A landslide prediction method based on multi-dimensional features

Abstract

The application discloses a landslide prediction method based on multi-dimensional features and belongs to the technical field of image processing. The application acquires HH-HV and VV-VH incremental deviation values and constructs an incremental deviation image; then, a background area is segmented from the incremental deviation image, and the difference between each pixel point and the background area in the polarization dominant degree is calculated to obtain the landslide contrast of HH-HV and VV-VH, and a landslide contrast image is generated; the incremental deviation image and the corresponding landslide contrast image are fused and enhanced to obtain HH-HV and VV-VH feature enhancement images; suspected landslide points are found out from the feature enhancement images, and associated weights are acquired to form an associated weight matrix; a landslide prediction neural network is used to process the feature enhancement images, and the associated weight matrix is used to apply weights to the feature images, so that the landslide risk value of a monitoring area is obtained. The application effectively improves the accuracy of landslide prediction.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a landslide prediction method based on multidimensional features. Background Technology

[0002] Every year, landslides cause countless casualties and economic losses worldwide, especially in areas with complex geological structures such as mountains and hills, where landslides occur more frequently.

[0003] Common landslide prediction methods often rely on single monitoring data or characteristic indicators, such as surface displacement monitoring and rainfall statistics. However, these methods are only suitable for small areas; for large-scale landslide monitoring, image monitoring is the optimal choice.

[0004] Existing SAR-based landslide prediction methods have significant limitations in the application of polarization features. Relying solely on a single polarization feature (such as using HH or HV polarization information alone) or simply combining polarization features results in insignificant landslide characteristics and low accuracy in landslide prediction. Summary of the Invention

[0005] In view of the above-mentioned shortcomings in the prior art, the present invention provides a landslide prediction method based on multidimensional features, which solves the problem of low accuracy in landslide prediction in the prior art.

[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a landslide prediction method based on multidimensional features, comprising the following steps:

[0007] Based on the synthetic aperture radar imagery of the monitored area, the incremental deviation values ​​of HH-HV and VV-VH are obtained, and HH-HV and VV-VH incremental deviation images are constructed.

[0008] Background regions are segmented from HH-HV and VV-VH incremental deviation images;

[0009] The difference in polarization dominance between each pixel and the background region is used to obtain the HH-HV and VV-VH landslide contrast, resulting in HH-HV and VV-VH landslide contrast images.

[0010] The HH-HV incremental deviation image and the HH-HV landslide contrast image are fused and enhanced to obtain the HH-HV feature-enhanced image. The VV-VH incremental deviation image and the VV-VH landslide contrast image are fused and enhanced to obtain the VV-VH feature-enhanced image.

[0011] Suspected landslide points were identified from HH-HV and VV-VH feature-enhanced images, and correlation weights were obtained to form a correlation weight matrix.

[0012] A landslide prediction neural network was used to process HH-HV and VV-VH feature-enhanced images, and a correlation weight matrix was used to apply weights to the feature maps to obtain the landslide risk values ​​of the monitored area.

[0013] Furthermore, the process of constructing the HH-HV incremental deviation image includes:

[0014] Based on the HH backscattering coefficient and HV backscattering coefficient of pixels in the synthetic aperture radar images at adjacent time points, the HH backscattering increment and HV backscattering increment are obtained.

[0015] The difference in polarization increments between HH and HV is obtained by subtracting the backscattering increment of HV from the backscattering increment of the same pixel and taking the absolute value of the subtraction result.

[0016] The mean difference of HH-HV polarization increments is obtained by averaging the differences at multiple times in the monitoring area during the stable period.

[0017] The HH-HV increment deviation value is obtained by subtracting the mean HH-HV polarization increment difference from the HH-HV polarization increment difference at the same pixel, and an HH-HV increment deviation image is constructed.

[0018] Furthermore, the process of constructing the VV-VH incremental deviation image includes:

[0019] Based on the VV backscattering coefficient and VH backscattering coefficient of pixels in the synthetic aperture radar images at adjacent time points, the VV backscattering increment and VH backscattering increment are obtained.

[0020] The VV-VH polarization increment difference is obtained by subtracting the VH backscattering increment from the VV backscattering increment of the same pixel and taking the absolute value of the subtraction result.

[0021] The average difference in VV-VH polarization increments is obtained by averaging the differences in VV-VH polarization increments at multiple times in the monitoring area during the stable period.

[0022] The VV-VH polarization increment difference is obtained by subtracting the mean VV-VH polarization increment difference from the VV-VH polarization increment difference at the same pixel, and a VV-VH increment deviation image is constructed.

[0023] Furthermore, the process of obtaining HH-HV and VV-VH landslide contrast images includes:

[0024] For each pixel, calculate the first polarization dominance, where the first polarization dominance is the ratio of the HH backscattering coefficient to the HV backscattering coefficient;

[0025] For each pixel, the second polarization dominance is calculated, where the second polarization dominance is the ratio of the VV backscattering coefficient to the VH backscattering coefficient.

[0026] The mean value of each first polarization dominance in the background region is taken, and the mean value of each second polarization dominance is taken.

[0027] Based on the first polarization dominance and the mean of the first polarization dominance of each pixel, the landslide contrast of HH-HV is obtained, and the HH-HV landslide contrast image is obtained.

[0028] The landslide contrast of VV-VH is obtained by taking the second polarization dominance and the mean of the second polarization dominance of each pixel, thus obtaining the VV-VH landslide contrast image.

[0029] Furthermore, the process of obtaining landslide contrast includes:

[0030] When the first polarization dominance is less than or equal to the mean of the first polarization dominance, the sliding contrast of the HH-HV of the pixel is assigned a value of 1;

[0031] When the first polarization dominance is greater than the mean of the first polarization dominance, the HH-HV sag contrast of this pixel is: , where C i,HH-HV O is the slope contrast of HH-HV for the i-th pixel. 1,i O represents the first polarization dominance of the i-th pixel. 1,avg Let i be the mean of the first polarization dominance, where i is a positive integer;

[0032] When the second polarization dominance is less than or equal to the mean of the second polarization dominance, the VV-VH sliding contrast of the pixel is assigned a value of 1;

[0033] When the second polarization dominance is greater than the mean of the second polarization dominance, the VV-VH slippage contrast of this pixel is: , where C i,VV-VH For the VV-VH slope contrast of the i-th pixel, O 2,i O represents the second polarization dominance of the i-th pixel. 2,avg This represents the mean of the second polarization dominance.

[0034] Furthermore, the fusion enhancement process includes: multiplying the HH-HV incremental deviation image and the HH-HV landslide contrast image pixel by pixel, and normalizing the multiplication result to obtain the HH-HV feature enhancement image; multiplying the VV-VH incremental deviation image and the VV-VH landslide contrast image element by element, and normalizing the multiplication result to obtain the VV-VH feature enhancement image.

[0035] Further, the process of obtaining the association weight includes: taking the average of each enhancement value in the HH-HV feature enhancement image to obtain the HH-HV enhancement mean; taking the average of each enhancement value in the VV-VH feature enhancement image to obtain the VV-VH enhancement mean; classifying pixels in the HH-HV feature enhancement image that are greater than the HH-HV enhancement mean as suspected landslide points; calculating the proportion of suspected landslide points in the neighborhood of pixel point A in the HH-HV feature enhancement image at the same pixel point location A to obtain a first ratio; classifying pixels in the VV-VH feature enhancement image that are greater than the VV-VH enhancement mean as suspected landslide points; calculating the proportion of suspected landslide points in the neighborhood of pixel point A in the VV-VH feature enhancement image at the same pixel point location A to obtain a second ratio; and using the average of the first ratio and the second ratio as the association weight for pixel point location A.

[0036] Furthermore, the landslide prediction neural network comprises, in sequence: a multi-polarization feature fusion module, a multi-scale feature extraction module, a first convolutional layer, and a fully connected layer.

[0037] Furthermore, the multi-polarity feature fusion module includes: a first scale-invariant feature extraction unit, a second scale-invariant feature extraction unit, a multiplier M1, a multiplier M2, a first average pooling layer, a second average pooling layer, and an adder A1;

[0038] The input of the first scale-invariant feature extraction unit is used to input the HH-HV feature-enhanced image;

[0039] The input of the second scale-invariant feature extraction unit is used to input the VV-VH feature enhancement image;

[0040] The first input of multiplier M1 is connected to the output of the first scale-invariant feature extraction unit, its second input is used to input the correlation weight matrix, and its output is connected to the input of the first average pooling layer.

[0041] The first input of multiplier M2 is connected to the output of the second scale-invariant feature extraction unit, its second input is used to input the correlation weight matrix, and its output is connected to the input of the second average pooling layer.

[0042] The input of adder A1 is connected to the output of the first average pooling layer and the output of the second average pooling layer, respectively, and its output is used as the output of the multi-polarization feature fusion module.

[0043] Furthermore, the multi-scale feature extraction module includes: a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first residual attention module, a second residual attention module, a third residual attention module, and a Concat layer;

[0044] The input of the second convolutional layer is connected to the input of the third and fourth convolutional layers, respectively, and serves as the input of the multi-scale feature extraction module.

[0045] The input of the first residual attention module is connected to the output of the second convolutional layer; the input of the second residual attention module is connected to the output of the third convolutional layer; the input of the third residual attention module is connected to the output of the fourth convolutional layer; the input of the Concat layer is connected to the outputs of the first, second, and third residual attention modules respectively, and its output serves as the output of the multi-scale feature extraction module.

[0046] The beneficial effects of this invention are as follows:

[0047] 1. This invention calculates the incremental deviation values ​​of HH-HV and VV-VH and constructs corresponding images. Combined with landslide contrast analysis, it fuses and enhances the incremental deviation features with contrast features, effectively amplifying the difference between the suspected landslide area and the background area. Compared with traditional single polarization features or simple combination methods, it can more clearly highlight the characteristics of the landslide area and solve the problem of indistinct landslide features.

[0048] 2. This invention introduces an association weight matrix, assigning an association weight to each pixel. This allows the neural network to focus on highly correlated regions when processing feature-enhanced images, reducing interference from irrelevant backgrounds. Simultaneously, the collaborative analysis of multi-dimensional polarization features (HH-HV, VV-VH) covers richer information on surface scattering characteristics, significantly improving the accuracy of prediction results. Attached Figure Description

[0049] Figure 1 This is a flowchart of a landslide prediction method based on multidimensional features;

[0050] Figure 2 This is a schematic diagram of the structure of a landslide prediction neural network;

[0051] Figure 3 This is a schematic diagram of the structure of the multi-polarization feature fusion module;

[0052] Figure 4 This is a schematic diagram of the multi-scale feature extraction module.

[0053] Figure 5 The diagram shows the structure of the first scale-invariant feature extraction unit and the second scale-invariant feature extraction unit.

[0054] Figure 6 This is a schematic diagram of the structure of the first residual attention module, the second residual attention module, and the third residual attention module. Detailed Implementation

[0055] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0056] like Figure 1 As shown, a landslide prediction method based on multidimensional features includes the following steps:

[0057] Based on the synthetic aperture radar imagery of the monitored area, the incremental deviation values ​​of HH-HV and VV-VH are obtained, and HH-HV and VV-VH incremental deviation images are constructed.

[0058] Background regions are segmented from HH-HV and VV-VH incremental deviation images;

[0059] The difference in polarization dominance between each pixel and the background region is used to obtain the HH-HV and VV-VH landslide contrast, resulting in HH-HV and VV-VH landslide contrast images.

[0060] The HH-HV incremental deviation image and the HH-HV landslide contrast image are fused and enhanced to obtain the HH-HV feature-enhanced image. The VV-VH incremental deviation image and the VV-VH landslide contrast image are fused and enhanced to obtain the VV-VH feature-enhanced image.

[0061] Suspected landslide points were identified from HH-HV and VV-VH feature-enhanced images, and correlation weights were obtained to form a correlation weight matrix.

[0062] A landslide prediction neural network was used to process HH-HV and VV-VH feature-enhanced images, and a correlation weight matrix was used to apply weights to the feature maps to obtain the landslide risk values ​​of the monitored area.

[0063] In this embodiment, the process of constructing the HH-HV incremental deviation image includes: subtracting the HH backscattering increment from the HV backscattering increment of the same pixel to obtain the HH-HV polarization increment difference; and subtracting the average HH-HV polarization increment difference during the stable period from the HH-HV polarization increment difference to obtain the HH-HV incremental deviation value.

[0064] The specific process includes: obtaining the HH backscattering increment and HV backscattering increment based on the HH backscattering coefficient and HV backscattering coefficient of the pixels in the synthetic aperture radar images at adjacent time points;

[0065] The difference in polarization increments between HH and HV is obtained by subtracting the backscattering increment of HV from the backscattering increment of the same pixel and taking the absolute value of the subtraction result.

[0066] The average difference of HH-HV polarization increments at multiple times in the monitoring area during the stable period (for the same pixel) is calculated to obtain the average difference of HH-HV polarization increments.

[0067] The HH-HV increment deviation value is obtained by subtracting the mean HH-HV polarization increment difference from the HH-HV polarization increment difference at the same pixel. The HH-HV increment deviation values ​​are then arranged according to the pixel position to construct the HH-HV increment deviation image.

[0068] The HH backscattering increment is: I HH,t =S HH,t - S HH,t-1 , among which, I HH,t S is the HH backscattering increment at time t. HH,t Let S be the HH backscattering coefficient at time t. HH,t-1 Let HH be the backscattering coefficient at time t-1.

[0069] The HV backscattering increment is: I HV,t =S HV,t - S HV,t-1 , among which, I HV,t S is the HV backscattering increment at time t. HV,t Let S be the HV backscattering coefficient at time t. HV,t-1 Let be the HV backscattering coefficient at time t-1.

[0070] The stable period is the time during which no landslides occur.

[0071] After a landslide, the interaction between the HH polarized wave and the loose material is enhanced, and the HH backscattering coefficient increases significantly. The loose landslide body (gravel, soil) is mainly scattered by the surface, the depolarization effect is weakened, and the HV backscattering coefficient decreases significantly.

[0072] This invention uses the HV backscattering increment of the same pixel point minus the HH backscattering increment, and then uses the difference in HH-HV polarization increments minus the mean value. By highlighting the landslide characteristics from the direction of HH and HV increment changes, it can better highlight the unique dynamic characteristics of the landslide area compared with a single polarization increment.

[0073] In this embodiment, the process of constructing the VV-VH incremental deviation image includes: subtracting the VV backscattering increment from the VH backscattering increment of the same pixel to obtain the VH-VV polarization increment difference; and subtracting the average VH-VV polarization increment difference during the stable period from the VH-VV polarization increment difference to obtain the VH-VV polarization increment deviation value.

[0074] The specific process includes: obtaining the VV backscattering increment and VH backscattering increment based on the VV backscattering coefficient and VH backscattering coefficient of the pixels in the synthetic aperture radar images at adjacent time points;

[0075] The VV-VH polarization increment difference is obtained by subtracting the VH backscattering increment from the VV backscattering increment of the same pixel and taking the absolute value of the subtraction result.

[0076] The average difference of VV-VH polarization increments at multiple times in the monitoring area during the stable period (for the same pixel) is calculated to obtain the average difference of VV-VH polarization increments.

[0077] The VV-VH polarization increment difference is obtained by subtracting the mean VV-VH polarization increment difference from the VV-VH polarization increment difference at the same pixel. The VV-VH increment deviation values ​​are then arranged according to the pixel position to construct a VV-VH increment deviation image.

[0078] The VV backscattering increment is: I VV,t =S VV,t - S VV,t-1 , among which, I VV,t Let S be the VV backscattering increment at time t. VV,t Let V be the backscattering coefficient at time t, and S be the backscattering coefficient. VV,t-1 Let V be the backscattering coefficient at time t-1.

[0079] The VH backscattering increment is: I VH,t =S VH,t - S VH,t-1 , among which, I VH,t S is the VH backscattering increment at time t. VH,t Let S be the VH backscattering coefficient at time t. VH,t-1 Let VH be the backscattering coefficient at time t-1.

[0080] Landslides cause a sharp increase in surface roughness, which enhances surface scattering of VV polarization. Landslides destroy vegetation and surface structure, weakening volume scattering and making surface scattering dominant. The scattering contribution of VH cross-polarization is significantly reduced.

[0081] This invention uses the VH backscattering increment of the same pixel point minus the VV backscattering increment, and then uses the VV-VH polarization increment difference minus the mean value to highlight the landslide characteristics from the direction of VH and VV increment changes. Compared with a single polarization increment, it can better highlight the unique dynamic characteristics of the landslide area.

[0082] The process of segmenting the background region from both HH-HV and VV-VH incremental deviation images includes:

[0083] The mean of the incremental deviations is obtained by averaging the incremental deviation values ​​in the incremental deviation image.

[0084] Pixels whose increments deviate from the mean are marked as background pixels;

[0085] The largest connected region will be selected from the connected regions formed by each background point as the background region.

[0086] When the incremental deviation image is an HH-HV incremental deviation image, the resulting background region is named the HH-HV background region. When the incremental deviation image is a VV-VH incremental deviation image, the resulting background region is named the VV-VH background region.

[0087] In this embodiment, the process of obtaining HH-HV and VV-VH landslide contrast images includes:

[0088] For each pixel, calculate the first polarization dominance, where the first polarization dominance is the ratio of the HH backscattering coefficient to the HV backscattering coefficient;

[0089] For each pixel, the second polarization dominance is calculated, where the second polarization dominance is the ratio of the VV backscattering coefficient to the VH backscattering coefficient.

[0090] The mean value of each first polarization dominance in the background region is taken, and the mean value of each second polarization dominance is taken.

[0091] Based on the first polarization dominance and the mean of the first polarization dominance of each pixel, the landslide contrast of HH-HV is obtained, and the HH-HV landslide contrast image is obtained.

[0092] The landslide contrast of VV-VH is obtained by taking the second polarization dominance and the mean of the second polarization dominance of each pixel, thus obtaining the VV-VH landslide contrast image.

[0093] When constructing the HH-HV landslide contrast image, the mean of the dominance of each pixel belonging to the HH-HV background area is calculated. Then, the landslide contrast of each pixel in HH-HV is obtained, and the landslide contrast is arranged according to the corresponding pixel position to obtain the HH-HV landslide contrast image.

[0094] When constructing the VV-VH landslide contrast image, the mean of the dominance of each pixel belonging to the VV-VH background area is calculated. Then, the landslide contrast of each pixel in VV-VH is obtained, and the landslide contrast is arranged according to the corresponding pixel position to obtain the VV-VH landslide contrast image.

[0095] Cross-polarization (HV and VH) scattering is more sensitive to "volume scattering" (multipath scattering from branches, leaves, and roots) of vegetation, while co-polarization (HH and VV) scattering is more sensitive to "surface scattering" from rough surfaces (such as bare soil and rocks).

[0096] Different polarization modes (HH, HV, VV, VH) exhibit varying responses to surface scattering mechanisms. HH and VV co-polarizations are more sensitive to scattering from rough surfaces, while HV and VH cross-polarizations are more sensitive to volume scattering from vegetation, etc. This relative relationship of scattering characteristics can be quantified by calculating polarization dominance (the ratio of co-polarization to cross-polarization). After a landslide, the surface transitions from stable "vegetation-soil" composite scattering to rough surface scattering dominated by "bare soil / gravel." The mean value of each dominance in the background area is calculated. Since no landslide occurred in the background area, the landslide contrast is obtained based on the difference between the polarization dominance of each pixel and the mean polarization dominance of the background area, highlighting the landslide situation.

[0097] In this embodiment, the process of obtaining landslide contrast includes:

[0098] When the first polarization dominance is less than or equal to the mean of the first polarization dominance, the sliding contrast of the HH-HV of the pixel is assigned a value of 1;

[0099] When the first polarization dominance is greater than the mean of the first polarization dominance, the HH-HV sag contrast of this pixel is: , where C i,HH-HV O is the slope contrast of HH-HV for the i-th pixel. 1,i O represents the first polarization dominance of the i-th pixel. 1,avg Let i be the mean of the first polarization dominance, where i is a positive integer;

[0100] When the second polarization dominance is less than or equal to the mean of the second polarization dominance, the VV-VH sliding contrast of the pixel is assigned a value of 1;

[0101] When the second polarization dominance is greater than the mean of the second polarization dominance, the VV-VH slippage contrast of this pixel is: , where C i,VV-VH For the VV-VH slope contrast of the i-th pixel, O 2,i O represents the second polarization dominance of the i-th pixel. 2,avg This represents the mean of the second polarization dominance.

[0102] When the polarization dominance is less than or equal to the mean, it is directly assigned a value of 1, which can quickly mark the region where the polarization characteristics conform to the stable background; when the polarization dominance is greater than the mean, the degree of deviation is quantified by the formula, which can accurately capture the region where the polarization dominance deviates abnormally due to landslide.

[0103] In this embodiment, the fusion enhancement process includes: multiplying the HH-HV incremental deviation image and the HH-HV landslide contrast image pixel by pixel, and normalizing the multiplication result to obtain the HH-HV feature enhancement image; multiplying the VV-VH incremental deviation image and the VV-VH landslide contrast image element by element, and normalizing the multiplication result to obtain the VV-VH feature enhancement image.

[0104] In this embodiment, the process of obtaining the association weight includes: taking the average of each enhancement value (which is the value after normalization of the multiplication result) in the HH-HV feature enhancement image to obtain the HH-HV enhancement mean; taking the average of each enhancement value in the VV-VH feature enhancement image to obtain the VV-VH enhancement mean; classifying pixels in the HH-HV feature enhancement image that are greater than the HH-HV enhancement mean as suspected landslide points; calculating the proportion of the number of suspected landslide points in the neighborhood of pixel point A in the HH-HV feature enhancement image at the same pixel point location A to obtain a first ratio; classifying pixels in the VV-VH feature enhancement image that are greater than the VV-VH enhancement mean as suspected landslide points; calculating the proportion of the number of suspected landslide points in the neighborhood of pixel point A in the VV-VH feature enhancement image at the same pixel point location A to obtain a second ratio; and using the average of the first ratio and the second ratio as the association weight of pixel point location A.

[0105] The first ratio is L HH-HV / 9,L HH-HV 9 represents the number of suspected landslide points within the neighborhood of pixel location A in the HH-HV feature-enhanced image, 9 represents the size of the neighborhood, and the second ratio is L. VV-VH / 9,L VV-VH This represents the number of suspected landslide points within the neighborhood of pixel location A in the VV-VH feature-enhanced image.

[0106] Arrange the associated weights according to their pixel positions to obtain the associated weight matrix.

[0107] This invention measures the correlation between a pixel and landslide features from a local perspective by calculating the proportion of suspected landslide points within the pixel's neighborhood. The invention calculates the ratio based on feature-enhanced images using both HH-HV and VV-VH polarization combinations, then averages the ratios to obtain the correlation weight. This fully integrates landslide feature information under different polarization modes. The correlation weight reflects the probability that a pixel is a landslide point; a higher correlation weight indicates a greater likelihood of a landslide at that pixel.

[0108] like Figure 2 As shown, the landslide prediction neural network includes, in sequence: a multi-polarization feature fusion module, a multi-scale feature extraction module, a first convolutional layer, and a fully connected layer.

[0109] The multi-polarization feature fusion module is used to input HH-HV and VV-VH feature-enhanced images. Then, the correlation weight matrix is ​​used to apply correlation weights to the corresponding features of the HH-HV feature-enhanced image and the corresponding features of the VV-VH feature-enhanced image. The HH-HV features and VV-VH features with the correlation weights are pooled and added together to obtain the multi-polarization fused features.

[0110] The multi-scale feature extraction module is used to extract multi-scale features from the multi-polar fusion features. After convolution processing by the first convolutional layer, the features are further extracted. The fully connected layer obtains the landslide risk value of the monitoring area based on the output of the first convolutional layer.

[0111] like Figure 3 As shown, the multi-polarization feature fusion module includes: a first scale-invariant feature extraction unit, a second scale-invariant feature extraction unit, a multiplier M1, a multiplier M2, a first average pooling layer, a second average pooling layer, and an adder A1;

[0112] The input terminal of the first scale-invariant feature extraction unit is used to input the HH-HV feature-enhanced image; the input terminal of the second scale-invariant feature extraction unit is used to input the VV-VH feature-enhanced image.

[0113] The first input of multiplier M1 is connected to the output of the first scale-invariant feature extraction unit, its second input is used to input the correlation weight matrix, and its output is connected to the input of the first average pooling layer.

[0114] The first input of multiplier M2 is connected to the output of the second scale-invariant feature extraction unit, its second input is used to input the correlation weight matrix, and its output is connected to the input of the second average pooling layer.

[0115] The input of adder A1 is connected to the output of the first average pooling layer and the output of the second average pooling layer, respectively, and its output is used as the output of the multi-polarization feature fusion module.

[0116] The first and second scale-invariant feature extraction units maintain consistent input and output sizes when extracting features from HH-HV and VV-VH enhanced images. This maximizes the preservation of spatial information in the original images and avoids feature misalignment or loss due to size changes. Multipliers M1 and M2 then apply correlation weights to corresponding locations. These correlation weights reflect the degree of association between pixels and the landslide, giving higher weights to features more relevant to the landslide, highlighting key features, suppressing background interference, and ensuring that subsequent feature fusion focuses more on the landslide area. Finally, the first and second average pooling layers extract mean features, and the HH-HV and VV-VH features are fused at adder A1.

[0117] like Figure 4 As shown, the multi-scale feature extraction module includes: a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first residual attention module, a second residual attention module, a third residual attention module, and a Concat layer;

[0118] The input of the second convolutional layer is connected to the input of the third and fourth convolutional layers, respectively, and serves as the input of the multi-scale feature extraction module.

[0119] The input of the first residual attention module is connected to the output of the second convolutional layer; the input of the second residual attention module is connected to the output of the third convolutional layer; the input of the third residual attention module is connected to the output of the fourth convolutional layer; the input of the Concat layer is connected to the outputs of the first, second, and third residual attention modules respectively, and its output serves as the output of the multi-scale feature extraction module.

[0120] This invention employs a multi-branch structure with second, third, and fourth convolutional layers to extract features at different scales from the input. Different convolutional layers can capture information of varying scales and levels of abstraction from the input, providing diverse feature inputs for the subsequent residual attention module and enriching the overall feature representation capability.

[0121] The second convolutional layer has a kernel size of 3×3, a stride of 1, and no padding; the third convolutional layer has a kernel size of 5×5, a stride of 1, and no padding; the fourth convolutional layer has a kernel size of 7×7, a stride of 1, and no padding.

[0122] like Figure 5 As shown, both the first and second scale-invariant feature extraction units include a fifth convolutional layer, a sixth convolutional layer, and a seventh convolutional layer. The kernel size of the fifth, sixth, and seventh convolutional layers is 3×3, the stride is 1, and the padding is 1.

[0123] The input and output sizes of the first-scale-invariant feature extraction unit and the second-scale-invariant feature extraction unit are the same.

[0124] like Figure 6 As shown, the first residual attention module, the second residual attention module, and the third residual attention module have the same structure, each including: a first convolutional block, a second convolutional block, a Sigmoid layer, a multiplier M3, and an adder A2;

[0125] Both the first and second convolutional blocks consist of: a convolutional layer, a ReLU activation function layer, and a global pooling layer.

[0126] The kernel size of the convolutional layer in the first convolutional block is 1×1, and the kernel size of the convolutional layer in the second convolutional block is 3×3. The kernel size of the first convolutional layer is 3×3.

[0127] Comparison table of landslide risk values ​​and conditions in the monitored area:

[0128]

[0129] This invention calculates the incremental deviation values ​​of HH-HV and VV-VH and constructs corresponding images. Combined with landslide contrast analysis, it fuses and enhances the incremental deviation features with contrast features, effectively amplifying the difference between suspected landslide areas and background areas. Compared to traditional single polarization features or simple combinations, this method more clearly highlights the characteristics of landslide areas, solving the problem of indistinct landslide features.

[0130] This invention introduces a correlation weight matrix, assigning a correlation weight to each pixel. This allows the neural network to focus on highly correlated regions when processing feature-enhanced images, reducing interference from irrelevant backgrounds. Simultaneously, the collaborative analysis of multi-dimensional polarization features (HH-HV, VV-VH) covers richer information on surface scattering characteristics, significantly improving the accuracy of prediction results.

[0131] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A landslide prediction method based on multidimensional features, characterized in that, Includes the following steps: Based on the synthetic aperture radar imagery of the monitored area, the incremental deviation values ​​of HH-HV and VV-VH are obtained, and HH-HV and VV-VH incremental deviation images are constructed. The HH-HV polarization increment difference is obtained by subtracting the HH backscattering increment from the HV backscattering increment of the same pixel. The HH-HV polarization increment difference is then subtracted from the mean HH-HV polarization increment difference during the stable period to obtain the HH-HV increment deviation value. The VH-VV polarization increment difference is obtained by subtracting the VV backscattering increment from the VH backscattering increment of the same pixel. The VH-VV polarization increment deviation is obtained by subtracting the average VH-VV polarization increment difference during the stable period from the VH-VV polarization increment difference. Background regions are segmented from HH-HV and VV-VH incremental deviation images; The difference in polarization dominance between each pixel and the background region is used to obtain the HH-HV and VV-VH landslide contrast, resulting in HH-HV and VV-VH landslide contrast images. The process of obtaining landslide contrast includes: When the first polarization dominance is less than or equal to the mean of the first polarization dominance, the sliding contrast of HH-HV for that pixel is assigned a value of 1. The first polarization dominance is the ratio of the backscattering coefficient of HH to the backscattering coefficient of HV. When the first polarization dominance is greater than the mean of the first polarization dominance, the HH-HV sag contrast of this pixel is: , where C i,HH-HV O is the slope contrast of HH-HV for the i-th pixel. 1,i O represents the first polarization dominance of the i-th pixel. 1,avg Let i be the mean of the first polarization dominance, where i is a positive integer; When the second polarization dominance is less than or equal to the mean of the second polarization dominance, the sliding contrast of VV-VH for that pixel is assigned a value of 1. The second polarization dominance is the ratio of the backscattering coefficient of VV to the backscattering coefficient of VH. When the second polarization dominance is greater than the mean of the second polarization dominance, the VV-VH slippage contrast of this pixel is: , where C i,VV-VH For the VV-VH slope contrast of the i-th pixel, O 2,i O represents the second polarization dominance of the i-th pixel. 2,avg The mean of the second polarization dominance; The HH-HV incremental deviation image and the HH-HV landslide contrast image are fused and enhanced to obtain the HH-HV feature-enhanced image. The VV-VH incremental deviation image and the VV-VH landslide contrast image are fused and enhanced to obtain the VV-VH feature-enhanced image. Suspected landslide points were identified from HH-HV and VV-VH feature-enhanced images, and correlation weights were obtained to form a correlation weight matrix. The process of obtaining the association weight includes: taking the average of each enhancement value in the HH-HV feature enhancement image to obtain the HH-HV enhancement mean; taking the average of each enhancement value in the VV-VH feature enhancement image to obtain the VV-VH enhancement mean; classifying pixels in the HH-HV feature enhancement image that are greater than the HH-HV enhancement mean as suspected landslide points; calculating the proportion of suspected landslide points in the neighborhood of pixel point A in the HH-HV feature enhancement image at the same pixel point location A to obtain a first ratio; classifying pixels in the VV-VH feature enhancement image that are greater than the VV-VH enhancement mean as suspected landslide points; calculating the proportion of suspected landslide points in the neighborhood of pixel point A in the VV-VH feature enhancement image at the same pixel point location A to obtain a second ratio; and using the average of the first ratio and the second ratio as the association weight for pixel point location A. A landslide prediction neural network was used to process HH-HV and VV-VH feature-enhanced images, and a correlation weight matrix was used to apply weights to the feature maps to obtain the landslide risk values ​​of the monitored area.

2. The landslide prediction method based on multidimensional features according to claim 1, characterized in that, The process of constructing the HH-HV incremental deviation image includes: Based on the HH backscattering coefficient and HV backscattering coefficient of pixels in the synthetic aperture radar images at adjacent time points, the HH backscattering increment and HV backscattering increment are obtained. The difference in polarization increments between HH and HV is obtained by subtracting the backscattering increment of HV from the backscattering increment of the same pixel and taking the absolute value of the subtraction result. The mean difference of HH-HV polarization increments is obtained by averaging the differences at multiple times in the monitoring area during the stable period. The HH-HV increment deviation value is obtained by subtracting the mean HH-HV polarization increment difference from the HH-HV polarization increment difference at the same pixel, and an HH-HV increment deviation image is constructed.

3. The landslide prediction method based on multidimensional features according to claim 1, characterized in that, The process of constructing the VV-VH incremental deviation image includes: Based on the VV backscattering coefficient and VH backscattering coefficient of pixels in the synthetic aperture radar images at adjacent time points, the VV backscattering increment and VH backscattering increment are obtained. The VV-VH polarization increment difference is obtained by subtracting the VH backscattering increment from the VV backscattering increment of the same pixel and taking the absolute value of the subtraction result. The average difference in VV-VH polarization increments is obtained by averaging the differences in VV-VH polarization increments at multiple times in the monitoring area during the stable period. The VV-VH polarization increment difference is obtained by subtracting the mean VV-VH polarization increment difference from the VV-VH polarization increment difference at the same pixel, and a VV-VH increment deviation image is constructed.

4. The landslide prediction method based on multidimensional features according to claim 1, characterized in that, The process of obtaining HH-HV and VV-VH landslide contrast images includes: Calculate the first polarization dominance for each pixel; Calculate the second polarization dominance for each pixel; The mean value of each first polarization dominance in the background region is taken, and the mean value of each second polarization dominance is taken. Based on the first polarization dominance and the mean of the first polarization dominance of each pixel, the landslide contrast of HH-HV is obtained, and the HH-HV landslide contrast image is obtained. The landslide contrast of VV-VH is obtained by taking the second polarization dominance and the mean of the second polarization dominance of each pixel, thus obtaining the VV-VH landslide contrast image.

5. The landslide prediction method based on multidimensional features according to claim 1, characterized in that, The fusion enhancement process includes: multiplying the HH-HV incremental deviation image and the HH-HV landslide contrast image pixel by pixel, and normalizing the multiplication result to obtain the HH-HV feature enhancement image; multiplying the VV-VH incremental deviation image and the VV-VH landslide contrast image element by element, and normalizing the multiplication result to obtain the VV-VH feature enhancement image.

6. The landslide prediction method based on multidimensional features according to claim 1, characterized in that, The landslide prediction neural network consists of the following components connected in sequence: a multi-polarization feature fusion module, a multi-scale feature extraction module, a first convolutional layer, and a fully connected layer.

7. The landslide prediction method based on multidimensional features according to claim 6, characterized in that, The multi-polarization feature fusion module includes: a first scale-invariant feature extraction unit, a second scale-invariant feature extraction unit, a multiplier M1, a multiplier M2, a first average pooling layer, a second average pooling layer, and an adder A1; The input of the first scale-invariant feature extraction unit is used to input the HH-HV feature-enhanced image; The input of the second scale-invariant feature extraction unit is used to input the VV-VH feature enhancement image; The first input of multiplier M1 is connected to the output of the first scale-invariant feature extraction unit, its second input is used to input the correlation weight matrix, and its output is connected to the input of the first average pooling layer. The first input of multiplier M2 is connected to the output of the second scale-invariant feature extraction unit, its second input is used to input the correlation weight matrix, and its output is connected to the input of the second average pooling layer. The input of adder A1 is connected to the output of the first average pooling layer and the output of the second average pooling layer, respectively, and its output is used as the output of the multi-polarization feature fusion module.

8. The landslide prediction method based on multidimensional features according to claim 6, characterized in that, The multi-scale feature extraction module includes: a second convolutional layer, a third convolutional layer, a fourth convolutional layer, a first residual attention module, a second residual attention module, a third residual attention module, and a Concat layer; The input of the second convolutional layer is connected to the input of the third and fourth convolutional layers, respectively, and serves as the input of the multi-scale feature extraction module. The input of the first residual attention module is connected to the output of the second convolutional layer; the input of the second residual attention module is connected to the output of the third convolutional layer; the input of the third residual attention module is connected to the output of the fourth convolutional layer; the input of the Concat layer is connected to the outputs of the first, second, and third residual attention modules respectively, and its output serves as the output of the multi-scale feature extraction module.

Images