A landslide disaster early warning method

Abstract

The present invention discloses a landslide disaster early warning method, which belongs to the field of image processing technology. The present invention collects synthetic aperture radar images of a mountain monitoring area under HV and HH polarization modes to obtain HV images and HH images; finds the HV and HH images corresponding to the starting time of the landslide event, and based on the difference in radar backscattering amplitude and phase between the starting time and the subsequent HV images, extracts dual-polarization amplitude change values ​​and phase change values ​​respectively, and constructs a dual-polarization amplitude and phase change image; counts the number of times a pixel point is marked as a landslide point, and constructs an amplitude and phase position confidence distribution matrix; finally, uses a landslide disaster early warning neural network to process the dual-polarization amplitude and phase change image, and combines the confidence applied by the amplitude and phase position confidence distribution matrix to obtain a landslide disaster risk value, effectively solving the problem of low accuracy of landslide disaster early warning in the existing technology.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a landslide disaster early warning method. Background Art

[0002] Landslides, a common geological disaster, are characterized by suddenness and destructive power, posing a serious threat to human life and property, infrastructure, and the ecological environment. In mountainous and hilly areas, landslides can not only bury homes, block traffic, and damage farmland, but can also trigger secondary disasters, resulting in significant losses to regional economic and social development. Therefore, timely and accurate landslide warnings have become a critical issue in geological disaster prevention and control.

[0003] With the development of remote sensing technology, synthetic aperture radar (SAR) has shown great potential in landslide monitoring due to its all-day, all-weather, and high-resolution observation capabilities. However, single-polarization SAR images contain limited information and cannot fully capture the complex changes in surface characteristics during landslides, resulting in low early warning accuracy. Summary of the Invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a landslide disaster early warning method that solves the problem of low accuracy of landslide disaster early warning in the prior art.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a landslide disaster early warning method, comprising:

[0006] In the HV and HH polarization modes, synthetic aperture radar images of the mountain monitoring area are collected respectively to obtain HV images and HH images;

[0007] Find the HV image and HH image at the start of the landslide event in each HV image and HH image;

[0008] According to the difference in radar backscatter amplitude between the image at the start time and the image after the start time, the dual-polarization amplitude change value is extracted and the dual-polarization amplitude change image is constructed;

[0009] According to the phase difference between the image at the start time and the image after the start time, the dual-polarization phase change value is extracted to construct a dual-polarization phase change image;

[0010] Count the number of times the same pixel is marked as the first landslide point and the second landslide point at multiple consecutive moments after the landslide event occurs, and construct the amplitude confidence distribution matrix and the phase position confidence distribution matrix;

[0011] The landslide disaster early warning neural network is used to process the dual-polarization amplitude change image and the dual-polarization phase change image. The landslide disaster risk value is obtained based on the confidence imposed by the amplitude confidence distribution matrix and the phase position confidence distribution matrix.

[0012] Furthermore, the process of constructing the dual-polarization amplitude change image includes:

[0013] The HV image at the start of the landslide event is subtracted from the HV image at time t to obtain the first radar backscatter amplitude difference of each pixel at time t, where t is the time number after the landslide event occurs;

[0014] Subtract the HH image at time t from the HH image at the start of the landslide event to obtain the second radar backscatter amplitude difference of each pixel at time t;

[0015] The first radar backscattering amplitude difference and the second radar backscattering amplitude difference at the same pixel position are added and normalized to obtain the dual-polarization amplitude change value at time t;

[0016] The dual-polarization amplitude change value of each pixel at time t is arranged according to the position of the pixel point to obtain the dual-polarization amplitude change image at time t.

[0017] Furthermore, the process of constructing the dual-polarization phase change image includes:

[0018] The HV image at time t is subtracted from the HV image at the start of the landslide event in phase to obtain the first phase difference of each pixel at time t, where t is the number of the time after the landslide event occurs;

[0019] Subtract the phase of the HH image at time t from the HH image at the start of the landslide event to obtain the second phase difference of each pixel at time t;

[0020] The first phase difference and the second phase difference at the same pixel position are added and normalized to obtain the dual-polarization phase change value at time t;

[0021] The dual-polarization phase change value of each pixel at time t is arranged according to the position of the pixel point to obtain the dual-polarization phase change image at time t.

[0022] Furthermore, the process of constructing the amplitude confidence distribution matrix and the phase position confidence distribution matrix includes:

[0023] In each dual-polarization amplitude change image, the pixel whose dual-polarization amplitude change value is greater than the amplitude change threshold is marked as the first landslide point;

[0024] In the dual-polarization amplitude change images corresponding to multiple consecutive moments after the landslide event occurs, the number of times the same pixel position is marked as the first landslide point is recorded;

[0025] According to the number of times the same pixel position is marked as the first landslide point, the amplitude confidence is calculated, and the amplitude confidence of each pixel is arranged according to the position of the pixel to obtain the amplitude confidence distribution matrix;

[0026] In each dual-polarization phase change image, the pixel whose dual-polarization phase change value is greater than the phase change threshold is marked as the second landslide point;

[0027] In the dual-polarization phase change images corresponding to multiple consecutive moments after the landslide event occurs, the number of times the same pixel position is marked as the second landslide point is recorded;

[0028] The phase position reliability is calculated based on the number of times the same pixel position is marked as the second landslide point. The phase position reliability of each pixel is arranged according to the position of the pixel to obtain the phase position reliability distribution matrix.

[0029] Furthermore, the process of calculating the amplitude confidence includes: taking the ratio of the number of times the same pixel position is marked as the first landslide point to the total time, when the ratio is greater than or equal to 0.5, the amplitude confidence is equal to the ratio, and when the ratio is less than 0.5, the amplitude confidence is 0;

[0030] The process of calculating the phase position reliability includes: taking the ratio of the number of times the same pixel position is marked as the second landslide point to the total time. When the ratio is greater than or equal to 0.5, the phase position reliability is equal to the ratio; when the ratio is less than 0.5, the phase position reliability is 0.

[0031] Furthermore, the landslide disaster warning neural network includes: an amplitude confidence applying unit, a phase position confidence applying unit, an amplitude feature extraction unit, a phase feature extraction unit, an adder A1, a first scale attention unit, a second scale attention unit, a Concat layer, a first convolution layer and a fully connected layer;

[0032] The first input end of the amplitude confidence applying unit is used to input the dual-polarization amplitude change image, the second input end thereof is used to input the amplitude confidence distribution matrix, and the output end thereof is connected to the input end of the amplitude feature extraction unit; the first input end of the phase position confidence applying unit is used to input the dual-polarization phase change image, the second input end thereof is used to input the phase position confidence distribution matrix, and the output end thereof is connected to the input end of the phase feature extraction unit;

[0033] The input end of adder A1 is connected to the output end of the amplitude feature extraction unit and the output end of the phase feature extraction unit respectively, and its output end is connected to the input end of the first scale attention unit and the input end of the second scale attention unit respectively; the input end of the Concat layer is connected to the output end of the first scale attention unit and the output end of the second scale attention unit respectively, and its output end is connected to the input end of the first convolutional layer;

[0034] The input end of the fully connected layer is connected to the output end of the first convolutional layer, and its output end serves as the output end of the landslide disaster warning neural network.

[0035] Furthermore, the amplitude confidence applying unit includes: a second convolutional layer and a multiplier M1;

[0036] The input end of the second convolutional layer serves as the first input end of the amplitude confidence applying unit; the first input end of the multiplier M1 is connected to the output end of the second convolutional layer, the second input end of the multiplier M1 serves as the second input end of the amplitude confidence applying unit, and the output end of the multiplier M1 serves as the output end of the amplitude confidence applying unit;

[0037] The phase position confidence applying unit includes: a third convolutional layer and a multiplier M2;

[0038] The input end of the third convolutional layer serves as the first input end of the phase position confidence application unit;

[0039] The first input terminal of the multiplier M2 is connected to the output terminal of the third convolutional layer, the second input terminal thereof serves as the second input terminal of the phase position reliability applying unit, and the output terminal thereof serves as the output terminal of the phase position reliability applying unit.

[0040] Furthermore, the amplitude feature extraction unit and the phase feature extraction unit each include: a first convolution block, a second convolution block and a Maxpool layer;

[0041] The input end of the first convolution block serves as the input end of the amplitude feature extraction unit or the phase feature extraction unit, and the output end thereof is connected to the input end of the second convolution block;

[0042] The input end of the Maxpool layer is connected to the output end of the second convolution block, and its output end serves as the output end of the amplitude feature extraction unit or the phase feature extraction unit.

[0043] Furthermore, the first-scale attention unit and the second-scale attention unit each include: a third convolution block, a fourth convolution block, a Sigmoid activation layer, and a multiplier M3;

[0044] The input end of the third convolution block serves as the input end of the first scale attention unit or the second scale attention unit, and its output end is connected to the first input end of the multiplier M3 and the input end of the fourth convolution block respectively;

[0045] The input end of the Sigmoid activation layer is connected to the output end of the fourth convolutional block, and its output end is connected to the second input end of the multiplier M3;

[0046] The output end of the multiplier M3 serves as the output end of the first-scale attention unit or the second-scale attention unit.

[0047] Furthermore, in the first-scale attention unit, the convolution kernel size of the third convolution block is 3×3, and the convolution kernel size of the fourth convolution block is 1×1;

[0048] In the second-scale attention unit, the convolution kernel size of the third convolution block is 5×5, and the convolution kernel size of the fourth convolution block is 1×1.

[0049] The beneficial effects of the present invention are:

[0050] 1. The present invention uses both HV and HH polarization modes to collect synthetic aperture radar images of the mountain monitoring area. Compared with the traditional single polarization mode, dual-polarization data can capture the surface scattering characteristics from different angles and fully reflect the changes in physical properties during the landslide process.

[0051] 2. This method extracts the changes in radar backscatter amplitude and phase between the initial and subsequent moments to construct dual-polarization amplitude and phase change images, respectively. This allows for precise capture of subtle changes in surface features before and after a landslide. Amplitude changes reflect changes in surface structure, while phase changes reveal surface deformation information. This approach avoids misjudgments and missed detections caused by single-feature analysis, improving the reliability of early warning results.

[0052] 3. The present invention counts the number of times a pixel is marked as a landslide point, constructs an amplitude confidence distribution matrix and a phase confidence distribution matrix, and quantitatively assesses the landslide potential of different pixels. During the landslide hazard early warning neural network processing, the confidence level imposed by the confidence matrix enables the network to fully consider the credibility of the data during risk assessment, prioritize high-confidence areas, and rationally allocate computing resources, further improving the accuracy of landslide hazard risk calculations. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a flow chart of a landslide disaster early warning method;

[0054] Figure 2 This is a schematic diagram of the structure of the landslide disaster early warning neural network;

[0055] Figure 3 This is a structural diagram of the amplitude confidence applying unit;

[0056] Figure 4 It is a structural diagram of the phase position reliability application unit;

[0057] Figure 5 Schematic diagram of the structure of the amplitude feature extraction unit and the phase feature extraction unit;

[0058] Figure 6 Schematic diagram of the structure of the first-scale attention unit and the second-scale attention unit. DETAILED DESCRIPTION

[0059] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0060] like Figure 1 As shown, a landslide disaster early warning method includes:

[0061] In the HV and HH polarization modes, synthetic aperture radar images of the mountain monitoring area are collected respectively to obtain HV images and HH images;

[0062] Find the HV image and HH image at the start of the landslide event in each HV image and HH image;

[0063] According to the difference in radar backscatter amplitude between the image at the start time and the image after the start time, the dual-polarization amplitude change value is extracted and the dual-polarization amplitude change image is constructed;

[0064] According to the phase difference between the image at the start time and the image after the start time, the dual-polarization phase change value is extracted to construct a dual-polarization phase change image;

[0065] Count the number of times the same pixel is marked as the first landslide point and the second landslide point at multiple consecutive moments after the landslide event occurs, and construct the amplitude confidence distribution matrix and the phase position confidence distribution matrix;

[0066] The landslide disaster early warning neural network is used to process the dual-polarization amplitude change image and the dual-polarization phase change image. The landslide disaster risk value is obtained based on the confidence imposed by the amplitude confidence distribution matrix and the phase position confidence distribution matrix.

[0067] In this embodiment, in the HV polarization mode, the synthetic aperture radar image of the mountain monitoring area collected is an HV image. In the HH polarization mode, the synthetic aperture radar image of the mountain monitoring area collected is an HH image.

[0068] In the present invention, each HV image and HH image is an image taken by the same radar at different time points on the same area.

[0069] Radar backscatter amplitude is a measure of signal strength expressed as the absolute value of the complex signal.

[0070] In this embodiment, the process of finding the image at the starting time of the landslide event includes:

[0071] A historical period without landslides was selected, and images of the mountain monitoring area under HV and HH polarization modes were collected during this period. The same type of images at every two adjacent moments were subtracted to calculate the radar backscatter amplitude difference of each pixel point. The radar backscatter amplitude difference of each pixel point was then averaged to obtain the daily radar backscatter amplitude difference.

[0072] The same subtraction and averaging operations are performed on the two most recently acquired images at adjacent times to obtain the latest radar backscatter amplitude difference. If the latest radar backscatter amplitude difference is greater than 1.5 times the daily radar backscatter amplitude difference for the corresponding polarization mode, either of the two most recently acquired images is considered to be the image at the start of the landslide event.

[0073] In this embodiment, the process of constructing the dual-polarization amplitude change image includes:

[0074] The HV image at the start of the landslide event is subtracted from the HV image at time t to obtain the first radar backscatter amplitude difference of each pixel at time t, where t is the time number after the landslide event occurs;

[0075] Subtract the HH image at time t from the HH image at the start of the landslide event to obtain the second radar backscatter amplitude difference of each pixel at time t;

[0076] The first radar backscattering amplitude difference and the second radar backscattering amplitude difference at the same pixel position are added and normalized to obtain the dual-polarization amplitude change value at time t;

[0077] The dual-polarization amplitude change value of each pixel at time t is arranged according to the position of the pixel point to obtain the dual-polarization amplitude change image at time t.

[0078] HV polarization is sensitive to vegetation volume scattering. Landslides can cause dramatic changes in HV scattering amplitude due to vegetation collapse and soil displacement. HH polarization is more sensitive to surface scattering (e.g., bare rock and soil). Changes in surface roughness caused by landslides significantly affect HH ​​scattering. By integrating the difference in HV and HH amplitudes, we can simultaneously capture both landslide characteristics: "volume scattering failure" and "surface scattering abrupt changes."

[0079] In the mountain monitoring area, after a landslide, vegetation collapses, exposing the soil and causing HV body scattering to subside. This is manifested as a sudden decrease in radar backscatter amplitude and an increase in HH radar backscatter amplitude. Therefore, the HV image at the start of the landslide event is subtracted from the HV image at time t, and the HH image at time t is subtracted from the HH image at the start of the landslide event. The difference in the first radar backscatter amplitude and the second radar backscatter amplitude are added together to amplify the affected area.

[0080] In this embodiment, the calculation formula for the dual-polarization amplitude change value at time t is: , where ε t,i is the change value of the dual-polarization amplitude at the position of the i-th pixel at the t-th time, A 1,t,i is the first radar backscattering amplitude difference at the i-th pixel position at time t, A 2,t,i is the second radar backscattering amplitude difference at the i-th pixel position at time t, A max is the maximum radar backscatter amplitude difference to be set, i is a positive integer.

[0081] In this embodiment, the process of constructing the dual-polarization phase change image includes:

[0082] The HV image at time t is subtracted from the HV image at the start of the landslide event in phase to obtain the first phase difference of each pixel at time t, where t is the number of the time after the landslide event occurs;

[0083] Subtract the phase of the HH image at time t from the HH image at the start of the landslide event to obtain the second phase difference of each pixel at time t;

[0084] The first phase difference and the second phase difference at the same pixel position are added and normalized to obtain the dual-polarization phase change value at time t;

[0085] The dual-polarization phase change value of each pixel at time t is arranged according to the position of the pixel point to obtain the dual-polarization phase change image at time t.

[0086] By separately calculating the phase difference under HH and HV polarizations and summing them, this method can more comprehensively reflect actual surface changes. HH polarization is sensitive to surface structure (such as soil and rock), while HV polarization is more sensitive to volume scattering such as vegetation. The complementary information from these two methods can highlight subtle surface features such as landslide precursors and slow slides, improving the ability to monitor subtle deformations.

[0087] The calculation formula of the dual-polarization phase change value at time t is: , where μ t,i is the dual-polarization phase change value of the i-th pixel position at the t-th time, | | is the absolute value, θ1,t,i is the first phase difference of the position of the i-th pixel at the t-th time, θ 2,t,i is the second phase difference of the position of the i-th pixel at the t-th time, θ max The maximum phase difference is set.

[0088] In this embodiment, the process of constructing the amplitude confidence distribution matrix and the phase position confidence distribution matrix includes:

[0089] In each dual-polarization amplitude change image, the pixel whose dual-polarization amplitude change value is greater than the amplitude change threshold is marked as the first landslide point;

[0090] In the dual-polarization amplitude change images corresponding to multiple consecutive moments after the landslide event occurs, the number of times the same pixel position is marked as the first landslide point is recorded;

[0091] According to the number of times the same pixel position is marked as the first landslide point, the amplitude confidence is calculated, and the amplitude confidence of each pixel is arranged according to the position of the pixel to obtain the amplitude confidence distribution matrix;

[0092] In each dual-polarization phase change image, the pixel whose dual-polarization phase change value is greater than the phase change threshold is marked as the second landslide point;

[0093] In the dual-polarization phase change images corresponding to multiple consecutive moments after the landslide event occurs, the number of times the same pixel position is marked as the second landslide point is recorded;

[0094] The phase position reliability is calculated based on the number of times the same pixel position is marked as the second landslide point. The phase position reliability of each pixel is arranged according to the position of the pixel to obtain the phase position reliability distribution matrix.

[0095] In this embodiment, the amplitude change threshold is a threshold set for the dual-polarization amplitude change value, and the phase change threshold is a threshold set for the dual-polarization phase change value. The thresholds are adjusted and set according to experiments and requirements.

[0096] Each dual-polarization amplitude change image and dual-polarization phase change image is the difference between the image at each moment after the landslide and the image at the start of the landslide. Therefore, the landslide point is marked in each dual-polarization amplitude change image and dual-polarization phase change image. The number of times that the same pixel position is marked as a landslide point is counted to obtain a confidence level, which is used to assess the possibility that the pixel position is a landslide point. Low confidence level (few times) indicates that the pixel position may be noise; high confidence level (many times) indicates that the landslide has entered an "active phase" with continuous and significant deformation.

[0097] In this embodiment, the process of calculating the amplitude confidence includes: taking the ratio of the number of times the same pixel position is marked as the first landslide point to the total time (the total time is equal to the number of dual-polarization phase change images corresponding to multiple consecutive time points after the landslide event occurs); when the ratio is greater than or equal to 0.5, the amplitude confidence is equal to the ratio; when the ratio is less than 0.5, the amplitude confidence is 0;

[0098] The process of calculating the phase position reliability includes taking the ratio of the number of times the same pixel position is marked as the second landslide point to the total time (the total time is equal to the number of dual-polarization phase change images corresponding to multiple consecutive time points after the landslide event occurs). When the ratio is greater than or equal to 0.5, the phase position reliability is equal to the ratio; when the ratio is less than 0.5, the phase position reliability is 0.

[0099] In this embodiment, the starting time is recorded as time t0. To predict the landslide hazard risk value at time T after the landslide event, a dual-polarization amplitude change image and a dual-polarization phase change image between time t0 and time T are constructed. According to the number of times the same pixel point is marked as the first landslide point and the second landslide point between time t0 and time T, an amplitude confidence distribution matrix and a phase position confidence distribution matrix are constructed. The dual-polarization amplitude change image and the dual-polarization phase change image at time T, as well as the amplitude confidence distribution matrix and the phase position confidence distribution matrix are input into the landslide hazard early warning neural network to obtain the landslide hazard risk value.

[0100] like Figure 2 As shown, the landslide disaster warning neural network includes: an amplitude confidence applying unit, a phase position confidence applying unit, an amplitude feature extraction unit, a phase feature extraction unit, an adder A1, a first scale attention unit, a second scale attention unit, a Concat layer, a first convolution layer and a fully connected layer;

[0101] The first input end of the amplitude confidence applying unit is used to input the dual-polarization amplitude change image, the second input end thereof is used to input the amplitude confidence distribution matrix, and the output end thereof is connected to the input end of the amplitude feature extraction unit; the first input end of the phase position confidence applying unit is used to input the dual-polarization phase change image, the second input end thereof is used to input the phase position confidence distribution matrix, and the output end thereof is connected to the input end of the phase feature extraction unit;

[0102] The input end of adder A1 is connected to the output end of the amplitude feature extraction unit and the output end of the phase feature extraction unit respectively, and its output end is connected to the input end of the first scale attention unit and the input end of the second scale attention unit respectively; the input end of the Concat layer is connected to the output end of the first scale attention unit and the output end of the second scale attention unit respectively, and its output end is connected to the input end of the first convolutional layer;

[0103] The input end of the fully connected layer is connected to the output end of the first convolutional layer, and its output end serves as the output end of the landslide disaster warning neural network.

[0104] The present invention applies the confidence of the amplitude confidence distribution matrix to the features of the dual-polarization amplitude change image through an amplitude confidence application unit, thereby improving the attention of important amplitude change features. The present invention applies the confidence of the phase position confidence distribution matrix to the features of the dual-polarization phase change image through a phase position confidence application unit, thereby improving the attention of important phase change features. The features are further extracted through an amplitude feature extraction unit and a phase feature extraction unit. The two types of features are added element-wise at adder A1. The first scale attention unit and the second scale attention unit are used to apply attention to features of different scales, and the attention of important features is adaptively improved. The outputs of the two attention units are spliced ​​through the Concat layer, and then a convolution operation is performed through the first convolution layer. The fully connected layer predicts the landslide hazard risk value based on the output features of the first convolution layer.

[0105] In this embodiment, the convolution kernel size of the first convolutional layer is 3×3.

[0106] like Figure 3 As shown, the amplitude confidence applying unit includes: a second convolutional layer and a multiplier M1;

[0107] The input end of the second convolutional layer serves as the first input end of the amplitude confidence applying unit; the first input end of the multiplier M1 is connected to the output end of the second convolutional layer, its second input end serves as the second input end of the amplitude confidence applying unit, and its output end serves as the output end of the amplitude confidence applying unit.

[0108] like Figure 4 As shown, the phase position confidence applying unit includes: a third convolutional layer and a multiplier M2;

[0109] The input end of the third convolutional layer serves as the first input end of the phase position confidence application unit;

[0110] The first input terminal of the multiplier M2 is connected to the output terminal of the third convolutional layer, the second input terminal thereof serves as the second input terminal of the phase position reliability applying unit, and the output terminal thereof serves as the output terminal of the phase position reliability applying unit.

[0111] In this embodiment, the convolution kernel sizes of the second convolution layer and the third convolution layer are 1×1.

[0112] The second and third convolutional layers extract features from the dual-polarization amplitude / phase change images, exploring the scattering anomalies and phase distortion patterns caused by landslides. The amplitude / phase position confidence distribution matrix quantifies the stability of the landslide data by using the number of times the data is marked at multiple moments (a higher number indicates a higher confidence level and a higher probability of a real landslide). Multiplying the convolutional layer output pixel by pixel with M1 / M2 effectively assigns weights to the features at each location—high-confidence areas are strengthened, while low-confidence areas (likely noise or short-term disturbances) are weakened.

[0113] like Figure 5 As shown, the amplitude feature extraction unit and the phase feature extraction unit both include: a first convolution block, a second convolution block and a Maxpool layer;

[0114] The input end of the first convolution block serves as the input end of the amplitude feature extraction unit or the phase feature extraction unit, and the output end thereof is connected to the input end of the second convolution block;

[0115] The input end of the Maxpool layer is connected to the output end of the second convolution block, and its output end serves as the output end of the amplitude feature extraction unit or the phase feature extraction unit.

[0116] like Figure 6 As shown, both the first-scale attention unit and the second-scale attention unit include: a third convolution block, a fourth convolution block, a Sigmoid activation layer, and a multiplier M3;

[0117] The input end of the third convolution block serves as the input end of the first scale attention unit or the second scale attention unit, and its output end is connected to the first input end of the multiplier M3 and the input end of the fourth convolution block respectively;

[0118] The input end of the Sigmoid activation layer is connected to the output end of the fourth convolutional block, and its output end is connected to the second input end of the multiplier M3;

[0119] The output end of the multiplier M3 serves as the output end of the first-scale attention unit or the second-scale attention unit.

[0120] In this embodiment, in the first-scale attention unit, the convolution kernel size of the third convolution block is 3×3, and the convolution kernel size of the fourth convolution block is 1×1;

[0121] In the second-scale attention unit, the convolution kernel size of the third convolution block is 5×5, and the convolution kernel size of the fourth convolution block is 1×1.

[0122] The first scale (3×3 convolution block) focuses on small-scale features, and the second scale (5×5 convolution block) captures large-scale features. Attention weights are generated through the fourth convolution block and the Sigmoid activation layer, adaptively enhancing landslide-related features and weakening background noise.

[0123] Each convolution block includes a convolution layer, an activation function ReLU, and a batch normalization layer.

[0124] In this embodiment, if the landslide hazard risk value is in the range of 81-100, the mountain is experiencing a large-scale landslide, and a first-level warning is issued; if the landslide hazard risk value is in the range of 61-80, the mountain is experiencing a medium-scale landslide, and a second-level warning is issued; if the landslide hazard risk value is in the range of 41-60, the mountain is experiencing a small-scale landslide, and a third-level warning is issued; if the landslide hazard risk value is in the range of 0-40, the mountain is in a non-landslide state, and no warning is issued.

[0125] The present invention uses HV and HH polarization methods to collect synthetic aperture radar images of the mountain monitoring area. Compared with the traditional single polarization method, dual-polarization data can capture the surface scattering characteristics from different angles and fully reflect the changes in physical properties during the landslide process.

[0126] By extracting the changes in radar backscatter amplitude and phase between the initial and subsequent moments, this method constructs dual-polarization amplitude and phase change images, respectively. This allows for precise capture of subtle changes in surface features before and after a landslide. Amplitude changes reflect changes in surface structure, while phase changes reveal surface deformation information. This approach avoids misjudgments and missed detections caused by single-feature analysis, improving the reliability of early warning results.

[0127] The present invention counts the number of times a pixel is flagged as a landslide point, constructs an amplitude confidence distribution matrix and a phase confidence distribution matrix, and quantitatively assesses the landslide potential of different pixel points. During the landslide hazard early warning neural network processing, the confidence level imposed by the confidence matrix enables the network to fully consider the credibility of the data during risk assessment, prioritize high-confidence areas, and rationally allocate computing resources, further improving the accuracy of landslide hazard risk calculations.

[0128] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A landslide disaster early warning method, characterized in that: include: In the HV and HH polarization modes, synthetic aperture radar images of the mountain monitoring area are collected respectively to obtain HV images and HH images; Find the HV image and HH image at the start of the landslide event in each HV image and HH image; According to the difference in radar backscatter amplitude between the image at the start time and the image after the start time, the dual-polarization amplitude change value is extracted and the dual-polarization amplitude change image is constructed; According to the phase difference between the image at the start time and the image after the start time, the dual-polarization phase change value is extracted to construct a dual-polarization phase change image; Count the number of times the same pixel is marked as the first landslide point and the second landslide point at multiple consecutive moments after the landslide event occurs, and construct the amplitude confidence distribution matrix and the phase position confidence distribution matrix; The landslide disaster early warning neural network is used to process the dual-polarization amplitude change image and the dual-polarization phase change image. The landslide disaster risk value is obtained based on the confidence imposed by the amplitude confidence distribution matrix and the phase position confidence distribution matrix.

2. The landslide disaster early warning method according to claim 1, characterized in that: The process of constructing a dual-polarization amplitude change image includes: The HV image at the start of the landslide event is subtracted from the HV image at time t to obtain the first radar backscatter amplitude difference of each pixel at time t, where t is the time number after the landslide event occurs; Subtract the HH image at time t from the HH image at the start of the landslide event to obtain the second radar backscatter amplitude difference of each pixel at time t; The first radar backscattering amplitude difference and the second radar backscattering amplitude difference at the same pixel position are added and normalized to obtain the dual-polarization amplitude change value at time t; The dual-polarization amplitude change value of each pixel at time t is arranged according to the position of the pixel point to obtain the dual-polarization amplitude change image at time t.

3. The landslide disaster early warning method according to claim 1, characterized in that: The process of constructing a dual-polarization phase change image includes: The HV image at time t is subtracted from the HV image at the start of the landslide event in phase to obtain the first phase difference of each pixel at time t, where t is the number of the time after the landslide event occurs; Subtract the phase of the HH image at time t from the HH image at the start of the landslide event to obtain the second phase difference of each pixel at time t; The first phase difference and the second phase difference at the same pixel position are added and normalized to obtain the dual-polarization phase change value at time t; The dual-polarization phase change value of each pixel at time t is arranged according to the position of the pixel point to obtain the dual-polarization phase change image at time t.

4. The landslide disaster early warning method according to claim 1, characterized in that: The process of constructing the amplitude confidence distribution matrix and the phase position confidence distribution matrix includes: In each dual-polarization amplitude change image, the pixel whose dual-polarization amplitude change value is greater than the amplitude change threshold is marked as the first landslide point; In the dual-polarization amplitude change images corresponding to multiple consecutive moments after the landslide event occurs, the number of times the same pixel position is marked as the first landslide point is recorded; According to the number of times the same pixel position is marked as the first landslide point, the amplitude confidence is calculated, and the amplitude confidence of each pixel is arranged according to the position of the pixel to obtain the amplitude confidence distribution matrix; In each dual-polarization phase change image, the pixel whose dual-polarization phase change value is greater than the phase change threshold is marked as the second landslide point; In the dual-polarization phase change images corresponding to multiple consecutive moments after the landslide event occurs, the number of times the same pixel position is marked as the second landslide point is recorded; The phase position reliability is calculated based on the number of times the same pixel position is marked as the second landslide point. The phase position reliability of each pixel is arranged according to the position of the pixel to obtain the phase position reliability distribution matrix.

5. The landslide disaster early warning method according to claim 4, characterized in that: The process of calculating the amplitude confidence includes: taking the ratio of the number of times the same pixel position is marked as the first landslide point to the total time, when the ratio is greater than or equal to 0.5, the amplitude confidence is equal to the ratio, when the ratio is less than 0.5, the amplitude confidence is 0; The process of calculating the phase position reliability includes: taking the ratio of the number of times the same pixel position is marked as the second landslide point to the total time. When the ratio is greater than or equal to 0.5, the phase position reliability is equal to the ratio; when the ratio is less than 0.5, the phase position reliability is 0.

6. The landslide disaster early warning method according to claim 1, characterized in that: The landslide disaster warning neural network includes: an amplitude confidence applying unit, a phase position confidence applying unit, an amplitude feature extraction unit, a phase feature extraction unit, an adder A1, a first scale attention unit, a second scale attention unit, a Concat layer, a first convolution layer and a fully connected layer; The first input end of the amplitude confidence applying unit is used to input the dual-polarization amplitude change image, the second input end thereof is used to input the amplitude confidence distribution matrix, and the output end thereof is connected to the input end of the amplitude feature extraction unit; the first input end of the phase position confidence applying unit is used to input the dual-polarization phase change image, the second input end thereof is used to input the phase position confidence distribution matrix, and the output end thereof is connected to the input end of the phase feature extraction unit; The input end of adder A1 is connected to the output end of the amplitude feature extraction unit and the output end of the phase feature extraction unit respectively, and its output end is connected to the input end of the first scale attention unit and the input end of the second scale attention unit respectively; the input end of the Concat layer is connected to the output end of the first scale attention unit and the output end of the second scale attention unit respectively, and its output end is connected to the input end of the first convolutional layer; The input end of the fully connected layer is connected to the output end of the first convolutional layer, and its output end serves as the output end of the landslide disaster warning neural network.

7. The landslide disaster early warning method according to claim 6, characterized in that: The amplitude confidence applying unit includes: a second convolutional layer and a multiplier M1; The input end of the second convolutional layer serves as the first input end of the amplitude confidence applying unit; the first input end of the multiplier M1 is connected to the output end of the second convolutional layer, the second input end of the multiplier M1 serves as the second input end of the amplitude confidence applying unit, and the output end of the multiplier M1 serves as the output end of the amplitude confidence applying unit; The phase position confidence applying unit includes: a third convolutional layer and a multiplier M2; The input end of the third convolutional layer serves as the first input end of the phase position confidence application unit; The first input terminal of the multiplier M2 is connected to the output terminal of the third convolutional layer, the second input terminal thereof serves as the second input terminal of the phase position reliability applying unit, and the output terminal thereof serves as the output terminal of the phase position reliability applying unit.

8. The landslide disaster early warning method according to claim 6, characterized in that: The amplitude feature extraction unit and the phase feature extraction unit both include: a first convolution block, a second convolution block and a Maxpool layer; The input end of the first convolution block serves as the input end of the amplitude feature extraction unit or the phase feature extraction unit, and the output end thereof is connected to the input end of the second convolution block; The input end of the Maxpool layer is connected to the output end of the second convolution block, and its output end serves as the output end of the amplitude feature extraction unit or the phase feature extraction unit.

9. The landslide disaster early warning method according to claim 6, characterized in that: Both the first-scale attention unit and the second-scale attention unit include: a third convolution block, a fourth convolution block, a Sigmoid activation layer, and a multiplier M3; The input end of the third convolution block serves as the input end of the first scale attention unit or the second scale attention unit, and its output end is connected to the first input end of the multiplier M3 and the input end of the fourth convolution block respectively; The input end of the Sigmoid activation layer is connected to the output end of the fourth convolutional block, and its output end is connected to the second input end of the multiplier M3; The output end of the multiplier M3 serves as the output end of the first-scale attention unit or the second-scale attention unit.

10. The landslide disaster early warning method according to claim 9, characterized in that: In the first-scale attention unit, the convolution kernel size of the third convolution block is 3×3, and the convolution kernel size of the fourth convolution block is 1×1; In the second-scale attention unit, the convolution kernel size of the third convolution block is 5×5, and the convolution kernel size of the fourth convolution block is 1×1.

Images