75.42%, MIoU 86.71%, Recall 85.03%, Precision 86.96%, F1-Score 85.99%. This result shows that CETransUNet exhibits good landslide localization ability in complex environments and maintains high recognition accuracy even for fragmented and dense landslides.

[0042] This invention also proposes a landslide identification system integrating deep learning, comprising: The data acquisition module is used to acquire landslide remote sensing image data; The landslide identification module is used to identify landslide remote sensing image data using a pre-trained landslide semantic segmentation model, and outputs binary segmentation results of landslide areas and non-landslide areas. The landslide semantic segmentation model uses the TransUNet model as its backbone, including an encoder, a skip connection layer, and a decoder. A coordinate attention module is introduced in the skip connection layer, and an edge-guided attention module is introduced after the decoder. The encoder extracts high-resolution features from the landslide remote sensing image data; the coordinate attention module performs spatial filtering on the high-resolution features, and then weights and fuses the filtered features with the high-resolution features; the edge-guided attention module obtains the edge contour features of the high-resolution features, and then fuses the semantic features, edge contour features, and weighted fusion features predicted by the decoder to output the landslide target range features, which are then binarized to obtain the binary segmentation result.

[0043] The CETransUNet model constructed in this invention introduces a coordinate attention module in the skip connection layer. By independently calculating the attention weights in the horizontal and vertical directions and fusing the original feature information, it narrows the semantic gap between the encoder and decoder.

[0044] This invention introduces an edge-guided attention module after the decoder, which integrates features from the encoder, edge contour features obtained by calculating the Laplacian function, and semantic features predicted by the decoder itself. By integrating these three feature information, the edge feature information is enhanced.

[0045] This invention employs a regular grid partitioning strategy to segment images into fixed-size sub-blocks with black border filling, while preserving the original dimensions and row / column numbers to achieve standardized block processing. Based on row / column encoding and original size information parsing, precise positioning and stitching of the segmented images are achieved, and geometric dimensions are restored through edge cropping. Simultaneously, it ensures robust synthesis by handling multiple formats and abnormal paths. The entire process is highly automated and effectively suppresses noise interference and weak boundary issues, effectively solving the stitching seam problem that occurs when stitching images in the study area, thus improving recognition accuracy.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0047] Furthermore, unless otherwise stated, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. All references to this specification are incorporated by way of citation to disclose and describe methods relating to those references. In the event of any conflict with any incorporated reference, the content of this specification shall prevail.

Claims

1. A landslide identification method integrating deep learning, characterized in that, Includes the following steps: Acquire landslide remote sensing image data; A pre-trained landslide semantic segmentation model is used to identify landslide remote sensing image data and output binary segmentation results of landslide areas and non-landslide areas. The landslide semantic segmentation model uses the TransUNet model as its skeleton, including an encoder, a skip connection layer, and a decoder. A coordinate attention module is introduced in the skip connection layer, and an edge-guided attention module is introduced after the decoder. The encoder extracts high-resolution features from landslide remote sensing image data; a coordinate attention module is introduced to spatially filter the high-resolution features, and the filtered features are weighted and fused with the high-resolution features; an edge-guided attention module is used to obtain the edge contour features of the high-resolution features, and the semantic features, edge contour features, and weighted fusion features predicted by the decoder are fused together to output the landslide target range features and perform binarization mapping to obtain the binary segmentation result.

2. The landslide identification method integrating deep learning according to claim 1, characterized in that, The method employs a pre-trained landslide semantic segmentation model to identify landslide remote sensing image data and outputs binary segmentation results for landslide and non-landslide areas, specifically including: A regular grid partitioning strategy is adopted. By determining the coverage grid coordinates of the original image of the landslide sample in the landslide remote sensing image of the study area, the original image is divided into sub-patterns of fixed size consistent with the input of the landslide semantic segmentation model. When the edge part of the landslide remote sensing image of the study area is insufficient to be divided into a complete sub-pattern, the boundary area of ​​the sub-pattern is filled with black to fill it into a complete sub-pattern, and the row and column numbers of each sub-pattern are determined. After normalizing the multiple sub-maps, they are input into the trained landslide semantic segmentation model, which outputs a binary map corresponding to each sub-map, where 1 represents a landslide area and 0 represents a non-landslide area. Based on the row and column numbers of the sub-plots, a regular block recombination strategy is adopted to recombine the binary images corresponding to all sub-plots, remove the edge fill parts, and finally obtain a complete binary image with the same size as the landslide sample in the original study area.

3. The landslide identification method integrating deep learning according to claim 1, characterized in that, The extraction of high-resolution features from landslide remote sensing image data via the encoder specifically includes: The encoder includes a ResNet50 convolutional network and a Transformer. During the encoding process, multiple convolutional layers of the ResNet50 convolutional network are used to extract feature maps of landslide remote sensing image data from shallow to deep layers through stepwise downsampling. The deepest feature map is used as input to the Transformer. The Transformer establishes a global context relationship between the feature maps, thereby extracting high-resolution features from the landslide remote sensing image data.

4. The landslide identification method integrating deep learning according to claim 3, characterized in that, The process of introducing a coordinate attention module to spatially filter high-resolution features and then weightedly fusing the filtered features with the high-resolution features specifically includes: By introducing a coordinate attention module, the attention weights of high-resolution features in the horizontal and vertical directions are calculated separately, and then weighted and fused to obtain the coordinate attention-weighted features, i.e. the selected features. The coordinate attention-weighted features are combined with the high-resolution features extracted by the encoder by element-wise addition or concatenation to obtain weighted fused features.

5. The landslide identification method integrating deep learning according to claim 4, characterized in that, The process involves acquiring high-resolution edge contour features through an edge-guided attention module, fusing the semantic features, edge contour features, and weighted fusion features predicted by the decoder, outputting landslide target range features, and performing binarization mapping to obtain a binary segmentation result. Specifically, this includes: The weighted fused features are input into the decoder, and the semantic features of the generated weighted fused features are predicted through upsampling and convolution operations. By introducing an edge-guided attention module, the edge contour features of the high-resolution feature map are obtained after convolution with the Laplacian operator. The semantic features, edge contour features and weighted fusion features are then weighted and fused through an attention mechanism. Finally, the fused features that combine boundary information, spatial positioning and semantic information are output, namely the target range features. The target range features are mapped to binary segmentation results by thresholding or classifying, where 0 represents non-landslide areas and 1 represents landslide areas.

6. The landslide identification method integrating deep learning according to claim 1, characterized in that, The acquisition of landslide remote sensing image data also includes data augmentation of the landslide remote sensing image dataset, specifically including: The acquired landslide remote sensing image data was divided into training set, validation set and test set. Each dataset was then randomly rotated at any angle, Gaussian noise was added, horizontal / vertical flipping was performed and brightness and contrast were adjusted, and then normalized to obtain the normalized training set, validation set and test set. The normalized training set is input into the landslide semantic segmentation model for model training.

7. The landslide identification method integrating deep learning according to claim 6, characterized in that, The step of inputting the normalized training set into the landslide semantic segmentation model for model training also includes inputting the test set into the trained landslide semantic segmentation model and outputting the binary segmentation results of the landslide area and non-landslide area through verification.

8. A landslide identification system integrating deep learning, characterized in that, include: The data acquisition module is used to acquire landslide remote sensing image data; The landslide identification module is used to identify landslide remote sensing image data using a pre-trained landslide semantic segmentation model, and outputs binary segmentation results of landslide areas and non-landslide areas. The landslide semantic segmentation model uses the TransUNet model as its backbone, including an encoder, a skip connection layer, and a decoder. A coordinate attention module is introduced in the skip connection layer, and an edge-guided attention module is introduced after the decoder. The encoder extracts high-resolution features from the landslide remote sensing image data; the coordinate attention module performs spatial filtering on the high-resolution features, and then weights and fuses the filtered features with the high-resolution features; the edge-guided attention module obtains the edge contour features of the high-resolution features, and then fuses the semantic features, edge contour features, and weighted fusion features predicted by the decoder to output the landslide target range features, which are then binarized to obtain the binary segmentation result.

Images