A landslide identification method based on multi-source remote sensing image fusion

Abstract

The invention discloses a landslide identification method based on multi-source remote sensing image fusion, and relates to the technical field of remote sensing image processing. The method comprises: selecting multi-source remote sensing images, constructing vectorized labels, and obtaining binary label data through vectorized labels; using sliding windows to crop multi-source remote sensing images and binary label data, and using computer vision enhancement technology to process multi-source remote sensing images and binary label data; constructing a landslide intelligent identification model based on the MiT architecture; implementing prediction using an ignoring edge prediction method, and obtaining multi-source remote sensing image recognition results according to the prediction results; defining geographic space coordinates according to the multi-source remote sensing image recognition results, and performing binary label vectorization to obtain landslide vector labels. The invention can not only better capture the details and shape characteristics of landslides, but also significantly improve the recognition ability of landslide intelligent identification models for landslides under different terrain and geomorphic conditions.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing image processing, and in particular to a landslide identification method based on multi-source remote sensing image fusion. Background Art

[0002] Landslides are an important and widely studied phenomenon in geological research, and are usually regarded as the product of geological processes occurring in a specific space in the earth's crust. In the long geological history, landslides are formed by various geological processes and are a typical manifestation of geological bodies. They can be regarded as the movement caused by the instability of strata, rock masses, volcanic bodies, or other geological structures on the earth's surface. The formation of landslides is closely related to the genetic characteristics of geological bodies. Their material composition, morphological characteristics, and spatial distribution patterns all show different geological characteristics. Landslides caused by extreme events often cause damage to mountain settlements, and sometimes the degree of landslide disasters may exceed the event itself, because they may interrupt roads and other traffic arteries, making it difficult to effectively respond to emergency situations. Therefore, after an extreme event occurs, rapid and accurate identification of landslides is crucial for timely quantitative assessment of disasters, emergency rescue, and subsequent risk management, especially in mountainous areas with complex environments and inconvenient transportation. Compared with time-consuming and labor-intensive field surveys, remote sensing images cover a large area of ​​earth observation and are not easily affected by infrastructure damaged by disasters, providing an excellent opportunity to explore a vast geographic space.

[0003] Remote sensing technology is an advanced technology that can detect and perceive targets or natural phenomena at a long distance without direct contact with the target. It not only has the ability to obtain spectral information of ground objects at a long distance and over a large range, but also can deeply study surface objects, providing geologists with a unique perspective and way of thinking. Remote sensing images can not only obtain the morphology and distribution characteristics of the surface landscape, but also obtain the detailed composition and structure of the objects, thereby achieving a more profound identification of objects. In geological interpretation work, remote sensing technology is widely used, which not only provides geologists with a new working perspective, but also provides a clever solution to geological problems when ground surveys are limited. Its application can not only efficiently save the time and a large amount of capital investment required for survey work, but also effectively avoid the geological dangers that may be caused by obstruction of the survey route, reduce the labor intensity of survey work, and become an important means to enhance the soft power of geological work. The landslide mapping methods based on remote sensing technology are roughly divided into methods based on manual interpretation and methods based on intelligent recognition. The manual interpretation method is the most commonly used method since the birth of remote sensing technology.

[0004] Deep learning is an emerging research direction in the field of machine learning. It aims to simulate the operating mechanism of the human brain through deep neural networks and solve various problems through learning. In the semantic segmentation of remote sensing images, deep learning has demonstrated a strong feature learning ability, which gradually deepens from shallow to deep layers, gradually extracts shallow features such as image shape and color, and finally realizes the abstraction of deep features such as category information and semantic information, thereby achieving accurate semantic segmentation of remote sensing images. Deep learning not only solves the problem that traditional image segmentation methods rely on manual feature selection, but also greatly improves recognition efficiency, making target recognition gradually tend to semi-automatic or even fully automated.

[0005] With the innovation of remote sensing technology, remote sensing images are obtained through images from various sensors and sources, including various optical remote sensing images. A series of features (such as shape, size, color, texture, pattern, terrain, etc.) can be used as standards and references for identifying landslides. However, the results of manual interpretation are easily affected by the subjectivity of operators. In addition, interpreting large-scale landslides requires extremely high labor and time costs. In terms of intelligent identification, in recent years, researchers have spent a lot of time training artificial intelligence models to map landslides through remote sensing images. Landslide mapping usually requires dual-phase remote sensing images obtained by the same sensor. This demand may be difficult to meet immediately after an emergency occurs, thus affecting work efficiency and emergency response capabilities.

[0006] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0007] In view of the problems in the related art, the present invention proposes a landslide identification method based on multi-source remote sensing image fusion to overcome the above technical problems existing in the existing related art.

[0008] To this end, the specific technical solution adopted by the present invention is as follows:

[0009] A landslide identification method based on multi-source remote sensing image fusion, the landslide identification method based on multi-source remote sensing image fusion comprises the following steps:

[0010] S1, select multi-source remote sensing images, construct vectorized labels, and obtain binary label data through vectorized labels;

[0011] S2, using a sliding window to crop multi-source remote sensing images and binary label data, and using computer vision enhancement technology to process multi-source remote sensing images and binary label data to obtain a data set;

[0012] S3. Based on the MiT architecture, a landslide intelligent identification model is constructed and trained according to the obtained data set;

[0013] S4. Based on the trained landslide intelligent recognition model, the prediction is implemented using the edge ignoring prediction method, and the multi-source remote sensing image recognition results are obtained according to the prediction results;

[0014] S5. Define the geospatial coordinates according to the multi-source remote sensing image recognition results, and perform binary label vectorization to obtain the landslide vector label.

[0015] Furthermore, selecting multi-source remote sensing images, constructing vectorized labels, and obtaining binary label data through vectorized labels include the following steps:

[0016] S11, selecting multi-source remote sensing images, and vectorizing and labeling the landslide areas in the multi-source remote sensing images to construct vectorized labels;

[0017] S12. Interpret multi-source remote sensing images through ArcGIS software and combine them with vectorized labels to obtain vector label data;

[0018] S13. Use the GDAL library to convert the acquired vector label data into binary label data.

[0019] Furthermore, the multi-source remote sensing images and binary label data are cropped using a sliding window, and the multi-source remote sensing images and binary label data are processed using computer vision enhancement technology to obtain a data set, including the following steps:

[0020] S21. Based on the GDAL library, multi-source remote sensing images and binary label data are cropped into remote sensing tiles using sliding windows;

[0021] S22, performing geometric processing on the cropped remote sensing tiles to obtain a geometric data set;

[0022] S23. Enhance the obtained geometric data set by using the Open-CV library to obtain a data set.

[0023] Furthermore, the remote sensing tile size corresponding to the multi-source remote sensing image is 1024*1024*3;

[0024] The remote sensing tile size corresponding to the binary label data is 1024*1024*1.

[0025] Furthermore, based on the MiT architecture, a landslide intelligent identification model is constructed. Training the landslide intelligent identification model according to the obtained data set includes the following steps:

[0026] S31. Based on the MiT architecture, a GRN normalization model is constructed using the global normalization method.

[0027] S32. Based on the constructed GRN normalization model, the MiT-B4 architecture is used as the backbone network to construct the LDGFormer algorithm;

[0028] S33. Based on the constructed LDGFormer algorithm, the landslide intelligent identification model was constructed using the Cross-Entropy loss function and the adamW optimizer;

[0029] S34. Based on the obtained data set, the landslide intelligent identification model is trained.

[0030] Furthermore, the calculation formula of the global normalization method is:

[0031] ,

[0032] ,

[0033] ,

[0034] In the formula, Represented as normalized features;

[0035] Expressed as The normalized features of the matrix;

[0036] Represented as a set of matrices;

[0037] Represented as a set of matrices in one-dimensional space;

[0038] Represented as a set of matrices in three-dimensional space;

[0039] Indicated as attached to Independent matrices of a matrix set;

[0040] Represented as the height of a matrix;

[0041] Represented as the width of the matrix;

[0042] Represented as the length of the matrix;

[0043] Expressed as The characteristics of each channel;

[0044] Expressed as The normalized features of the matrix;

[0045] It is represented as the regularized channels;

[0046] Represented as the matrix set matrices;

[0047] Represented as the matrix set matrices;

[0048] Represented as the matrix set matrices;

[0049] Represented as the final output matrix;

[0050] Indicates a parameter to prevent division by zero;

[0051] Indicates a parameter to prevent division by zero.

[0052] Furthermore, based on the constructed GRN normalization model, the MiT-B4 architecture is used as the backbone network to construct the LDGFormer algorithm, which includes the following steps:

[0053] S321, divide the multi-source remote sensing image into several image blocks through the MiT-B4 architecture;

[0054] S322, using the MiT-B4 architecture to process a number of image blocks, and obtaining information interaction between the MiT-B4 architecture and the image blocks;

[0055] S323, exchanging information between a number of image blocks and extracting image features;

[0056] S324, mapping the image features to the category distribution probability, and performing classification prediction;

[0057] S325. Combine the constructed GRN normalization model to construct the LDGFormer algorithm.

[0058] Furthermore, the formula of the Cross-Entropy loss function is:

[0059] ,

[0060] In the formula, represents the loss function;

[0061] represents the number of samples;

[0062] Indicates belonging to The true label value of the sample;

[0063] Indicates belonging to The predicted label value of the sample;

[0064] Represents a natural number.

[0065] Furthermore, based on the trained landslide intelligent recognition model, the prediction is implemented using the edge ignoring prediction method, and the multi-source remote sensing image recognition results are obtained according to the prediction results, including the following steps:

[0066] S41. Use the GDAL library to establish an edge-ignoring prediction method;

[0067] S42, cropping multi-source remote sensing images and predicting using edge-ignoring prediction method;

[0068] S43, based on the prediction results, using morphological filtering algorithm to filter the multi-source remote sensing images;

[0069] S44. Combined with the hole filling algorithm in the Open-CV library, the multi-source remote sensing image after filtering is repaired to obtain the recognition result of the multi-source remote sensing image.

[0070] Furthermore, defining the geospatial coordinates according to the multi-source remote sensing image recognition results and performing binary label vectorization to obtain the landslide vector label includes the following steps:

[0071] S51. Use the GDAL library to perform geospatial mapping between the recognition results of the multi-source remote sensing images and the multi-source remote sensing images, and define the spatial coordinate system of the multi-source remote sensing images;

[0072] S52, performing vectorization conversion on the defined multi-source remote sensing image, deleting the background vector value of non-landslide, and obtaining the final landslide identification label.

[0073] The beneficial effects of the present invention are:

[0074] 1. The present invention establishes an LDGFormer algorithm for landslide identification through a GRN normalization model based on a global normalization method and combined with a MiT architecture. The LDGFormer algorithm can not only better capture the details and shape characteristics of landslides, but also significantly improve the recognition ability of the landslide intelligent recognition model for landslides under different terrain and geomorphic conditions.

[0075] 2. The present invention uses computer vision enhancement technology and combines high-resolution remote sensing images from multiple data sources for data training to obtain an intelligent landslide recognition model with good segmentation effect, so that it can effectively perform inference calculations on high-resolution remote sensing images, identify and extract landslide information therefrom, and help achieve the goal of automatically identifying and accurately interpreting large-scale landslides induced by earthquakes or extreme climatic conditions in high-resolution remote sensing images, improve the response speed and accuracy to disaster situations, and thus protect people’s lives and property.

[0076] 3. The present invention ignores the edge prediction method, combines the morphological filtering algorithm with the hole filling algorithm, effectively removes the noise points existing in the multi-source remote sensing images, effectively improves the clarity and accuracy of the data, and lays a good foundation for the subsequent landslide identification work; at the same time, by using the GDAL definition to identify the geographic spatial coordinates, the final landslide vector label is obtained, which further improves the data processing efficiency and availability. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0078] Figure 1 is a flow chart of a landslide identification method based on multi-source remote sensing image fusion according to an embodiment of the present invention;

[0079] Figure 2 is an example diagram of a landslide image sample block in a landslide identification method based on multi-source remote sensing image fusion according to an embodiment of the present invention;

[0080] Figure 3 is an example diagram of landslide sample interpretation labels corresponding to landslide image sample blocks in a landslide identification method based on multi-source remote sensing image fusion according to an embodiment of the present invention;

[0081] Figure 4 is a structural schematic diagram of an LDGFormer algorithm in a landslide identification method based on multi-source remote sensing image fusion according to an embodiment of the present invention;

[0082] Figure 5 is a structural schematic diagram of a GRN normalization model in a landslide identification method based on multi-source remote sensing image fusion according to an embodiment of the present invention;

[0083] Figure 6is a schematic diagram of a method for ignoring edge prediction in a landslide identification method based on multi-source remote sensing image fusion according to an embodiment of the present invention;

[0084] Figure 7 is a schematic diagram of a landslide identification result in a landslide identification method based on multi-source remote sensing image fusion according to an embodiment of the present invention;

[0085] Figure 8 It is a schematic diagram of landslide label results obtained in a landslide identification method based on multi-source remote sensing image fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0086] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention and are mainly used to illustrate the embodiments. They can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, ordinary technicians in this field should be able to understand other possible implementation methods and advantages of the present invention.

[0087] According to an embodiment of the present invention, a landslide identification method based on multi-source remote sensing image fusion is provided.

[0088] The present invention is further described with reference to the accompanying drawings and specific embodiments. Figure 1-Figure 3 As shown, a landslide identification method based on multi-source remote sensing image fusion according to an embodiment of the present invention comprises the following steps:

[0089] S1. Select multi-source remote sensing images, construct vectorized labels, and obtain binary label data through vectorized labels.

[0090] In the description of the present invention, selecting multi-source remote sensing images, constructing vectorized labels, and obtaining binary label data through the vectorized labels include the following steps:

[0091] S11. Select multi-source remote sensing images, perform vector annotation on landslide areas in the multi-source remote sensing images, and construct vector labels.

[0092] It should be added that the vectorized labels were constructed manually with large-scale landslides near human settlements and infrastructure projects, which were triggered by rainfall or earthquakes, as the main targets.

[0093] S12. Interpret multi-source remote sensing images through ArcGIS software and combine them with vectorized labels to obtain vector label data.

[0094] S13. Use the GDAL library to convert the acquired vector label data into binary label data.

[0095] It should be added that the GDAL library is used to convert the acquired vector label data into binary label data consistent with the corresponding multi-source remote sensing images, with the same spatial resolution, the landslide pixel value is set to 1, the background pixel value is set to 0, and the geospatial coordinate system is WGS1984.

[0096] S2. Use a sliding window to crop multi-source remote sensing images and binary label data, and use computer vision enhancement technology to process multi-source remote sensing images and binary label data to obtain a data set.

[0097] In the description of the present invention, using a sliding window to crop multi-source remote sensing images and binary label data, and using computer vision enhancement technology to process the multi-source remote sensing images and binary label data to obtain a data set includes the following steps:

[0098] S21. Based on the GDAL library, a sliding window is used to crop multi-source remote sensing images and binary label data into remote sensing tiles.

[0099] In the description of the present invention, the remote sensing tile size corresponding to the multi-source remote sensing image is 1024*1024*3;

[0100] The remote sensing tile size corresponding to the binary label data is 1024*1024*1.

[0101] S22. Perform geometric processing on the cropped remote sensing tiles to obtain a geometric data set.

[0102] It should be noted that the geometric processing includes performing geometric scale symmetry, flipping and rotation transformation on the cropped remote sensing tiles.

[0103] S23. Enhance the obtained geometric data set by using the Open-CV library to obtain a data set.

[0104] It should be noted that the enhancement processing includes spectral stretching, color change, and contrast transformation of the geometric data set.

[0105] S3. Based on the MiT architecture, a landslide intelligent identification model is constructed and trained according to the obtained data set.

[0106] In the description of the present invention, based on the MiT architecture, a landslide intelligent recognition model is constructed, and training the landslide intelligent recognition model according to the obtained data set includes the following steps:

[0107] S31. Based on the MiT architecture, a GRN normalization model is constructed using the global normalization method.

[0108] In the description of the present invention, the calculation formula of the global normalization method is:

[0109] ,

[0110] ,

[0111] ,

[0112] In the formula, Represented as normalized features;

[0113] Expressed as The normalized features of the matrix;

[0114] Represented as a set of matrices;

[0115] Represented as a set of matrices in one-dimensional space;

[0116] Represented as a set of matrices in three-dimensional space;

[0117] Indicated as attached to Independent matrices of a matrix set;

[0118] Represented as the height of a matrix;

[0119] Represented as the width of the matrix;

[0120] Represented as the length of the matrix;

[0121] Expressed as The characteristics of each channel;

[0122] Expressed as The normalized features of the matrix;

[0123] It is represented as the regularized channels;

[0124] Represented as the matrix set matrices;

[0125] Represented as the matrix set matrices;

[0126] Represented as the matrix set matrices;

[0127] Represented as the final output matrix;

[0128] Indicates a parameter to prevent division by zero;

[0129] Indicates a parameter to prevent division by zero.

[0130] It should be noted that the GRN normalization model is Figure 5 As shown in the figure, F1, F2, F3, and F4 represent the fully connected layers after each encoder, and GRN represents the global normalization model.

[0131] S32. Based on the constructed GRN normalized model, the MiT-B4 architecture is used as the backbone network to construct the LDGFormer algorithm.

[0132] It should be noted that the MiT-B4 architecture contains four groups of basic modules, each of which consists of image segmentation, visual attention module, hybrid mechanism module, classification head and multi-layer Transformer encoder.

[0133] In the description of the present invention, based on the constructed GRN normalization model, the MiT-B4 architecture is used as the backbone network, and the construction of the LDGFormer algorithm includes the following steps:

[0134] S321. Divide the multi-source remote sensing image into several image blocks through the MiT-B4 architecture.

[0135] It should be noted that the input image is divided into multiple small blocks through image segmentation, each small block is usually a pixel area of ​​a fixed size, and the image blocks are used as input sequences.

[0136] S322. Use the MiT-B4 architecture to process several image blocks and obtain information interaction between the MiT-B4 architecture and the image blocks.

[0137] It should be noted that the visual attention module enables the MiT-B4 architecture to dynamically focus on and establish complex relationships between image blocks when processing image block sequences.

[0138] S323, exchanging information between several image blocks and extracting image features.

[0139] It should be noted that the hybrid mechanism module enables information exchange between different image blocks to promote global feature learning and representation, which helps the model better capture the global structure and semantic information of the image.

[0140] S324, mapping the image features to the category distribution probability and performing classification prediction.

[0141] It should be noted that the MiT-B4 architecture is usually composed of multiple Transformer encoders, each of which contains multiple attention heads and feedforward neural networks for feature extraction and representation learning of image block sequences at different levels of abstraction. A classification head for image classification tasks is usually added, which is responsible for mapping the extracted image features to the category probability distribution for classification prediction.

[0142] S325. Combine the constructed GRN normalization model to construct the LDGFormer algorithm.

[0143] It should be noted that the LDGFormer algorithm is Figure 4 As shown in the figure, H and W represent the height and width of the matrix; C, C1, C2, C3, and C4 represent the lengths of different matrices; MatMul represents the normalization layer; Softmax represents the classification activation layer; F1, F2, F3, and F4 represent the fully connected layers after each encoder; and GRN represents the global normalization model.

[0144] S33. Based on the constructed LDGFormer algorithm, the landslide intelligent identification model is constructed using the Cross-Entropy loss function and adamW optimizer.

[0145] In the description of the present invention, the formula of the Cross-Entropy loss function is:

[0146] ,

[0147] In the formula, represents the loss function;

[0148] represents the number of samples;

[0149] Indicates belonging to The true label value of the sample;

[0150] Indicates belonging to The predicted label value of the sample;

[0151] Represents a natural number.

[0152] S34. Based on the obtained data set, the landslide intelligent identification model is trained.

[0153] S4. Based on the trained landslide intelligent recognition model, the prediction is implemented using the edge ignoring prediction method, and the multi-source remote sensing image recognition results are obtained based on the prediction results.

[0154] In the description of the present invention, based on the trained landslide intelligent recognition model, the prediction is implemented using the edge ignoring prediction method, and the multi-source remote sensing image recognition result is obtained according to the prediction result, which includes the following steps:

[0155] S41. Use the GDAL library to establish an edge-ignoring prediction method.

[0156] S42. Crop multi-source remote sensing images and use the edge ignoring prediction method for prediction.

[0157] It should be noted that the edge-ignoring prediction method is to crop multi-source remote sensing images with overlap and adopt the edge-ignoring strategy when stitching. In the prediction step, a large image file in an independent geographic space is directly input to perform result prediction. Figure 6 As shown, A represents the circumscribed rectangle of the sliding window cropping solution, and a represents the inscribed rectangle of the sliding window cropping solution.

[0158] S43. Based on the prediction results, a morphological filtering algorithm is used to filter the multi-source remote sensing images.

[0159] S44. Combined with the hole filling algorithm in the Open-CV library, the multi-source remote sensing image after filtering is repaired to obtain the recognition result of the multi-source remote sensing image.

[0160] S5. Define the geospatial coordinates according to the multi-source remote sensing image recognition results, and perform binary label vectorization to obtain the landslide vector label.

[0161] In the description of the present invention, defining geographic spatial coordinates according to multi-source remote sensing image recognition results and performing binary label vectorization to obtain landslide vector labels includes the following steps:

[0162] S51. Use the GDAL library to perform geospatial mapping between the recognition results of the multi-source remote sensing images and the multi-source remote sensing images, and define the spatial coordinate system of the multi-source remote sensing images;

[0163] S52, performing vectorization conversion on the defined multi-source remote sensing image, deleting the background vector value of non-landslide, and obtaining the final landslide identification label.

[0164] It should be noted that the landslide label results obtained are as follows: Figure 7-Figure 8 As shown, Figure 8 The line pointed by the arrow in the middle represents the landslide identification label (landslide label), and the area surrounded by the line is the landslide area.

[0165] In summary, with the help of the above technical solution of the present invention, a LDGFormer algorithm for landslide identification is established through a GRN normalization model based on a global normalization method and combined with a MiT architecture. The LDGFormer algorithm can not only better capture the details and shape characteristics of landslides, but also significantly improve the recognition ability of landslide intelligent recognition models for landslides under different terrain and geomorphic conditions. Through computer vision enhancement technology, high-resolution remote sensing images from multiple data sources are combined for data training to obtain a landslide intelligent recognition model with good segmentation effect, so that high-resolution remote sensing images can be effectively inferred and calculated, and landslide information can be identified and extracted therefrom, which helps to achieve the goal of automatically identifying and interpreting large-scale landslides induced by earthquakes or extreme climate conditions in high-resolution remote sensing images with high precision, improve the response speed and accuracy to disaster situations, and thus protect people's lives and property. By ignoring the edge prediction method and combining the morphological filtering algorithm with the hole filling algorithm, the noise points in the multi-source remote sensing images can be effectively removed, which effectively improves the clarity and accuracy of the data and lays a good foundation for the subsequent landslide identification work; at the same time, by using the GDAL defined identified geographic spatial coordinates, the final landslide vector label is obtained, which further improves the data processing efficiency and availability.

[0166] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A landslide identification method based on multi-source remote sensing image fusion, characterized in that: The landslide identification method based on multi-source remote sensing image fusion includes the following steps: S1, select multi-source remote sensing images, construct vectorized labels, and obtain binary label data through vectorized labels; S2, using a sliding window to crop multi-source remote sensing images and binary label data, and using computer vision enhancement technology to process multi-source remote sensing images and binary label data to obtain a data set; S3. Based on the MiT architecture, a landslide intelligent identification model is constructed and trained according to the obtained data set; S4. Based on the trained landslide intelligent recognition model, the prediction is implemented using the edge ignoring prediction method, and the multi-source remote sensing image recognition results are obtained according to the prediction results; S5, defining geographic spatial coordinates according to the multi-source remote sensing image recognition results, and performing binary label vectorization to obtain landslide vector labels; The construction of a landslide intelligent recognition model based on the MiT architecture and training of the landslide intelligent recognition model according to the obtained data set include the following steps: S31. Based on the MiT architecture, a GRN normalization model is constructed using the global normalization method. S32. Based on the constructed GRN normalization model, the MiT-B4 architecture is used as the backbone network to construct the LDGFormer algorithm; S33. Based on the constructed LDGFormer algorithm, the landslide intelligent identification model was constructed using the Cross-Entropy loss function and the adamW optimizer; S34, training a landslide intelligent identification model based on the obtained data set; The calculation formula of the global normalization method is: , , , In the formula, Represented as normalized features; Expressed as The normalized features of the matrix; Represented as a set of matrices; Represented as a set of matrices in one-dimensional space; Represented as a set of matrices in three-dimensional space; Indicated as attached to Independent matrices of a matrix set; Represented as the height of a matrix; Represented as the width of the matrix; Represented as the length of the matrix; Expressed as The characteristics of each channel; Expressed as The normalized features of the matrix; It is represented as the regularized channels; Represented as the matrix set matrices; Represented as the matrix set matrices; Represented as the matrix set matrices; Represented as the final output matrix; Indicates a parameter to prevent division by zero; Indicates a parameter to prevent division by zero; The GRN normalization model constructed based on the MiT-B4 architecture is used as the backbone network to construct the LDGFormer algorithm, including the following steps: S321, divide the multi-source remote sensing image into several image blocks through the MiT-B4 architecture; S322, using the MiT-B4 architecture to process a number of image blocks, and obtaining information interaction between the MiT-B4 architecture and the image blocks; S323, exchanging information between a number of image blocks and extracting image features; S324, mapping the image features to the category distribution probability, and performing classification prediction; S325. Combine the constructed GRN normalization model to construct the LDGFormer algorithm; The formula of the Cross-Entropy loss function is: , In the formula, represents the loss function; represents the number of samples; Indicates belonging to The true label value of the sample; Indicates belonging to The predicted label value of the sample; Represents a natural number.

2. The landslide identification method based on multi-source remote sensing image fusion according to claim 1 is characterized in that: The method of selecting multi-source remote sensing images, constructing vectorized labels, and obtaining binary label data through the vectorized labels includes the following steps: S11, selecting multi-source remote sensing images, and vectorizing and labeling the landslide areas in the multi-source remote sensing images to construct vectorized labels; S12. Interpret multi-source remote sensing images through ArcGIS software and combine them with vectorized labels to obtain vector label data; S13. Use the GDAL library to convert the acquired vector label data into binary label data.

3. The landslide identification method based on multi-source remote sensing image fusion according to claim 1 is characterized in that: The method of using a sliding window to crop multi-source remote sensing images and binary label data, and using computer vision enhancement technology to process the multi-source remote sensing images and binary label data to obtain a data set includes the following steps: S21. Based on the GDAL library, multi-source remote sensing images and binary label data are cropped into remote sensing tiles using sliding windows; S22, performing geometric processing on the cropped remote sensing tiles to obtain a geometric data set; S23. Enhance the obtained geometric data set by using the Open-CV library to obtain a data set.

4. The landslide identification method based on multi-source remote sensing image fusion according to claim 3 is characterized in that: The remote sensing tile size corresponding to the multi-source remote sensing image is 1024*1024*3; The remote sensing tile size corresponding to the binary label data is 1024*1024*1.

5. The landslide identification method based on multi-source remote sensing image fusion according to claim 1 is characterized in that: The method of implementing prediction based on the trained landslide intelligent recognition model by using the edge ignoring prediction method and obtaining the multi-source remote sensing image recognition result according to the prediction result comprises the following steps: S41. Use the GDAL library to establish an edge-ignoring prediction method; S42, cropping multi-source remote sensing images and predicting using edge-ignoring prediction method; S43, based on the prediction results, using morphological filtering algorithm to filter the multi-source remote sensing images; S44. Combined with the hole filling algorithm in the Open-CV library, the multi-source remote sensing image after filtering is repaired to obtain the recognition result of the multi-source remote sensing image.

6. The landslide identification method based on multi-source remote sensing image fusion according to claim 1 is characterized in that: Defining geographic spatial coordinates according to multi-source remote sensing image recognition results and performing binary label vectorization to obtain landslide vector labels includes the following steps: S51. Use the GDAL library to perform geospatial mapping between the recognition results of the multi-source remote sensing images and the multi-source remote sensing images, and define the spatial coordinate system of the multi-source remote sensing images; S52, performing vectorization conversion on the defined multi-source remote sensing image, deleting the background vector value of non-landslide, and obtaining the final landslide identification label.

Images