A landslide risk prediction method based on a landslide disaster evaluation model

Abstract

The application discloses a landslide risk prediction method based on a landslide disaster evaluation model and belongs to the technical field of image processing. The application firstly extracts a remote sensing image of a region to be monitored, obtains a channel ground surface matching value and a channel soil matching value, and constructs an R, G and B channel soil ground surface contrast graph; then takes a plurality of adjacent time contrast graphs, obtains soil ground surface characteristic values according to soil ground surface contrast value changes, and constructs a characteristic graph; then constructs a cumulative landslide weight matrix according to a landslide weight coefficient of a suspected variable pixel point, and constructs a cumulative soil ground surface characteristic graph according to the soil ground surface characteristic values; finally, the application processes the cumulative landslide weight matrix and the cumulative soil ground surface characteristic graph corresponding to the three channels by using a landslide disaster evaluation model, and obtains a landslide risk score. The application improves the accuracy of landslide risk prediction.

Description

Technical Field

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

[0002] With the continuous development of remote sensing technology, landslide identification based on remote sensing images has become an important means of early warning for modern geological disasters. Currently, existing remote sensing landslide identification technologies mainly employ multi-source image fusion and feature extraction methods. A typical technical approach includes: first, acquiring surface images of the target area through multispectral or high-resolution satellite imagery; then, using image processing techniques to extract key geological features. Current landslide identification solutions based on remote sensing images analyze spectral and textural information from remote sensing images at a single moment to extract features of the landslide area and thus identify landslide signs. However, existing methods rely heavily on image data from single or limited time points, making it difficult to accurately capture subtle but continuous soil deformation changes. This is especially true for slowly developing potential landslide areas, where the small magnitude of changes often leads to delays and inaccuracies in landslide risk warnings, significantly limiting the effectiveness of disaster prevention. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, this invention provides a landslide risk prediction method based on a landslide disaster assessment model, which solves the problem of low accuracy in landslide risk early warning in existing technologies.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: a landslide risk prediction method based on a landslide disaster assessment model, comprising the following steps:

[0005] S1. Extract remote sensing images of the area to be monitored, obtain channel surface matching values ​​and channel soil matching values, and construct soil and surface comparison maps of R, G, and B channels;

[0006] S2. Take the channel soil surface comparison map at multiple adjacent time points, and obtain the soil surface characteristic value based on the change of soil surface comparison value in the channel soil surface comparison map at adjacent time points, and construct the soil surface characteristic map.

[0007] S3. Construct a cumulative landslide weight matrix based on the landslide weight coefficients of suspected changing pixels in the soil surface feature map at each time point;

[0008] S4. Construct a cumulative soil surface feature map based on the soil surface feature values ​​in the soil surface feature map at each time point;

[0009] S5. The landslide risk score is obtained by processing the cumulative landslide weight matrix and cumulative soil surface feature map corresponding to the three channels using the landslide disaster assessment model.

[0010] Furthermore, S1 includes the following sub-steps:

[0011] S11. Based on the matching values ​​of each channel value of each pixel in the remote sensing image and the corresponding channel values ​​of the stored surface pixels, obtain the R channel surface matching value, G channel surface matching value and B channel surface matching value;

[0012] S12. Based on the matching values ​​of each channel value of each pixel in the remote sensing image and the corresponding channel values ​​of the stored soil pixels, obtain the soil matching values ​​of the R channel, G channel, and B channel.

[0013] S13. Take the ratio of the soil matching value of the R channel to the surface matching value of the R channel as the R channel comparison value, replace the channel value of the original pixel with the R channel comparison value, and obtain the R channel soil and surface comparison map.

[0014] S14. Take the ratio of the soil matching value of the G channel to the surface matching value of the G channel as the G channel comparison value, replace the channel value of the original pixel with the G channel comparison value, and obtain the G channel soil and surface comparison map.

[0015] S15. Take the ratio of the soil matching value of channel B to the surface matching value of channel B as the contrast value of channel B. Replace the channel value of the original pixel with the contrast value of channel B to obtain the soil and surface contrast map of channel B.

[0016] Furthermore, S2 includes the following sub-steps:

[0017] S21. For the same channel, in the channel soil surface comparison map at adjacent time points, the channel soil surface comparison map at the later time point is subtracted from the channel soil surface comparison map at the earlier time point to obtain the soil surface comparison difference value of each pixel.

[0018] S22. At the same pixel, the ratio of the soil surface contrast difference to the soil surface contrast value at the previous time step is taken as the soil surface feature value of that pixel.

[0019] S23. Replace the soil surface feature values ​​with the channel values ​​of the original pixels to obtain the soil surface feature map.

[0020] Furthermore, S3 includes the following sub-steps:

[0021] S31. Take the average value of the soil surface feature values ​​on the soil surface feature map at each time point to obtain the soil surface feature threshold.

[0022] S32. Mark pixels whose soil surface feature values ​​are greater than the soil surface feature threshold as suspected variable pixels.

[0023] S33. Obtain the landslide weight coefficient for suspected variable pixels, set the landslide weight coefficient of non-suspected variable pixels to 0, arrange each landslide weight coefficient according to the corresponding pixel, and obtain the landslide weight coefficient matrix at each time.

[0024] S34. Add the landslide weight coefficient matrices of multiple consecutive time points to obtain the cumulative landslide weight matrix.

[0025] Furthermore, the process of obtaining the landslide weight coefficient in S33 is as follows: count the number of suspected changed pixels within the 5×5 neighborhood of the suspected changed pixel, and use the ratio of the number of suspected changed pixels in the neighborhood to the total number of pixels in the 5×5 neighborhood as the landslide weight coefficient.

[0026] Furthermore, S4 includes the following sub-steps:

[0027] S41. When the soil surface feature value of a pixel in the soil surface feature map is less than the soil surface feature threshold, set the soil surface feature value of that pixel to 0, and retain other soil surface feature values ​​to obtain a corrected soil surface feature map.

[0028] S42. Add the corrected soil surface feature maps from multiple consecutive time points to obtain the cumulative soil surface feature map.

[0029] Furthermore, the landslide hazard assessment model in S5 includes: three soil surface feature weight fusion units, three feature extraction units, an adder A1, a CNN network, and a fully connected layer;

[0030] The first input of each soil surface feature weight fusion unit is used to input the cumulative landslide weight matrix corresponding to one channel, and its second input is used to input the cumulative soil surface feature map corresponding to the same channel. The input of the feature extraction unit is connected to the output of the soil surface feature weight fusion unit. The input of adder A1 is connected to the output of the three feature extraction units respectively, and its output is connected to the input of the CNN network. The input of the fully connected layer is connected to the output of the CNN network, and its output is used as the output of the landslide disaster assessment model.

[0031] Furthermore, each soil surface feature weight fusion unit includes: a first convolutional block, a second convolutional block, a first maximum pooling layer, a second maximum pooling layer, a first average pooling layer, a second average pooling layer, a multiplier M1, a multiplier M2, and a first Concat layer;

[0032] The input of the first convolutional block is used to input the cumulative landslide weight matrix, and its output is connected to the input of the first max-pooling layer and the first average-pooling layer, respectively. The input of the second convolutional block is used to input the cumulative soil surface feature map, and its output is connected to the input of the second max-pooling layer and the second average-pooling layer, respectively. The input of multiplier M1 is connected to the output of the first max-pooling layer and the second max-pooling layer, respectively. The input of multiplier M2 is connected to the output of the first average-pooling layer and the second average-pooling layer, respectively. The input of the first Concat layer is connected to the output of multiplier M1 and the output of multiplier M2, respectively, and its output serves as the output of the soil surface feature weight fusion unit.

[0033] Furthermore, the feature extraction unit includes: a third convolutional block, a fourth convolutional block, a fifth convolutional block, a first residual block, a second residual block, and a second Concat layer;

[0034] The input of the third convolutional block is connected to the input of the fourth convolutional block and serves as the input of the feature extraction unit; the output of the third convolutional block is connected to the input of the first residual block; the output of the first residual block is connected to the input of the second residual block; the input of the second Concat layer is connected to the output of the fourth convolutional block and the output of the second residual block, respectively, and its output is connected to the input of the fifth convolutional block; the output of the fifth convolutional block serves as the output of the feature extraction unit.

[0035] The beneficial effects of this invention are as follows: By extracting remote sensing images of the area to be monitored and constructing soil surface comparison maps for R, G, and B channels, the channel value differences between soil and surface are transformed into comparison values, enhancing the identification of landslide features and making landslide signs easier to detect. Next, by taking soil surface comparison maps from multiple adjacent time points and obtaining soil surface feature values ​​based on changes in the soil surface comparison values, a soil surface feature map is constructed, which can capture soil changes at different times, overcoming the shortcomings of existing technologies in capturing minute continuous deformations. Then, by constructing a cumulative landslide weight matrix and a cumulative soil surface feature map, the landslide situation and soil surface feature value changes at each pixel location over a period of time are recorded. Finally, a landslide disaster assessment model is used to process the cumulative data corresponding to the three channels, and by integrating the information from each channel, a more accurate landslide risk score is obtained, effectively improving the timeliness and accuracy of landslide risk warning and solving the problem of low warning accuracy in existing technologies.

[0036] This invention significantly enhances the saliency of landslide characteristics by constructing soil surface comparison values, making them easier to detect. At the same time, by combining landslide weight coefficients and soil surface characteristic values ​​over a period of time, it further improves the sensitivity of landslide identification, effectively solving the problem of untimely and inaccurate risk warnings caused by small deformation amplitude when facing slowly developing potential landslide areas in existing technologies. Attached Figure Description

[0037] Figure 1 This is a flowchart of a landslide risk prediction method based on a landslide disaster assessment model;

[0038] Figure 2 This is a schematic diagram of a landslide.

[0039] Figure 3 for Figure 2 R, G, and B channel values ​​within a 3×3 area of ​​the landslide region;

[0040] Figure 4 for Figure 2 R, G, and B channel values ​​within a 3×3 area in the landslide region of Central Africa;

[0041] Figure 5 This is a schematic diagram of the landslide hazard assessment model.

[0042] Figure 6 A schematic diagram of the structure of the soil surface feature weighting fusion unit;

[0043] Figure 7 This is a schematic diagram of the feature extraction unit.

[0044] Figure 8 This is a schematic diagram of the structure of the first residual block and the second residual block. Detailed Implementation

[0045] 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.

[0046] like Figure 1 As shown, a landslide risk prediction method based on a landslide hazard assessment model includes the following steps:

[0047] S1. Extract remote sensing images of the area to be monitored, obtain channel surface matching values ​​and channel soil matching values, and construct soil and surface comparison maps of R, G, and B channels;

[0048] S2. Take the channel soil surface comparison map at multiple adjacent time points, and obtain the soil surface characteristic value based on the change of soil surface comparison value in the channel soil surface comparison map at adjacent time points, and construct the soil surface characteristic map.

[0049] S3. Construct a cumulative landslide weight matrix based on the landslide weight coefficients of suspected changing pixels in the soil surface feature map at each time point;

[0050] S4. Construct a cumulative soil surface feature map based on the soil surface feature values ​​in the soil surface feature map at each time point;

[0051] S5. The landslide risk score is obtained by processing the cumulative landslide weight matrix and cumulative soil surface feature map corresponding to the three channels using the landslide disaster assessment model.

[0052] In this embodiment, S1 includes the following sub-steps:

[0053] S11. Based on the matching values ​​of each channel value of each pixel in the remote sensing image and the corresponding channel values ​​of the stored surface pixels, obtain the R channel surface matching value, G channel surface matching value and B channel surface matching value;

[0054] S12. Based on the matching values ​​of each channel value of each pixel in the remote sensing image and the corresponding channel values ​​of the stored soil pixels, obtain the soil matching values ​​of the R channel, G channel, and B channel.

[0055] S13. Take the ratio of the soil matching value of the R channel to the surface matching value of the R channel as the R channel comparison value, replace the channel value of the original pixel with the R channel comparison value, and obtain the R channel soil and surface comparison map.

[0056] S14. Take the ratio of the soil matching value of the G channel to the surface matching value of the G channel as the G channel comparison value, replace the channel value of the original pixel with the G channel comparison value, and obtain the G channel soil and surface comparison map.

[0057] S15. Take the ratio of the soil matching value of channel B to the surface matching value of channel B as the contrast value of channel B. Replace the channel value of the original pixel with the contrast value of channel B to obtain the soil and surface contrast map of channel B.

[0058] In this embodiment, S11 specifically involves: obtaining an R-channel surface matching value based on the matching value between the R-channel value of each pixel in the remote sensing image and the R-channel value of the stored surface pixel; obtaining a G-channel surface matching value based on the matching value between the G-channel value of each pixel in the remote sensing image and the G-channel value of the stored surface pixel; and obtaining a B-channel surface matching value based on the matching value between the B-channel value of each pixel in the remote sensing image and the B-channel value of the stored surface pixel.

[0059] In this embodiment, S12 specifically involves: obtaining an R-channel soil matching value based on the matching value between the R-channel value of each pixel in the remote sensing image and the R-channel value of the stored soil pixel; obtaining a G-channel soil matching value based on the matching value between the G-channel value of each pixel in the remote sensing image and the G-channel value of the stored soil pixel; and obtaining a B-channel soil matching value based on the matching value between the B-channel value of each pixel in the remote sensing image and the B-channel value of the stored soil pixel.

[0060] In this embodiment, the formula for calculating the matching value between two channel values ​​is: , where γ i v is the matching value of the i-th pixel in the remote sensing image. i v is the channel value of the i-th pixel in the remote sensing image. s,i Let | be the channel value of the i-th stored pixel, and | be the absolute value.

[0061] In this embodiment, the stored surface pixels are pre-stored pixels of the non-landslide surface, and the stored soil pixels are pre-stored pixels of the soil after a landslide.

[0062] This embodiment extracts Figure 2 The R-channel values, G-channel values, and B-channel values ​​for a 3×3 area within the landslide region are shown below. Figure 3 As shown.

[0063] This embodiment extracts Figure 2 The R-channel values, G-channel values, and B-channel values ​​of the 3×3 area of ​​the landslide region in Central Africa are shown below. Figure 4 As shown. By comparison Figure 3 and Figure 3 It is evident that there are significant differences in the R, G, and B channel values ​​between landslide areas and non-landslide areas. In remote sensing images, different land features (landslide areas and non-landslide areas) exhibit different reflectance and absorption characteristics in the R, G, and B channels (corresponding to different spectral bands) due to differences in material composition and structure. The spectral responses of soil and rock materials in landslide areas differ from those in non-landslide areas, which is reflected in the differences in the R, G, and B channel values.

[0064] This invention constructs a soil-surface contrast map using R, G, and B channels, transforming the differences in spectral channel values ​​between soil and surface in remote sensing images into contrast values. Based on the principle of spectral characteristics of land features, different land features have inherent differences in reflectance across spectral channels. By calculating the contrast values, the differences in spectral features between soil and surface can be effectively amplified, greatly enhancing the recognizability of landslide features in images. Compared to existing technologies that analyze images at a single moment, this invention can more accurately capture landslide signs.

[0065] In this embodiment, S2 is performed on the same channel, and S2 includes the following sub-steps:

[0066] S21. For the same channel, in the channel soil surface comparison map at adjacent time points, the channel soil surface comparison map at the later time point is subtracted from the channel soil surface comparison map at the earlier time point to obtain the soil surface comparison difference value of each pixel.

[0067] S22. At the same pixel, the ratio of the soil surface contrast difference to the soil surface contrast value at the previous time step is taken as the soil surface feature value of that pixel.

[0068] S23. Replace the soil surface feature values ​​with the channel values ​​of the original pixels to obtain the soil surface feature map.

[0069] In this embodiment, for process S2, the channel soil surface comparison images of the same channel at 7 time points t-5, t-4, t-3, t-2, t-1, t, and t+1 are selected. The channel soil surface comparison images of adjacent time points t-4 and t-5, t-3 and t-4, t-2 and t-3, t-1 and t-2, t and t-1, and t+1 and t are subtracted. For the same pixel position, the soil surface comparison difference corresponding to 6 time points is obtained. After processing according to S22 and S23, the soil surface feature maps of 6 time points can be obtained.

[0070] In this embodiment, the formula for calculating the soil surface feature value of a pixel is: , where α t,i Let C be the soil surface feature value of the i-th pixel in the soil surface contrast image at time t. t,i Let C be the soil surface contrast value of the i-th pixel in the soil surface contrast image at time t. t+1,i Let be the soil surface contrast value of the i-th pixel in the soil surface contrast image at time t+1, where i is a positive integer and || represents the absolute value.

[0071] In remote sensing monitoring, landslide occurrence is a time-varying process, and subtle soil deformations will manifest differently in remote sensing image channel data at different times. S21 calculates the soil surface contrast difference by subtracting the soil surface contrast images of adjacent time points in the same channel, transforming the potential changes in channel values ​​caused by landslides into differences, and accurately capturing the changes of pixels at the same location over time. Meanwhile, S22 calculates soil surface feature values ​​by using the soil surface contrast value from the previous time point as the denominator for ratio calculation, which converts the difference into a relative degree of change, eliminating the influence of differences in the magnitude of original channel values ​​in different areas.

[0072] In this embodiment, S3 is performed on the same channel, and S3 includes the following sub-steps:

[0073] S31. Take the average value of the soil surface feature values ​​on the soil surface feature map at each time point to obtain the soil surface feature threshold.

[0074] S32. Mark pixels whose soil surface feature values ​​are greater than the soil surface feature threshold as suspected variable pixels.

[0075] S33. Obtain the landslide weight coefficient for suspected changed pixels, set the landslide weight coefficient of non-suspected changed pixels to 0, arrange each landslide weight coefficient according to the corresponding pixel, and obtain the landslide weight coefficient matrix at each time. That is, arrange the landslide weight coefficients of all pixels according to the spatial position correspondence of the pixels in the image, and construct the landslide weight coefficient matrix corresponding to each time.

[0076] S34. Add the landslide weight coefficient matrices of multiple consecutive time points to obtain the cumulative landslide weight matrix.

[0077] In this embodiment, the time period of 10 seconds or 5 seconds can be selected in S34.

[0078] In this embodiment, the process of obtaining the landslide weight coefficient in S33 is as follows: count the number of suspected changed pixels within the 5×5 neighborhood of the suspected changed pixel, and use the ratio of the number of suspected changed pixels in the neighborhood to the total number of pixels in the 5×5 neighborhood as the landslide weight coefficient.

[0079] The expression for the landslide weight coefficient is: , where θ is the landslide weight coefficient and N is the number of suspected changing pixels in the neighborhood.

[0080] When a landslide occurs, its impact is not limited to isolated single pixels, but rather has a regional and continuous effect. The sliding of the landslide body will cause displacement or deformation of the surrounding soil and rock within a certain range. If a pixel belongs to a real landslide area, there will often be more similar points of change in its neighborhood; conversely, if it is an isolated anomaly caused by random noise or irrelevant interference, the number of similar points in its neighborhood is usually smaller.

[0081] For slowly developing potential landslide areas, where the landslide area gradually increases and occurs locally within the monitoring area, this invention takes the average value of the soil surface feature values ​​on the soil surface feature map at each time point to obtain the soil surface feature threshold. Pixels with values ​​greater than the soil surface feature threshold are then filtered out, i.e., the pixels that have changed in adjacent time points are obtained.

[0082] This invention adds up the landslide weight coefficient matrices of multiple consecutive time points, and can display the position of each pixel point over a period of time in a cumulative landslide weight matrix.

[0083] In this embodiment, S4 is performed on the same channel, and S4 includes the following sub-steps:

[0084] S41. When the soil surface feature value of a pixel in the soil surface feature map is less than the soil surface feature threshold, set the soil surface feature value of that pixel to 0, and retain other soil surface feature values ​​to obtain a corrected soil surface feature map.

[0085] S42. Add the corrected soil surface feature maps from multiple consecutive time points to obtain the cumulative soil surface feature map.

[0086] Remote sensing image data contains a large amount of background noise and random fluctuations unrelated to landslides. The soil surface feature values ​​corresponding to these are usually small and within the main range of the data distribution. However, the feature values ​​corresponding to feature changes caused by landslides often deviate from the main distribution and exceed a threshold. By setting feature values ​​below the threshold to 0, background noise and irrelevant interference information can be effectively filtered out, retaining only the effective signals related to landslides, highlighting the features of the landslide area, and making landslide information more significant. The occurrence and development of landslides is a dynamic time process, and an image at a single moment may not fully reflect the overall picture and development trend of a landslide. By accumulating the modified feature maps at multiple moments, landslide-related information at different moments can be superimposed, strengthening the temporal dimension of landslide features. For example, if the soil surface feature value of pixel A on the soil surface feature map at time t is 0 (no landslide has occurred at this time), and a landslide occurs at pixel A at time t+1, then by adding the values, the location of the landslide over a period of time can be highlighted.

[0087] like Figure 5 As shown, the landslide disaster assessment model in S5 includes: three soil surface feature weight fusion units, three feature extraction units, adder A1, CNN network and fully connected layer;

[0088] The first input of each soil surface feature weight fusion unit is used to input the cumulative landslide weight matrix corresponding to one channel, and its second input is used to input the cumulative soil surface feature map corresponding to the same channel. The input of the feature extraction unit is connected to the output of the soil surface feature weight fusion unit. The input of adder A1 is connected to the output of the three feature extraction units respectively, and its output is connected to the input of the CNN network. The input of the fully connected layer is connected to the output of the CNN network, and its output is used as the output of the landslide disaster assessment model.

[0089] The cumulative landslide weight matrix reflects the probability of a pixel experiencing a landslide over a period of time, while the cumulative soil surface feature map records the cumulative changes in soil surface features over time. Through fusion, the model can enhance landslide features from two dimensions: "cumulative probability" and "cumulative change trend," making the differences between landslide and non-landslide areas more significant. Shallow features are extracted by a feature extraction unit, and the features from the three channels are fused using an adder A1. The CNN network further extracts deep features, and the fully connected layer performs the final mapping of the features, outputting a landslide risk score.

[0090] like Figure 6 As shown, each soil surface feature weight fusion unit includes: a first convolutional block, a second convolutional block, a first maximum pooling layer, a second maximum pooling layer, a first average pooling layer, a second average pooling layer, a multiplier M1, a multiplier M2, and a first Concat layer.

[0091] The input of the first convolutional block is used to input the cumulative landslide weight matrix, and its output is connected to the input of the first max-pooling layer and the first average-pooling layer, respectively. The input of the second convolutional block is used to input the cumulative soil surface feature map, and its output is connected to the input of the second max-pooling layer and the second average-pooling layer, respectively. The input of multiplier M1 is connected to the output of the first max-pooling layer and the second max-pooling layer, respectively. The input of multiplier M2 is connected to the output of the first average-pooling layer and the second average-pooling layer, respectively. The input of the first Concat layer is connected to the output of multiplier M1 and the output of multiplier M2, respectively, and its output serves as the output of the soil surface feature weight fusion unit.

[0092] This invention sets up max pooling and average pooling layers after the first and second convolutional blocks to extract salient features and global mean features, respectively. Then, multiplier M1 multiplies the outputs of the first and second max pooling layers element-wise to achieve fusion enhancement of the two types of salient features. Multiplier M2 multiplies the outputs of the first and second average pooling layers element-wise to achieve fusion enhancement of the two types of global mean features. Finally, the output is concatenated through the first concat layer.

[0093] like Figure 7 As shown, the feature extraction unit includes: a third convolutional block, a fourth convolutional block, a fifth convolutional block, a first residual block, a second residual block, and a second Concat layer;

[0094] The input of the third convolutional block is connected to the input of the fourth convolutional block and serves as the input of the feature extraction unit; the output of the third convolutional block is connected to the input of the first residual block; the output of the first residual block is connected to the input of the second residual block; the input of the second Concat layer is connected to the output of the fourth convolutional block and the output of the second residual block, respectively, and its output is connected to the input of the fifth convolutional block; the output of the fifth convolutional block serves as the output of the feature extraction unit.

[0095] In this embodiment, the first, second, third, fourth, and fifth convolutional blocks each include a convolutional layer, a ReLU layer, and a normalization layer. The kernel size of the convolutional layers in the first and second convolutional blocks is 1×1, the kernel size of the convolutional layers in the third convolutional block is 3×3, the kernel size of the convolutional layers in the fourth convolutional block is 5×5, and the kernel size of the convolutional layers in the fifth convolutional block is 3×3.

[0096] like Figure 8 As shown, both the first residual block and the second residual block include: a first convolutional layer, a first ReLU layer, a second convolutional layer, an adder A2, and a second ReLU layer. The kernel size of the first convolutional layer and the second convolutional layer is 3×3.

[0097] The third and fourth convolutional blocks extract features at different scales. The third convolutional block is used to extract local detail features. The first and second residual blocks are set to enhance the transmission of useful features and gradually amplify these weak features, making them easier for the model to capture. The fourth convolutional block is used to extract global features. The second concat layer concatenates the features of the two paths to achieve the fusion of detail features and global features.

[0098] In terms of dynamic monitoring, this invention obtains soil surface characteristic values ​​and constructs a characteristic map based on the changes in soil surface comparison values ​​in adjacent time-series soil surface comparison maps. Landslide occurrence is a dynamic evolutionary process, with subtle soil deformations accumulating gradually. By analyzing multi-time-series data, this invention can capture the continuous changing trends of soil surface characteristics, transforming minute but continuous soil deformations into quantifiable characteristic values. Whether it's a slowly developing potential landslide area or the early signs of a sudden landslide, everything can be recorded, effectively solving the problems of existing technologies' difficulty in monitoring subtle continuous changes and delayed early warning.

[0099] In the comprehensive assessment phase, this invention assigns landslide weight coefficients to suspected variable pixels in the soil surface feature map at various time points, constructing a cumulative landslide weight matrix and a cumulative soil surface feature map. These data are then comprehensively processed using a landslide hazard assessment model. Different regions have varying degrees of potential impact on landslide occurrence; the weight coefficients highlight key hazard areas. The cumulative matrix and feature map constructed from multi-time-point data comprehensively reflect the landslide risk accumulation process at each pixel location over a period of time. The landslide hazard assessment model performs comprehensive calculations based on multi-channel, multi-time-period data, fully considering the characteristic differences under different spectral channels, comprehensively quantifying landslide risk from both spatial and temporal dimensions, and improving the accuracy of landslide risk scoring.

[0100] 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 risk prediction method based on a landslide hazard assessment model, characterized in that, Includes the following steps: S1. Extract remote sensing images of the area to be monitored, obtain channel surface matching values ​​and channel soil matching values, and construct soil and surface comparison maps of R, G, and B channels; S2. Take the channel soil surface comparison map at multiple adjacent time points, and obtain the soil surface characteristic value based on the change of soil surface comparison value in the channel soil surface comparison map at adjacent time points, and construct the soil surface characteristic map. S3. Construct a cumulative landslide weight matrix based on the landslide weight coefficients of suspected changing pixels in the soil surface feature map at each time point; S4. Construct a cumulative soil surface feature map based on the soil surface feature values ​​in the soil surface feature map at each time point; S5. The landslide risk score is obtained by processing the cumulative landslide weight matrix and cumulative soil surface feature map corresponding to the three channels using the landslide disaster assessment model. The landslide disaster assessment model in S5 includes: three soil surface feature weight fusion units, three feature extraction units, an adder A1, a CNN network, and a fully connected layer; The first input of each soil surface feature weight fusion unit is used to input the cumulative landslide weight matrix corresponding to one channel, and its second input is used to input the cumulative soil surface feature map corresponding to the same channel; the input of the feature extraction unit is connected to the output of the soil surface feature weight fusion unit; the input of adder A1 is connected to the output of the three feature extraction units respectively, and its output is connected to the input of the CNN network; the input of the fully connected layer is connected to the output of the CNN network, and its output serves as the output of the landslide disaster assessment model; Each soil surface feature weight fusion unit includes: a first convolutional block, a second convolutional block, a first maximum pooling layer, a second maximum pooling layer, a first average pooling layer, a second average pooling layer, a multiplier M1, a multiplier M2, and a first Concat layer; The input of the first convolutional block is used to input the cumulative landslide weight matrix, and its output is connected to the input of the first max-pooling layer and the first average-pooling layer, respectively. The input of the second convolutional block is used to input the cumulative soil surface feature map, and its output is connected to the input of the second max-pooling layer and the second average-pooling layer, respectively. The input of multiplier M1 is connected to the output of the first max-pooling layer and the second max-pooling layer, respectively. The input of multiplier M2 is connected to the output of the first average-pooling layer and the second average-pooling layer, respectively. The input of the first Concat layer is connected to the output of multiplier M1 and the output of multiplier M2, respectively, and its output serves as the output of the soil surface feature weight fusion unit. The feature extraction unit includes: a third convolutional block, a fourth convolutional block, a fifth convolutional block, a first residual block, a second residual block, and a second Concat layer; The input of the third convolutional block is connected to the input of the fourth convolutional block and serves as the input of the feature extraction unit; the output of the third convolutional block is connected to the input of the first residual block; the output of the first residual block is connected to the input of the second residual block; the input of the second Concat layer is connected to the output of the fourth convolutional block and the output of the second residual block, respectively, and its output is connected to the input of the fifth convolutional block; the output of the fifth convolutional block serves as the output of the feature extraction unit.

2. The landslide risk prediction method based on a landslide disaster assessment model according to claim 1, characterized in that, S1 includes the following steps: S11. Based on the matching values ​​of each channel value of each pixel in the remote sensing image and the corresponding channel values ​​of the stored surface pixels, obtain the R channel surface matching value, G channel surface matching value and B channel surface matching value; S12. Based on the matching values ​​of each channel value of each pixel in the remote sensing image and the corresponding channel values ​​of the stored soil pixels, obtain the soil matching values ​​of the R channel, G channel, and B channel. S13. Take the ratio of the soil matching value of the R channel to the surface matching value of the R channel as the R channel comparison value, replace the channel value of the original pixel with the R channel comparison value, and obtain the R channel soil and surface comparison map. S14. Take the ratio of the soil matching value of the G channel to the surface matching value of the G channel as the G channel comparison value, replace the channel value of the original pixel with the G channel comparison value, and obtain the G channel soil and surface comparison map. S15. Take the ratio of the soil matching value of channel B to the surface matching value of channel B as the contrast value of channel B. Replace the channel value of the original pixel with the contrast value of channel B to obtain the soil and surface contrast map of channel B.

3. The landslide risk prediction method based on a landslide disaster assessment model according to claim 1, characterized in that, S2 includes the following steps: S21. For the same channel, in the channel soil surface comparison map at adjacent time points, the channel soil surface comparison map at the later time point is subtracted from the channel soil surface comparison map at the earlier time point to obtain the soil surface comparison difference value of each pixel. S22. At the same pixel, the ratio of the soil surface contrast difference to the soil surface contrast value at the previous time step is taken as the soil surface feature value of that pixel. S23. Replace the soil surface feature values ​​with the channel values ​​of the original pixels to obtain the soil surface feature map.

4. The landslide risk prediction method based on a landslide hazard assessment model according to claim 1, characterized in that, S3 includes the following steps: S31. Take the average value of the soil surface feature values ​​on the soil surface feature map at each time point to obtain the soil surface feature threshold. S32. Mark pixels whose soil surface feature values ​​are greater than the soil surface feature threshold as suspected variable pixels. S33. Obtain the landslide weight coefficient for suspected variable pixels, set the landslide weight coefficient of non-suspected variable pixels to 0, arrange each landslide weight coefficient according to the corresponding pixel, and obtain the landslide weight coefficient matrix at each time. S34. Add the landslide weight coefficient matrices of multiple consecutive time points to obtain the cumulative landslide weight matrix.

5. The landslide risk prediction method based on a landslide disaster assessment model according to claim 4, characterized in that, The process of obtaining the landslide weight coefficient in S33 is as follows: count the number of suspected changed pixels within the 5×5 neighborhood of the suspected changed pixels, and take the ratio of the number of suspected changed pixels in the neighborhood to the total number of pixels in the 5×5 neighborhood as the landslide weight coefficient.

6. The landslide risk prediction method based on a landslide disaster assessment model according to claim 4, characterized in that, S4 includes the following sub-steps: S41. When the soil surface feature value of a pixel in the soil surface feature map is less than the soil surface feature threshold, set the soil surface feature value of that pixel to 0, and retain other soil surface feature values ​​to obtain a corrected soil surface feature map. S42. Add the corrected soil surface feature maps from multiple consecutive time points to obtain the cumulative soil surface feature map.

Images