Landslide remote sensing image semantic segmentation method based on bilateral segmentation network

Abstract

The invention relates to the technical field of semantic segmentation, in particular to a bilateral segmentation network-based landslide remote sensing image semantic segmentation method, which comprises the following steps of: S2, constructing a detail branch landslide image feature extraction network, and extracting a feature map of detail branches; s3, establishing a landslide semantic branch feature extraction network, and extracting a feature map of semantic branches; and S4, fusing the detail branches and the semantic branches by adopting a bilateral aggregation guide layer to form a landslide semantic segmentation model, and performing landslide semantic segmentation. According to the scheme, the problems that the multi-scale feature extraction and fusion capability of a traditional convolutional neural network architecture model is weak, the mined image semantic information is poor and the long-distance dependency relationship is difficult to capture can be solved, so that the landslide semantic segmentation error rate is reduced, and the accuracy and recognition efficiency of a segmentation result are improved.

Description

Technical Field

[0001] The invention relates to the technical field of semantic segmentation, and in particular to a semantic segmentation method of landslide remote sensing images based on a bilateral segmentation network. Background Art

[0002] Landslide is a common geological disaster that threatens people's lives and property. In order to prevent the harm caused by landslides, landslide risks are identified through monitoring and analysis. In the existing technology, landslide semantic segmentation is often used to accurately separate the landslide area from other landforms, providing important information for the geological disaster early warning system, which helps to timely discover and monitor potential landslide risks, thereby reminding people to take corresponding prevention and control measures.

[0003] With the development of artificial intelligence and computer vision technology, deep learning has been widely used in the field of landslide semantic segmentation and has achieved remarkable results. In the task of landslide semantic segmentation, deep learning models can automatically learn the semantic information in images and achieve accurate pixel-level segmentation. Traditional landslide semantic segmentation networks include classic convolutional neural network architectures, such as fully connected neural networks, U-Net, SegNet, etc., which mainly achieve pixel-level semantic segmentation through convolution and upsampling operations.

[0004] However, the traditional landslide semantic segmentation network structure usually only uses a single scale information for segmentation, and has weak multi-scale feature extraction and fusion capabilities for landslide areas. When extracting features, the existing deep learning model only uses simple convolution or pooling operations, which cannot fully mine the semantic information in the image. In addition, the traditional convolutional neural network structure has difficulty in feature extraction when dealing with long-distance dependencies, and landslide areas usually have a large spatial range, which requires considering the long-distance relationship between pixels. Summary of the invention

[0005] The present invention aims to provide a semantic segmentation method for landslide remote sensing images based on a bilateral segmentation network, which can solve the problems of weak multi-scale feature extraction and fusion capabilities of traditional convolutional neural network architecture models, poor image semantic information mining and difficulty in capturing long-distance dependencies, so as to reduce the error rate of landslide semantic segmentation and improve the accuracy and recognition efficiency of segmentation results.

[0006] The present invention provides the following basic scheme: a semantic segmentation method of landslide remote sensing images based on a bilateral segmentation network, including the following contents:

[0007] S2, constructing a detail branch landslide image feature extraction network to extract a feature map of the detail branch;

[0008] S3, establishing a landslide semantic branch feature extraction network to extract feature maps of semantic branches;

[0009] S4. A bilateral aggregation guided layer is used to fuse the detail branch and the semantic branch to form a landslide semantic segmentation model for landslide semantic segmentation.

[0010] Furthermore, it also includes: S1, constructing a landslide semantic segmentation dataset;

[0011] Said S1 comprises:

[0012] A landslide dataset is obtained for experiments, wherein the landslide dataset includes: images, shape files of landslide boundaries and digital elevation models; the images include landslide images and their corresponding labels;

[0013] Preprocess the image; the preprocessing includes: adjusting the resolution and performing data enhancement;

[0014] The landslide dataset was normalized and standardized, and the digital elevation model and image were fused.

[0015] Furthermore, the detail branch landslide image feature extraction network includes: a plurality of convolutional layers, each convolutional layer is connected in sequence, and the number of output channels of the convolutional layer is half of the number of output channels of the previous convolutional layer.

[0016] Further, the S2 includes:

[0017] S201, performing detail normalization processing on the input multi-channel landslide image:

[0018]

[0019] Among them, x ijkl is the index value of the image at (i, j, k, l), B is the batch size, C is the number of channels, H is the image length, W is the image width, μ j is the batch mean, σ j is the standard deviation, and ε is a value ranging from 10 -6 to 10 -12 A positive number between

[0020] The detail normalized feature values ​​are:

[0021]

[0022] S202, the landslide image after detail normalization is convolved through sequentially connected convolutional layers, and the number of output channels of each convolutional layer is half of the number of input channels.

[0023] Furthermore, the landslide semantic branch feature extraction network includes a backbone module and a context embedding module;

[0024] The backbone module is used to process the input image and increase the number of channels while reducing the spatial dimension; the input image is a landslide image in the landslide semantic segmentation dataset;

[0025] The context embedding module is used to capture the context information of the input image and generate surface features; the input image is the feature map output by the backbone module;

[0026] The backbone module processes the input image and increases the number of channels while reducing the spatial dimension. The specific calculation steps are as follows:

[0027] Perform c on the input image 1 ×c 1 Convolution operation:

[0028]

[0029] X 1 Perform batch normalization and use the ReLU activation function for calculation:

[0030] X 2 =ReLU(BN(X 1 ));

[0031] Among them, BN is the batch normalization operation, and the formula is:

[0032]

[0033] Among them, μ and σ 2 are the mean and variance respectively, γ and β are learning parameters;

[0034] X 2 Carry out 2 ×c 2 Convolution operation:

[0035]

[0036] X 3 Perform batch normalization and use the ReLU activation function for calculation:

[0037] X 4 =ReLU(BN(X 3 ));

[0038] X 4 Carry out 4 ×c 4 Convolution operation:

[0039]

[0040] X 5 Perform batch normalization and use the ReLU activation function for calculation:

[0041] X 6 =ReLU(BN(X 5 ));

[0042] X 2 Carry out 7 ×c 7 The maximum pooling operation:

[0043]

[0044] X 6 And the output X of the maximum pooling layer 7 To connect:

[0045] X 8 =Concat(X 6 ,X 7 );

[0046] X 8 Carry out 9 ×c 9 Convolution operation:

[0047]

[0048] X 9 Perform batch normalization and use the ReLU activation function to calculate and obtain the final output:

[0049] Output = ReLU(BN(X 9 ));

[0050] Through the context embedding module, the context information of the input image is captured to generate surface features. The specific calculation steps are as follows:

[0051] Perform GAPooling global average pooling calculation on the input feature map:

[0052]

[0053] Among them, f(i,j,c) is the value of the cth channel of the input feature map at position (i,j), f gap (c) is the value of the cth channel after pooling;

[0054] Perform batch normalization on the pooled data:

[0055] X bn =BN(f gap (c));

[0056] X bn Carry out conv1 ×c conv1Convolution operation:

[0057]

[0058] X conv1 Perform batch normalization and use the ReLU activation function for calculation:

[0059] X relu =ReLU(X conv1 );

[0060] X after broadcast relu Add to the input feature map f(i,j,c):

[0061] X sum =X relu +f(i,j,c);

[0062] X sum Carry out output ×c output Convolution operation:

[0063]

[0064] Further, the S4 comprises:

[0065] S401, performing depth-separable convolution calculation on the feature map of the detail branch extracted by the detail branch landslide image feature extraction network, and then performing calculation by multi-scale parallel dilated convolution:

[0066]

[0067] Among them, Y is the calculation result of multi-scale parallel dilated convolution, K i is the i-th convolution kernel, X is the calculation result of the depthwise separable convolution, and N is the number of parallel convolution kernels;

[0068] Perform element-wise multiplication of Y with the upsampled result of the semantic branch:

[0069] X de_mul =Y×X upsam ;

[0070] Among them, X de_mul is the result after convolution of detail branch, X upsam This is the result after upsampling of the semantic branch;

[0071] S402, perform a depth-separable convolution on the feature map of the semantic branch extracted by the landslide semantic branch feature extraction network, and then perform calculations through multi-scale parallel dilated convolutions, and then perform a Sigmoid function on the result X after the detail branch convolution. semProcess it and perform an element-wise product with the average pooling result of the detail branch:

[0072] X sem_mul =Sigmoid(X sem )×X apooling ;

[0073] Among them, X sem_mul is the result after convolution of detail branch, X apooling It is the result after average pooling of semantic branch;

[0074] S403, adding the operation result of the detail branch in S401 and the operation result of the semantic branch in S402 pixel by pixel, and performing a convolution operation on the addition result to obtain the final output result.

[0075] Furthermore, it also includes: S5, using the landslide semantic segmentation dataset to train and test the landslide semantic segmentation model, including:

[0076] S501, setting parameters of a landslide semantic segmentation model, and using a training set of a landslide semantic segmentation data set to train the landslide semantic segmentation model;

[0077] S502, performing semantic segmentation on a test set of the landslide semantic segmentation dataset using the fitted landslide semantic segmentation model to obtain a segmentation result;

[0078] S503: Evaluate the training result of the landslide semantic segmentation model according to the evaluation index. If the evaluation index meets the preset requirements, output the landslide semantic segmentation model.

[0079] Furthermore, it also includes:

[0080] S6. Use the landslide semantic segmentation model to identify the landslide contours, and calculate NDVI to analyze the vegetation coverage and growth status to assist in landslide identification, including:

[0081] Acquire remote sensing image data in near-infrared and red bands including landslide contours;

[0082] According to the remote sensing data source, select the corresponding near-infrared band and red light band;

[0083] The NDVI value is calculated according to the formula NDVI = (NIR-R) / (NIR+R), where NIR is the reflectance of the near-infrared band and R is the reflectance of the red light band; and the reflectance of the two bands is normalized during the calculation process to obtain a value between -1 and 1;

[0084] The calculated NDVI values ​​are visualized and statistically analyzed, and the analysis data are extracted to assist in identifying landslides; the larger the NDVI value, the higher the vegetation coverage.

[0085] Further, it also includes: S7, using a digital high-rise model to analyze the slope of the landslide contour map, including:

[0086] Get a digital elevation model dataset; a digital elevation model dataset is a raster dataset in which each pixel represents the elevation of a specific point on the surface;

[0087] Calculate the slope of the raster data and obtain a new raster dataset, in which the value of each pixel represents the slope at that point;

[0088] The possibility of landslide is determined by the slope. If the slope is greater than 10 degrees and less than 45 degrees, the possibility of landslide is judged to be high.

[0089] Further, the method further includes: S8, monitoring rainfall conditions, obtaining rainfall parameters, and adjusting the landslide assignment parameters according to the rainfall conditions and rainfall parameters, including:

[0090] Analyze whether there is rainfall within the preset time before the current time point. If not, the landslide risk in the area is judged to be low, and the landslide parameter assignment of the area is reduced according to the preset reduction ratio;

[0091] If yes, then determine whether the rainfall is greater than the preset rainfall, the rainfall intensity is greater than the preset rainfall intensity, and the rainfall time is greater than the preset rainfall time. If any of the judgments is yes, then increase the assigned parameter of the landslide in the area according to the preset increase ratio;

[0092] The preset rising ratio is the adjustment ratio P0;

[0093] P0=w1×P1+w2×P2+w3×P3

[0094] Among them, P0 represents the adjusted probability coefficient of landslide occurrence, P1, P2 and P3 represent rainfall, rainfall intensity and rainfall time respectively, and w1, w2 and w3 represent the weights of rainfall, rainfall intensity and rainfall time respectively.

[0095] Further, it also includes: S9, constructing a comprehensive judgment model for identifying and distinguishing landslide areas and construction sites by combining the change rate of landslide edges, the change rate of pixels in suspected areas, and human activity trajectories;

[0096] A comprehensive judgment model is constructed to identify and distinguish landslide areas and construction sites by combining the change rate of landslide edges, the change rate of pixels in suspected areas, and the trajectory of human activities.

[0097] Specifically include:

[0098] Conduct landslide edge change rate analysis, including:

[0099] By analyzing the remote sensing images of continuous time series, the position change rate of the pixel points at the edge of the landslide is calculated;

[0100] Image processing technology is used to track the movement trajectory of pixels at the edge of the landslide in multiple images;

[0101] Analyze the position change rate of the pixel points at the edge of the landslide, identify the area where the landslide is active, and use it as the suspected landslide area. The active area of ​​the landslide is characterized by the movement rate of the pixel points at the edge of the landslide being higher than the preset movement rate.

[0102] Conduct pixel change rate analysis in the suspected landslide area, including:

[0103] Mark suspected landslide areas in remote sensing images;

[0104] Calculate the change rate of the pixels at the edge of the landslide in the suspected landslide area, including changes in brightness, color, and texture;

[0105] According to the change rate of the pixel points at the edge of the landslide, judging whether the landslide area is a landslide area or a construction site, including: judging whether the change rate of the pixel points at the edge of the landslide is greater than a preset change rate and changes continuously, if so, judging it as a landslide area, if the change rate of the pixel points at the edge of the landslide changes intermittently, judging it as a construction site;

[0106] Conduct human activity trajectory analysis, including:

[0107] Collect and analyze data on human activities near the landslide area;

[0108] Remote sensing image data and GIS technology are used to extract information on human activity trajectories, including activity types, scope, and intensity. The change rate of landslide edges, the change rate of pixels in suspected areas, and human activity trajectories are combined to construct a comprehensive judgment model to distinguish between landslide activities and construction sites.

[0109] Training and validation of the comprehensive judgment model, including:

[0110] A remote sensing image dataset containing landslides and construction sites is used to train the comprehensive judgment model;

[0111] Validate the model’s performance through cross-validation and independent test sets;

[0112] The trained comprehensive judgment model is applied, and the application data is obtained as feedback data to adjust and optimize the comprehensive judgment model.

[0113] Beneficial effects: First, this scheme processes the detail features in the landslide image through detail normalization, uses the detail branch of the backbone network to extract the detail information in the landslide image, and captures the detail features such as edges and textures in the landslide image.

[0114] Then, the semantic branch of the backbone network is used to extract the semantic information of the geological structure of the landslide in the landslide image, which helps the network understand the semantic meaning of different areas in the image, thereby improving the ability of the landslide semantic segmentation network (i.e., the landslide semantic segmentation model) to understand the semantic meaning of different areas in the image;

[0115] Finally, the bilateral guided aggregation layer is used to fuse the detail information (feature map of detail branch) and semantic information (feature map of semantic branch) of the landslide, and the multi-scale parallel dilated convolution is used to consider the information at different scales to improve the representation ability and performance of the network; and the segmentation head is used to generate the final landslide image segmentation result. The problems of weak multi-scale feature extraction and fusion capabilities of the traditional convolutional neural network architecture model, poor image semantic information mining, and difficulty in capturing long-distance dependencies are solved to reduce the error rate of landslide semantic segmentation and improve the accuracy and recognition efficiency of segmentation results. BRIEF DESCRIPTION OF THE DRAWINGS

[0116] Figure 1 It is a flow chart of an embodiment of a method for semantic segmentation of landslide remote sensing images based on a bilateral segmentation network according to the present invention;

[0117] Figure 2 It is a structural schematic diagram of the backbone module in an embodiment of the landslide remote sensing image semantic segmentation method based on the bilateral segmentation network of the present invention;

[0118] Figure 3 It is a structural schematic diagram of a context embedding module in an embodiment of a method for semantic segmentation of landslide remote sensing images based on a bilateral segmentation network of the present invention;

[0119] Figure 4 It is a structural schematic diagram of a bilateral aggregation guide layer in an embodiment of a method for semantic segmentation of landslide remote sensing images based on a bilateral segmentation network of the present invention;

[0120] Figure 5 It is a schematic diagram of landslide labels in an embodiment of a method for semantic segmentation of landslide remote sensing images based on a bilateral segmentation network of the present invention;

[0121] Figure 6 It is a schematic diagram of landslide segmentation results in an embodiment of the method for semantic segmentation of landslide remote sensing images based on a bilateral segmentation network of the present invention. DETAILED DESCRIPTION

[0122] The following is further described in detail through specific implementation methods:

[0123] Embodiment 1

[0124] This embodiment is basically as shown in the attached Figure 1 As shown in the figure: The semantic segmentation method of landslide remote sensing image based on bilateral segmentation network includes the following contents:

[0125] S1. Construct a landslide semantic segmentation dataset;

[0126] Specifically, a landslide dataset is obtained for experimentation, wherein the landslide dataset includes: images, shape files of landslide boundaries, and digital elevation models; the images include landslide images and their corresponding labels; in this embodiment, a landslide dataset of Bijie City, Guizhou Province, from the Wuhan University Photogrammetry and Computer Vision Open Source Database is selected for experimentation, and the landslide dataset includes: images, shape files of landslide boundaries, and digital elevation models; wherein the images are satellite optical images, including: 770 landslide images and corresponding labels, all from TripleSat satellite images from May to August 2018;

[0127] Preprocessing the images; the preprocessing includes: adjusting the resolution and performing data enhancement; in this embodiment, the resolution of the images is unified to 256×256, and the original images in the landslide data set are enhanced from the original 770 to 3080 through data enhancement operations such as symmetry, cropping, rotation, and composite transformation;

[0128] The landslide data set is normalized and standardized, and the digital elevation model and the image are fused. In this embodiment, the landslide data set is normalized and standardized, and the digital elevation model and the satellite optical image are fused, and the shape of the fused image is (4, 256, 256).

[0129] S2. Construct a detail branch landslide image feature extraction network to extract a feature map of the detail branch. The detail branch landslide image feature extraction network includes: a plurality of convolutional layers, each of which is connected in sequence, and the number of output channels of the convolutional layer is half of the number of output channels of the previous convolutional layer. The structure of the detail branch landslide image feature extraction network in this embodiment is shown in Table 1:

[0130] Table 1: Basic structure of detail branch

[0131]

[0132]

[0133] The specific construction process is as follows:

[0134] S201, performing detail normalization processing on the input multi-channel landslide image:

[0135]

[0136] Among them, x ijkl is the index value of the image at (i, j, k, l), B is the batch size, H is the image length, W is the image width, μ j is the batch mean, σ j is the standard deviation, and ε is a value ranging from 10 -6 to 10 -12 In floating-point calculations, the value of ε is usually related to the precision of the floating-point number. For single-precision floating-point numbers (32 bits) and double-precision floating-point numbers (64 bits), the value range of ε is 10 -6 to 10 -12 between

[0137] The detail normalized feature values ​​are:

[0138]

[0139] S202, the landslide image after detail normalization is convolved through sequentially connected convolutional layers, and the number of output channels of each convolutional layer is half of the number of input channels, specifically including:

[0140] The landslide image after detail normalization is convolved through the first convolutional layer C1 to obtain a feature map; the number of output channels of this convolutional layer is half of the number of input channels α, and the input image is downsampled;

[0141] The feature map output by the first convolutional layer C1 is input into the second convolutional layer C2 to obtain a feature map with reduced dimension. The number of output channels of C2 is half of the number of output channels of C1, which further reduces the dimension of the feature map.

[0142] The feature map output by the second convolutional layer C2 is input into the third convolutional layer C3 to obtain a feature map with reduced dimension. The number of output channels of C3 is half of the number of output channels of C2, and the dimension of the feature map is further reduced.

[0143] S3, establishing a landslide semantic branch feature extraction network to extract feature maps of semantic branches;

[0144] The landslide semantic branch feature extraction network includes a backbone module and a context embedding module;

[0145] The backbone module is used to process the input image (i.e., landslide image) and increase the number of channels while reducing the spatial dimension; wherein the input image is a landslide image in the landslide semantic segmentation dataset;

[0146] The context embedding module is used to capture the context information of the input image (i.e., landslide image) and generate surface features;

[0147] The specific calculation steps of the two modules are as follows:

[0148] The input image is processed by the backbone module and the number of channels is increased while reducing the spatial dimension. The structure of the backbone module is shown in the attached figure. Figure 2 As shown, the specific calculation steps are as follows:

[0149] Perform c on the input image 1 ×c 1 Convolution operation, in this embodiment, c 1 =3:

[0150]

[0151] X 1 Perform batch normalization and use the ReLU activation function for calculation:

[0152] X 2 =ReLU(BN(X 1 ));

[0153] Among them, BN is the batch normalization operation, and the formula is:

[0154]

[0155] Among them, μ and σ 2 are the mean and variance respectively, γ and β are learning parameters;

[0156] X 2 Carry out 2 ×c 2 Convolution operation, in this embodiment, c 2 =1:

[0157]

[0158] X 3 Perform batch normalization and use the ReLU activation function for calculation:

[0159] X 4 =ReLU(BN(X 3 ));

[0160] X 4 Carry out 4 ×c 4 Convolution operation, in this embodiment, c 4 =1:

[0161]

[0162] X 5 Perform batch normalization and use the ReLU activation function for calculation:

[0163] X 6 =ReLU(BN(X5 ));

[0164] X 2 Carry out 7 ×c 7 The maximum pooling operation is performed. In this embodiment, c 7 =1:

[0165]

[0166] X 6 And the output X of the maximum pooling layer 7 To connect:

[0167] X 8 =Concat(X 6 ,X 7 );

[0168] X 8 Carry out 9 ×c 9 Convolution operation, in this embodiment, c 9 =1:

[0169]

[0170] X 9 Perform batch normalization and use the ReLU activation function to calculate and obtain the feature map as the final output:

[0171] Output = ReLU(BN(X 9 )).

[0172] The above convolution kernels can be sized according to requirements, and convolution kernels of different sizes can be mixed to meet the requirements of feature extraction at different levels;

[0173] Through the context embedding module, the context information of the input image is captured to generate richer surface features, where the input image is the landslide image input data, that is, the feature map output by the backbone module. The structure of the context embedding module is shown in the attached figure. Figure 3 As shown, the specific calculation steps are as follows:

[0174] Perform GAPooling global average pooling calculation on the input feature map:

[0175]

[0176] Among them, f(i,j,c) is the value of the cth channel of the input feature map at position (i,j), f gap(c) is the value of the cth channel after pooling; the input feature map is the feature representation extracted by the backbone module from the original input image (i.e., landslide image) through convolution operation, which reflects the local features of the image and is used for subsequent deep learning tasks;

[0177] Perform batch normalization on the pooled data:

[0178] X bn =BN(f gap (c));

[0179] X bn Carry out conv1 ×c conv1 Convolution operation, in this embodiment, c conv1 =1:

[0180]

[0181] X conv1 Perform batch normalization and use the ReLU activation function for calculation:

[0182] X relu =ReLU(X conv1 );

[0183] X after broadcast relu Add to the input feature map f(i,j,c):

[0184] X sum =X relu +f(i,j,c);

[0185] X sum Carry out output ×c output Convolution operation, in this embodiment, c output =3:

[0186]

[0187] S4, using the bilateral aggregation guidance layer to fuse the detail branch and the semantic branch to form a landslide semantic segmentation model, and perform landslide semantic segmentation; the structure of the bilateral aggregation guidance layer is as shown in the attached figure. Figure 4 As shown;

[0188] The specific process is as follows:

[0189] S401, performing depth-separable convolution calculation on the feature map of the detail branch extracted by the detail branch landslide image feature extraction network, and then performing calculation by multi-scale parallel dilated convolution:

[0190]

[0191] Among them, Y is the calculation result of multi-scale parallel dilated convolution, K i is the i-th convolution kernel, X is the calculation result of the depthwise separable convolution, and N is the number of parallel convolution kernels;

[0192] Perform element-wise multiplication of Y with the upsampled result of the semantic branch:

[0193] X de_mul =Y×X upsam ;

[0194] Among them, X de_mul is the result after convolution of detail branch, X upsam It is the result after upsampling of the semantic branch; the upsampling methods include but are not limited to: bilinear interpolation, transposed convolution (deconvolution), and the upsampling of the semantic branch is for the intermediate feature map to restore the resolution of the feature map so as to fuse it with the feature maps of other branches to generate the final output result.

[0195] S402, perform a depth-separable convolution on the feature map of the semantic branch extracted by the landslide semantic branch feature extraction network, and then perform calculations through multi-scale parallel dilated convolutions, and then perform a Sigmoid function on the result X after the detail branch convolution. sem Process it and perform an element-wise product with the average pooling result of the detail branch:

[0196] X sem_mul =Sigmoid(X sem )×X apooling ;

[0197] Among them, X sem_mul is the result after convolution of detail branch, X apooling It is the result of average pooling of the semantic branch.

[0198] S403, adding the operation result of the detail branch in S401 and the operation result of the semantic branch in S402 pixel by pixel, and performing a convolution operation on the addition result to obtain the final output result.

[0199] S5. Use the landslide semantic segmentation dataset to train and test the landslide semantic segmentation model;

[0200] The specific process is as follows:

[0201] S501, setting parameters of a landslide semantic segmentation model, and using a training set of a landslide semantic segmentation data set to train the landslide semantic segmentation model; in this embodiment, the code is written in Python language, and the model parameters are set as shown in Table 2;

[0202] Table 2 Model training parameter settings

[0203]

[0204] S502: Perform semantic segmentation on the test set of the landslide semantic segmentation data set using the fitted landslide semantic segmentation model to obtain a segmentation result. Figure 5 and Figure 6 As shown in the figure, the segmentation results show that the bilateral segmentation network can effectively perform semantic segmentation on landslide remote sensing satellites. Figure 5 For the landslide label, Figure 6 is the landslide segmentation result;

[0205] S503, evaluating the training result of the landslide semantic segmentation model according to the evaluation index, and if the evaluation index meets the preset requirements, outputting the landslide semantic segmentation model for landslide semantic segmentation;

[0206] The evaluation indicators in this embodiment include: F1 score, accuracy, Kappa index and IOU; the training results of the model are evaluated, and the experimental results obtained are shown in Table 3;

[0207] Table 3 Model evaluation indicators

[0208]

[0209] Embodiment 2

[0210] This embodiment is basically the same as the above embodiment, except that it further includes:

[0211] S6. Use the landslide semantic segmentation model to identify the landslide contours, and calculate NDVI to analyze the vegetation coverage and growth status to assist in landslide identification;

[0212] The specific process is as follows:

[0213] Remote sensing image data of near infrared band and red light band including landslide contour spots are obtained; in this embodiment, remote sensing image data are obtained from a satellite remote sensing platform (such as Landsat, Sentinel-2, etc.) or an unmanned aerial vehicle remote sensing platform.

[0214] According to the remote sensing data source, select the corresponding near infrared band and red light band; the remote sensing data source includes remote sensing image data;

[0215] The NDVI value is calculated according to the formula NDVI = (NIR-R) / (NIR+R), where NIR is the reflectance of the near infrared band and R is the reflectance of the red light band; and the reflectance of the two bands is normalized during the calculation process to obtain a value between -1 and 1; in this embodiment, the python programming language can be used for programming;

[0216] The calculated NDVI values ​​are visualized and statistically analyzed, and the analysis data are extracted to assist in identifying landslides. When the NDVI value approaches zero, it means that the surface coverage within the range is rock or bare soil and other areas without vegetation. That is, the larger the NDVI value, the higher the vegetation coverage rate.

[0217] Embodiment 3

[0218] This embodiment is basically the same as the above embodiment, except that it further includes:

[0219] S7. Use digital high-rise model (DEM) to analyze the slope of landslide contour patches;

[0220] The specific process is as follows:

[0221] Obtain a digital elevation model (DEM) dataset; a digital elevation model (DEM) dataset is a raster dataset in which each pixel (or pixel) represents the elevation of a specific point on the surface; in this embodiment, a digital elevation model (DEM) dataset is obtained through an open source data platform;

[0222] Calculate the slope of the raster data and obtain a new raster data set, in which the value of each pixel represents the slope at the point; in this embodiment, the DEM data (i.e., raster data) in the digital elevation model data set is imported into the selected programming environment, and the slope of the raster data is calculated by programming in Python language. After the calculation is completed, a new raster data set is obtained, in which the value of each pixel represents the slope at the point;

[0223] The possibility of landslide is determined by the slope; specifically, the slope is one of the important factors affecting the occurrence of landslides. If the slope is greater than 10 degrees and less than 45 degrees, the possibility of landslide is judged to be high, because such slopes are more prone to landslides, especially slopes that are steep at the bottom, gentle in the middle, and steep at the top, and have a circular shape at the top, which are favorable terrain for landslides.

[0224] Vegetation has high reflectivity in the near-infrared band and relatively low reflectivity in the red band. Based on this characteristic, by obtaining remote sensing image data of specific bands containing landslide contours and applying the normalized difference vegetation index (NDVI) calculation formula, it is possible to quantitatively reflect the vegetation coverage and growth status. Different vegetation coverage conditions are closely related to the formation and stability of landslides. Generally speaking, the root system of vegetation has a reinforcing effect on the soil. The probability of landslides in areas with good vegetation coverage is relatively low, while the risk of landslides in areas with sparse or missing vegetation is higher. By accurately calculating the image data according to the formula, the NDVI value that characterizes the vegetation status can be obtained. For example, in forest-covered areas, the near-infrared band reflects a lot of light, and the red light band absorbs more, and the calculated NDVI value is close to 1; while in bare soil or rocky areas, the reflectivity of the two bands is similar, and the NDVI value is close to 0.

[0225] The introduction of NDVI calculation can accurately capture the vegetation information around the landslide area and integrate it into the landslide identification process as an important auxiliary judgment basis. Compared with relying solely on image semantic segmentation to judge landslides, combining vegetation status can eliminate the situation where areas with similar vegetation coverage but geological stability are misjudged as landslides. For example, some gentle slopes covered with low grass may be misjudged due to terrain light and shadow. Referring to NDVI, it is known that the roots of vegetation stabilize the soil, avoiding misidentification and improving the accuracy of locating landslide areas.

[0226] Taking vegetation characteristics as a reference, when the NDVI value shows high vegetation coverage, the system will not easily issue a landslide alarm even if some features of the image have a certain similarity with a landslide, such as shadows and textures, greatly reducing the probability of false alarms in areas with good vegetation and stable geology, and ensuring the reliability of early warning information.

[0227] Clearly distinguish between stable slopes covered with vegetation and potential landslide slopes. In complex mountain environments, there may be a mixture of various terrains and landforms. After calculating NDVI, the boundary between the landslide area and the surrounding normal vegetation area is clearer, and the two will not be confused, providing accurate basic data for subsequent disaster assessment and prevention.

[0228] Digital elevation model (DEM) records the elevation information of each point on the surface in the form of a grid. Based on this, the slope of the grid data is calculated through a specific algorithm. The terrain slope is one of the key factors affecting the occurrence of landslides. When the slope is within a certain range, the sliding force and anti-sliding force of the rock and soil under the action of gravity are unbalanced, and the possibility of landslides increases greatly. Generally speaking, the steeper the slope, the greater the impact of the gravity component on the rock and soil, and it is easier to start sliding; while the rock and soil on the too gentle slope is relatively stable.

[0229] For example, using the slope calculation method based on the triangulated network, the slope value is calculated according to the elevation difference and horizontal distance between adjacent grid points based on the trigonometric function relationship to quantify the slope corresponding to each pixel.

[0230] The possibility of landslides can be directly judged based on accurate slope data, providing key terrain basis for landslide identification models. Compared with judging only from visual image features, incorporating slope information can accurately focus on specific terrain areas prone to landslides, such as slopes with slopes of 10-45 degrees in mountainous areas, accurately identifying high-risk landslide areas, avoiding missing real hidden danger areas, and improving overall identification accuracy.

[0231] For gentle areas with a slope of less than 10 degrees, even if there are some light and shadow changes similar to landslides or slight signs of soil erosion in the image, the possibility of landslides can be ruled out in combination with the slope information, and no false alarms will be issued, reducing unnecessary waste of disaster prevention resources.

[0232] In mountainous areas, different terrains are intertwined, with small gullies formed by natural erosion, artificially built terraces and other complex terrains. Through slope analysis, areas that are similar to landslides due to undulating terrain but are actually stable can be distinguished from areas with real landslide potential, ensuring that the identification results are clear and accurate without causing confusion.

[0233] Embodiment 4

[0234] This embodiment is basically the same as the above embodiment, except that it further includes:

[0235] S8, monitoring rainfall conditions, obtaining rainfall parameters, and adjusting the landslide assignment parameters according to the rainfall conditions and rainfall parameters; the assignment parameters are set according to requirements, including: assignment or probability weight;

[0236] Rainfall is an important triggering factor for landslides, and its parameters such as rainfall amount, rainfall intensity and rainfall time have a significant impact on landslides;

[0237] Specific impacts include: Rainfall: heavy rainfall can oversaturate surface water, increasing the probability of soil loosening and rock settlement;

[0238] Rainfall intensity: High-intensity rainfall can be concentrated in small areas, increasing the likelihood of soil liquefaction and rock hydraulic shock;

[0239] Rainfall time: Long rainfall will cause the soil to be continuously saturated, increasing the risk of landslides, while short rainfall may reduce the probability of landslides;

[0240] Specifically, rainfall is an important triggering factor for landslides, and its multiple parameters - rainfall amount, rainfall intensity and rainfall time - play a pivotal role in landslide risk assessment. For areas where landslides are suspected to occur, it is particularly important to conduct a detailed analysis and verification of the rainfall conditions within a preset time period (such as 1-5 days). If a large amount of rainfall is monitored within the preset time period, the alertness to the landslide risk in the area should be immediately raised. Large amounts of rainfall can cause surface water to quickly reach or even exceed saturation, thereby significantly increasing the probability of soil loosening and rock settlement. In this case, the moisture content in the soil rises sharply, and the effective stress decreases, making the landslide risk in the area higher and lower. Soil is more prone to sliding; rainfall intensity is also a factor that cannot be ignored; high-intensity rainfall often forms strong runoff in a small area, which not only intensifies the soil liquefaction process, but also increases the hydraulic impact on the rock. Therefore, when high-intensity rainfall occurs in a short period of time, the identification weight is increased; rainfall time (i.e., rainfall duration) also has a significant impact on landslide risk. Long-term rainfall will keep the soil in a saturated state, weakening the friction between soil particles, thereby increasing the probability of landslides; on the contrary, if the rainfall time is short, the soil may drain quickly after the rainfall ends, reducing the saturation, and the weight should be reduced.

[0241] Based on the analysis of the above three key parameters, the suspected landslide area in the remote sensing image can be identified more accurately. The specific process is as follows:

[0242] Analyze whether there is rainfall within the preset time before the current time point. If not, the landslide risk in the area is judged to be low, and the assigned parameters of the landslide in the area are reduced according to the preset reduction ratio. That is, if there is no rainfall within the preset time (such as 1-5), it can be reasonably inferred that the landslide risk in the area is relatively low, so the assigned parameters of the suspected landslide can be appropriately reduced;

[0243] If so, determine whether the rainfall is greater than the preset rainfall, the rainfall intensity is greater than the preset rainfall, and the rainfall time is greater than the preset rainfall time. If any of the judgments is yes, the assigned parameters of the landslide in the area will be increased according to the preset increase ratio; if rainfall is monitored, especially when the rainfall is large, the intensity is high or the duration is long, the assigned parameters of the landslide in the area should be significantly increased, which will help to take preventive measures in time and reduce the losses caused by potential landslides.

[0244] The preset rising ratio is the adjustment ratio P0 (landslide adjustment coefficient);

[0245] P0=w1×P1+w2×P2+w3×P3;

[0246] Among them, P0 represents the adjusted probability coefficient of landslide occurrence, P1, P2 and P3 represent rainfall, rainfall intensity and rainfall time respectively, and w1, w2 and w3 represent the weights of rainfall, rainfall intensity and rainfall time respectively.

[0247] The triggering mechanism of rainfall on landslides is complex and critical. The amount of rainfall determines the amount of water infiltrating the surface. A large amount of rainfall causes the soil water content to rise rapidly, the weight to increase, the pore water pressure to increase, the effective stress to decrease, and the anti-sliding force to decrease. The intensity of rainfall affects the kinetic energy of raindrops impacting the surface. High-intensity rainfall can quickly destroy the soil structure, promote the formation of surface runoff, and aggravate erosion and soil loosening. The duration of rainfall is related to the continuous saturation of soil moisture. Long-term rainfall maintains high water content and continuously weakens soil stability. Based on these principles, by monitoring rainfall parameters within a preset time, the landslide value or probability weight is adjusted according to the weight distribution formula. For example, in a mountainous area, if the rainfall in a short period of time far exceeds the local soil infiltration capacity, the rainwater quickly gathers on the surface, and the rainfall intensity is high, the raindrops impact the slope surface, and the rainfall continues for 3 days. After calculation according to the formula, the landslide risk weight is greatly increased.

[0248] Dynamically consider rainfall factors and calibrate landslide risk judgment in real time according to actual rainfall conditions. Compared with fixed pattern recognition, it can accurately reflect changes in landslide risk before and after rainfall, such as the rising stage of soil moisture content after rain, accurately capture high-risk areas, and timely mark potential landslide bodies under the influence of heavy rainfall, without missing high-risk areas, thus improving the recognition accuracy of rainfall-related landslides.

[0249] During the dry and rainless period, the soil moisture is stable. Even if there is a small amount of surface change caused by weathering in the image, the low landslide risk can be determined based on rainfall monitoring and will not be misjudged as a landslide hazard, thus reducing false alarms and ensuring the credibility of the early warning system.

[0250] Distinguish between areas with risk changes caused by rainfall and naturally stable areas. In mountainous areas during the rainy season, some areas are greatly affected by rainfall, while others are less affected due to good drainage and stable geological conditions. Through precise weight adjustment, areas with high and low landslide risks are clearly defined, without confusing plots of different stability.

[0251] Embodiment 5

[0252] This embodiment is basically the same as the above embodiment, except that it further includes:

[0253] S9. Construct a comprehensive judgment model to identify and distinguish landslide areas and construction sites by combining the change rate of landslide edges, the change rate of pixels in suspected areas, and the trajectory of human activities;

[0254] A comprehensive judgment model is constructed to identify and distinguish landslide areas and construction sites by combining the change rate of landslide edges, the change rate of pixels in suspected areas, and the trajectory of human activities, thereby improving the accuracy of semantic segmentation of landslide remote sensing images.

[0255] The landslide edge change rate analysis:

[0256] By analyzing the remote sensing images of continuous time series, the position change rate of the pixel points at the edge of the landslide is calculated; where the position change rate = displacement / time;

[0257] Use image processing techniques, such as optical flow or image registration, to track the movement of pixels at the edge of the landslide in multiple images;

[0258] Analyze the position change rate of the pixel points at the edge of the landslide, identify the area where the landslide is active, and use it as the suspected landslide area. The active area of ​​the landslide is characterized by a position change rate of the pixel points at the edge of the landslide being higher than the preset position change rate, that is, it usually shows a higher pixel point position change rate.

[0259] Analysis of pixel change rate in suspected landslide area:

[0260] Marking suspected landslide areas in remote sensing images, where the suspected landslide areas may be landslides or construction sites; in this embodiment, the suspected landslide areas are identified and marked by color recognition and edge contour recognition;

[0261] Calculate the change rate of the pixels at the edge of the landslide in the suspected landslide area, including changes in brightness, color, texture and other characteristics;

[0262] According to the change rate of the pixel points at the edge of the landslide, it is judged whether the landslide area is a landslide area or a construction site; specifically, the landslide area is usually characterized by rapid and continuous changes in the pixel points, while the construction site may be characterized by intermittent and irregular changes; therefore, it is judged whether the change rate of the pixel points at the edge of the landslide is greater than the preset change rate and changes continuously (i.e., the brightness, color, and texture are greater than the corresponding preset values ​​and change continuously). If so, it is determined to be a landslide area; if the change rate of the pixel points at the edge of the landslide changes intermittently, it is determined to be a construction site;

[0263] Conduct human activity trajectory analysis, including:

[0264] Collect and analyze data on human activities near landslide areas, such as road construction, building construction, etc.;

[0265] Remote sensing image data and GIS technology are used to extract information on human activity trajectories, including activity type, scope, and intensity.

[0266] The landslide edge change rate, the suspected area pixel point change rate and the human activity trajectory are combined to build a comprehensive judgment model to distinguish landslide activities from construction sites; the comprehensive judgment model is used to analyze the landslide edge change rate and identify the landslide activity area as the suspected landslide area; the pixel point change rate analysis in the suspected landslide area is carried out to identify the landslide area and the construction site, and the identified landslide area and the construction site are compared and analyzed in combination with the human activity trajectory analysis to determine whether the identified landslide area is consistent with the construction site and the human activity trajectory, and the final distinction result between the landslide area and the construction site is output;

[0267] Training and validation of the comprehensive judgment model, including:

[0268] A remote sensing image dataset containing landslides and construction sites is used to train the comprehensive judgment model;

[0269] Verify the performance of the model through cross-validation and independent test sets to ensure its reliability and stability in practical applications;

[0270] Apply the trained comprehensive judgment model and obtain application data as feedback data to adjust and optimize the comprehensive judgment model, including:

[0271] Apply the trained comprehensive judgment model to actual landslide monitoring and identification tasks;

[0272] According to the actual application effect, the model parameters are continuously adjusted and optimized to improve its recognition accuracy and efficiency. Through the above adjustments and supplements, the landslide area and the construction site can be distinguished more accurately, providing more accurate information support for geological disaster warning and prevention. The model uses a classification algorithm based on machine learning, such as support vector machine (SVM) or deep neural network, to make the final decision and judgment, and takes the change rate of the landslide edge, the change rate of the pixels in the suspected area, and the quantitative characteristics of the human activity trajectory as input vectors to train the model to recognize two different modes of landslide activity and construction site. During the training process, a large amount of historical case data is used to optimize the model parameters so that it can accurately learn the combination mode of each feature in different scenarios. After sufficient training, the model can quickly and accurately output the judgment results when facing new monitoring data of suspected landslide areas.

[0273] The change rate of the landslide edge reflects the dynamic activity of the landslide body. Through the analysis of continuous time series remote sensing images, the displacement of pixels is tracked using image processing techniques such as optical flow method. When the landslide body is in an active state, the edge rock and soil body continues to deform and move, and the pixel position changes at a high rate. In terms of the change rate of pixels in the suspected area, the landslide area is in continuous motion, and the brightness, color, and texture change with the displacement, rolling, and exposure of fresh surfaces of the rock and soil body. Although there are local changes caused by personnel and mechanical operations at the construction site, the law is different from that of landslides, and most of them are intermittent and locally concentrated. At the same time, human activity trajectories are collected, because human engineering activities such as road construction and house construction may change the surface morphology, similar to the appearance of landslides, but combined with activity trajectory information, it can be judged whether it is a landslide from the type, scope, and intensity of the activity. These factors are combined to build a model, and each factor confirms and complements each other.

[0274] For example, the slope next to a road under construction in a mountainous area may be mistakenly judged as a landslide if only short-term image changes are observed. However, combined with the trajectory of human activities, frequent entry and exit of large machinery and material stacking may be found, which may be determined to be the impact of construction. Another example is that long-term monitoring shows that the edge pixels of a slope move steadily and the texture continues to change, but there is no sign of human construction in the corresponding area, which may be judged as a landslide activity area.

[0275] Multi-dimensional information fusion comprehensively considers landslide dynamics, surface change details and human interference factors to avoid the limitations of single-factor judgment. Whether it is the recovery of slowly developing old landslides, or newly formed landslide hazards, as well as scenes that are easily confused with construction sites, they can be accurately identified and landslide areas can be accurately identified, thereby improving the recognition success rate in complex environments.

[0276] For areas with human activities, such as short-term surface changes caused by temporary earth piles and excavations at construction sites, the comparison between the human activity trajectory and the pixel point change characteristics will prevent misjudgment as natural landslides, reduce unnecessary alarms, and maintain the efficiency and stability of the early warning system.

[0277] In complex scenarios such as urban-rural fringe areas and mountain development zones, it is necessary to clearly distinguish the surface differences caused by landslides, construction and normal geological evolution, accurately define the boundaries and nature of landslides, provide precise guidance for subsequent targeted prevention and control, and monitoring, and do not confuse surface changes of different causes and natures.

[0278] The above is only an embodiment of the present invention. The common sense such as the known specific structure and characteristics in the scheme is not described in detail here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field of the invention before the application date or priority date, can obtain all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A semantic segmentation method for landslide remote sensing images based on a bilateral segmentation network, characterized in that: It includes the following: S2, constructing a detail branch landslide image feature extraction network to extract a feature map of the detail branch; S3, establishing a landslide semantic branch feature extraction network to extract feature maps of semantic branches; S4. A bilateral aggregation guided layer is used to fuse the detail branch and the semantic branch to form a landslide semantic segmentation model for landslide semantic segmentation.

2. The landslide remote sensing image semantic segmentation method based on bilateral segmentation network according to claim 1 is characterized in that: Also includes: S1. Construct a landslide semantic segmentation dataset; Said S1 comprises: A landslide dataset is obtained for experiments, wherein the landslide dataset includes: images, shape files of landslide boundaries and digital elevation models; the images include landslide images and their corresponding labels; Preprocess the image; the preprocessing includes: adjusting the resolution and performing data enhancement; The landslide dataset was normalized and standardized, and the digital elevation model and image were fused.

3. The landslide remote sensing image semantic segmentation method based on bilateral segmentation network according to claim 1 is characterized in that: The detail branch landslide image feature extraction network comprises: a plurality of convolutional layers, each of which is connected in sequence, and the number of output channels of the convolutional layer is half of the number of output channels of the previous convolutional layer; The S2 comprises: S201, performing detail normalization processing on the input multi-channel landslide image: Among them, x ijkl is the index value of the image at (i, j, k, l), B is the batch size, C is the number of channels, H is the image length, W is the image width, μ j is the batch mean, σ j is the standard deviation, and ε is a value ranging from 10 -6 to 10 -12 A positive number between The detail normalized feature values ​​are: S202, the landslide image after detail normalization is convolved through sequentially connected convolutional layers, and the number of output channels of each convolutional layer is half of the number of input channels.

4. The method for semantic segmentation of landslide remote sensing images based on bilateral segmentation network according to claim 3 is characterized in that: The landslide semantic branch feature extraction network includes a backbone module and a context embedding module; The backbone module is used to process the input image and increase the number of channels while reducing the spatial dimension; the input image is a landslide image in the landslide semantic segmentation dataset; The context embedding module is used to capture the context information of the input image and generate surface features; The input image is the feature map output by the backbone module; The backbone module processes the input image and increases the number of channels while reducing the spatial dimension. The specific calculation steps are as follows: Perform a c1×c1 convolution operation on the input image: Perform batch normalization on X1 and use the ReLU activation function for calculation: X2 = ReLU(BN(X1)); Among them, BN is the batch normalization operation, and the formula is: Among them, μ and σ 2 are the mean and variance respectively, γ and β are learning parameters; Perform a c2×c2 convolution operation on X2: Perform batch normalization on X3 and use the ReLU activation function for calculation: X4 = ReLU(BN(X3)); Perform c4×c4 convolution operation on X4: Perform batch normalization on X5 and use the ReLU activation function for calculation: X6 = ReLU(BN(X5)); Perform a c7×c7 maximum pooling operation on X2: Connect X6 to the output of the maximum pooling layer X7: X8 = Concat(X6,X7); Perform c9×c9 convolution operation on X8: Perform batch normalization on X9 and use the ReLU activation function to calculate and obtain the final output: Output = ReLU(BN(X9)); Through the context embedding module, the context information of the input image is captured to generate surface features. The specific calculation steps are as follows: Perform GAPooling global average pooling calculation on the input feature map: Among them, f(i,j,c) is the value of the cth channel of the input feature map at position (i,j), f gap (c) is the value of the cth channel after pooling; Perform batch normalization on the pooled data: X bn =BN(f gap (c)); X bn Carry out conv1 ×c conv1 Convolution operation: X conv1 Perform batch normalization and use the ReLU activation function for calculation: X relu =ReLU(X conv1 ): X after broadcast relu Add to the input feature map f(i,j,c): X sum =X relu +f(i,j,c); X sum Carry out output ×c output Convolution operation:

5. The method for semantic segmentation of landslide remote sensing images based on bilateral segmentation network according to claim 4 is characterized in that: The S4 comprises: S401, performing depth-separable convolution calculation on the feature map of the detail branch extracted by the detail branch landslide image feature extraction network, and then performing calculation by multi-scale parallel dilated convolution: Among them, Y is the calculation result of multi-scale parallel dilated convolution, K i is the i-th convolution kernel, X is the calculation result of the depthwise separable convolution, and N is the number of parallel convolution kernels; Perform element-wise multiplication of Y with the upsampled result of the semantic branch: X de_mul =Y×X upsam ; Among them, X de_mul is the result after convolution of detail branch, X upsam This is the result after upsampling of the semantic branch; S402, perform a depth-separable convolution on the feature map of the semantic branch extracted by the landslide semantic branch feature extraction network, and then perform calculations through multi-scale parallel dilated convolutions, and then perform a Sigmoid function on the result X after the detail branch convolution. sem Process it and perform an element-wise product with the average pooling result of the detail branch: X sem_mul =Sigmoid(X sem )×X apooling ; Among them, X sem_mul is the result after convolution of detail branch, X apooling It is the result of average pooling of semantic branch; S403, adding the operation result of the detail branch in S401 and the operation result of the semantic branch in S402 pixel by pixel, and performing a convolution operation on the addition result to obtain the final output result.

6. The method for semantic segmentation of landslide remote sensing images based on bilateral segmentation network according to claim 5, characterized in that: Also includes: S5. Use the landslide semantic segmentation dataset to train and test the landslide semantic segmentation model, including: S501, setting parameters of a landslide semantic segmentation model, and using a training set of a landslide semantic segmentation data set to train the landslide semantic segmentation model; S502, performing semantic segmentation on a test set of the landslide semantic segmentation dataset using the fitted landslide semantic segmentation model to obtain a segmentation result; S503: Evaluate the training result of the landslide semantic segmentation model according to the evaluation index. If the evaluation index meets the preset requirements, output the landslide semantic segmentation model.

7. The method for semantic segmentation of landslide remote sensing images based on bilateral segmentation network according to claim 1, characterized in that: Also includes: S6. Use the landslide semantic segmentation model to identify the landslide contours, and calculate NDVI to analyze the vegetation coverage and growth status to assist in landslide identification, including: Acquire remote sensing image data in near-infrared and red bands including landslide contours; According to the remote sensing data source, select the corresponding near-infrared band and red light band; The NDVI value is calculated according to the formula NDVI = (NIR-R) / (NIR+R), where NIR is the reflectance of the near-infrared band and R is the reflectance of the red light band; and the reflectance of the two bands is normalized during the calculation process to obtain a value between -1 and 1; The calculated NDVI values ​​are visualized and statistically analyzed, and the analysis data are extracted to assist in identifying landslides; the larger the NDVI value, the higher the vegetation coverage.

8. The method for semantic segmentation of landslide remote sensing images based on bilateral segmentation network according to claim 7, characterized in that: Also includes: S7. Use the digital high-rise model to analyze the slope of the landslide contour map, including: Get a digital elevation model dataset; a digital elevation model dataset is a raster dataset in which each pixel represents the elevation of a specific point on the surface; Calculate the slope of the raster data and obtain a new raster dataset, in which the value of each pixel represents the slope at that point; The possibility of landslide is determined by the slope. If the slope is greater than 10 degrees and less than 45 degrees, the possibility of landslide is judged to be high.

9. The method for semantic segmentation of landslide remote sensing images based on bilateral segmentation network according to claim 8, characterized in that: Also includes: S8. Monitor rainfall conditions, obtain rainfall parameters, and adjust the landslide assignment parameters according to the rainfall conditions and rainfall parameters, including: Analyze whether there is rainfall within the preset time before the current time point. If not, the landslide risk in the area is judged to be low, and the assigned parameters of the landslide in the area are reduced according to the preset reduction ratio; If yes, then determine whether the rainfall is greater than the preset rainfall, the rainfall intensity is greater than the preset rainfall intensity, and the rainfall time is greater than the preset rainfall time. If any of the judgments is yes, then increase the assigned parameters of the landslide in the area according to the preset increase ratio; The preset rising ratio is the adjustment ratio P0; P0=w1×P1+w2×P2+w3×P3 Among them, P0 represents the adjusted probability coefficient of landslide occurrence, P1, P2 and P3 represent rainfall, rainfall intensity and rainfall time respectively, and w1, w2 and w3 represent the weights of rainfall, rainfall intensity and rainfall time respectively.

10. The landslide remote sensing image semantic segmentation method based on bilateral segmentation network according to claim 9, characterized in that: Also includes: S9. Construct a comprehensive judgment model to identify and distinguish landslide areas and construction sites by combining the change rate of landslide edges, the change rate of pixels in suspected areas, and the trajectory of human activities; A comprehensive judgment model is constructed to identify and distinguish landslide areas and construction sites by combining the change rate of landslide edges, the change rate of pixels in suspected areas, and the trajectory of human activities, including: Conduct landslide edge change rate analysis, including: By analyzing the remote sensing images of continuous time series, the position change rate of the pixel points at the edge of the landslide is calculated; Image processing technology is used to track the movement trajectory of pixels at the edge of the landslide in multiple images; Analyze the position change rate of the pixel points at the edge of the landslide, identify the area where the landslide is active, and use it as the suspected landslide area. The active area of ​​the landslide is characterized by the movement rate of the pixel points at the edge of the landslide being higher than the preset movement rate. Conduct pixel change rate analysis in the suspected landslide area, including: Mark suspected landslide areas in remote sensing images; Calculate the change rate of the pixels at the edge of the landslide in the suspected landslide area, including changes in brightness, color, and texture; According to the change rate of the pixel points at the edge of the landslide, judging whether the landslide area is a landslide area or a construction site, including: judging whether the change rate of the pixel points at the edge of the landslide is greater than a preset change rate and changes continuously, if so, judging it as a landslide area, if the change rate of the pixel points at the edge of the landslide changes intermittently, judging it as a construction site; Conduct human activity trajectory analysis, including: Collect and analyze data on human activities near the landslide area; Use remote sensing image data and GIS technology to extract information on human activity trajectories, including activity type, scope, and intensity; The change rate of the landslide edge, the change rate of the pixels in the suspected area, and the trajectory of human activities are combined to build a comprehensive judgment model to distinguish between landslide activities and construction sites; Training and validation of the comprehensive judgment model, including: A remote sensing image dataset containing landslides and construction sites is used to train the comprehensive judgment model; Validate the model’s performance through cross-validation and independent test sets; The trained comprehensive judgment model is applied, and the application data is obtained as feedback data to adjust and optimize the comprehensive judgment model.

Images