Debris flow risk assessment method and assessment system thereof

Abstract

The application discloses a kind of debris flow risk assessment method and its evaluation system, belong to image processing technical field.The application first collects monitoring area after rain multispectral image, extracts the wet loose system number and moisture rock and dirt coefficient of each pixel point;In the same pixel point, take the wet loose body contrast ratio and moisture rock and dirt contrast ratio of current and rain starting time, obtain suspected debris flow pixel point;Statistical change time and current time multispectral image in each suspected debris flow pixel point row, column of continuous suspected point quantity, construct flowability weight matrix;The above two kinds of contrast ratios of each pixel point at current time are respectively formed corresponding contrast image;Two kinds of images are handled using risk assessment neural network, based on current flowability weight matrix to feature exert weight, obtain debris flow risk value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a debris flow risk assessment method and an assessment system thereof. BACKGROUND

[0002] Traditional debris flow risk assessment methods mostly rely on geological survey data, historical disaster records and empirical formulas, and have limitations such as long evaluation period, poor timeliness and low spatial resolution. For example, based on the determination of soil mechanical parameters by field sampling, it is difficult to quickly reflect the dynamic changes of surface materials during rainfall process; and the warning model relying on hydrological station monitoring data is also difficult to accurately capture the micro-landform and material composition differences in small-scale areas, resulting in a large deviation between the evaluation results and the actual disaster occurrence.

[0003] With the rapid development of remote sensing technology and artificial intelligence, multispectral remote sensing provides a new technical path for real-time monitoring of surface wetness and loose material movement due to its advantages of synchronously obtaining surface material spectral characteristics and spatial distribution information. However, existing disaster assessment methods based on multispectral images often only use a single spectral index to determine the surface wetness, ignoring the influence of loose body structure changes and rock-water interaction on debris flow, and lacking dynamic quantification of material flow, resulting in low accuracy of risk assessment. SUMMARY

[0004] In view of the above deficiencies in the prior art, the present application provides a debris flow risk assessment method and an assessment system thereof to solve the problem of low accuracy of debris flow risk assessment in the prior art.

[0005] In order to achieve the above-mentioned application purposes, the technical scheme adopted by the present application is as follows: a debris flow risk assessment method, comprising the following steps:

[0006] Collecting multispectral images of a monitoring area after rainfall, and extracting the wet loose body coefficient and the water-rock coefficient of each pixel point;

[0007] At the same pixel point, taking the wet loose body contrast value and the water-rock contrast value at the current time and the starting time of rainfall to obtain a suspected debris flow pixel point;

[0008] Counting the number of continuous suspected debris flow pixel points in the row and column of each suspected debris flow pixel point in the multispectral images at the variable time and the current time, and constructing a flowability weight matrix;

[0009] The wet loose body contrast value of each pixel point at the current time is used to form a wet loose body contrast image, and the water-rock contrast value of each pixel point at the current time is used to form a water-rock contrast image;

[0010] The wet loose body contrast image and the moisture debris contrast image at the current time are processed by using a risk assessment neural network, and a weight of a feature is exerted based on a liquidity weight matrix at the current time, so as to obtain a debris flow risk value.

[0011] Further, the process of obtaining the wet loose system number comprises: subtracting the vegetation index NDVI of each pixel point from 1 to obtain a non-vegetation index NDVI, adding the MNDWI water body index of each pixel point to 1 to obtain an enhanced MNDWI water body index, multiplying the enhanced MNDWI water body index and the non-vegetation index NDVI of the same pixel point, and normalizing the multiplication result to obtain the wet loose system number of each pixel point.

[0012] Further, the process of obtaining the moisture debris coefficient comprises: subtracting the near-infrared band NIR reflectance of each pixel point from the short-wave infrared 2 band SWIR2 reflectance of each pixel point, and normalizing the subtraction result to obtain a debris factor; subtracting the red band Red reflectance from the green band Green reflectance of each pixel point, and normalizing the subtraction result to obtain a soil moisture factor, and multiplying the soil moisture factor and the debris factor to obtain the moisture debris coefficient.

[0013] Further, the process of obtaining the suspected debris flow pixel point comprises:

[0014] At the same pixel point, the ratio of the wet loose system number of the pixel point at the current time to the wet loose system number of the pixel point at the rain starting time is taken as the wet loose body contrast value;

[0015] At the same pixel point, the ratio of the moisture debris coefficient of the pixel point at the current time to the moisture debris coefficient of the pixel point at the rain starting time is taken as the moisture debris contrast value;

[0016] The wet loose body contrast values of the pixel points are averaged to obtain a wet loose body contrast average value;

[0017] The moisture debris contrast values of the pixel points are averaged to obtain a moisture debris contrast average value;

[0018] The pixel points with the wet loose body contrast value greater than the wet loose body contrast average value and the moisture debris contrast value greater than the moisture debris contrast average value are marked as suspected debris flow pixel points.

[0019] Further, the process of constructing the liquidity weight matrix comprises:

[0020] The time when each pixel point is marked as a suspected debris flow pixel point is marked as a variable time;

[0021] Taking each suspected debris flow pixel in the multispectral image at the time of change and the current time as the center, count the number of consecutive suspected debris flow pixels in the row to obtain the row suspected number, and count the number of consecutive suspected debris flow pixels in the column to obtain the column suspected number.

[0022] Based on the number of suspected rows and columns of suspected debris flow pixels at the time of change and the current time, obtain the mobility weight of the suspected debris flow pixel.

[0023] Reset the liquidity weights of other pixels to 0, and construct a liquidity weight matrix based on the liquidity weights of each suspected debris flow pixel.

[0024] Furthermore, the process of obtaining liquidity weights includes:

[0025] The lateral flow weight is obtained based on the difference in the number of suspected debris flow pixels at the current time and at the time of change.

[0026] The vertical flow weight is obtained based on the difference in the number of suspected debris flow pixels in the column between the current time and the time of change.

[0027] The average of the lateral and longitudinal flow weights for the same pixel is used to obtain the flow weight of the suspected debris flow pixel.

[0028] Furthermore, the process of obtaining the lateral flow weight includes: at the same suspected debris flow pixel location, subtracting the suspected row number at the time of change from the suspected row number at the current time to obtain the row suspected number difference; when the row suspected number difference is less than 0, assigning the lateral flow weight to 0; when the row suspected number difference is greater than 0, normalizing the row suspected number difference to obtain the lateral flow weight.

[0029] The process of obtaining the longitudinal flow weight includes: at the same suspected debris flow pixel location, subtracting the suspected column number at the time of change from the suspected column number at the current time to obtain the column suspected number difference. When the column suspected number difference is less than 0, the longitudinal flow weight is assigned a value of 0. When the column suspected number difference is greater than 0, the column suspected number difference is normalized to obtain the longitudinal flow weight.

[0030] Furthermore, the risk assessment neural network includes: a first stacked convolutional unit, a second stacked convolutional unit, a multiplier M1, a multiplier M2, a first feature extraction unit, a second feature extraction unit, an adder A1, and a fully connected layer;

[0031] The input of the first stacked convolutional unit is used to input a contrast image of moist loose material; the input of the second stacked convolutional unit is used to input a contrast image of water-bearing rock debris.

[0032] The first input of multiplier M1 is used to input the fluidity weights, its second input is connected to the output of the first stacked convolutional unit, and its output is connected to the input of the first feature extraction unit; the first input of multiplier M2 is used to input the fluidity weights, its second input is connected to the output of the second stacked convolutional unit, and its output is connected to the input of the second feature extraction unit.

[0033] The input of adder A1 is connected to the output of the first feature extraction unit and the output of the second feature extraction unit, and its output is connected to the input of the fully connected layer; the output of the fully connected layer serves as the output of the risk assessment neural network.

[0034] Furthermore, both the first feature extraction unit and the second feature extraction unit include: a first convolutional layer, a max pooling layer, an average pooling layer, an adder A2, a second convolutional layer, a sigmoid layer, and a multiplier M3;

[0035] The input of the first convolutional layer serves as the input of the first feature extraction unit and the second feature extraction unit, and its output is connected to the input of the max pooling layer and the input of the average pooling layer, respectively.

[0036] The input of adder A2 is connected to the output of max pooling layer and average pooling layer, respectively, and its output is connected to the first input of multiplier M3 and the input of the second convolutional layer, respectively. The output of the second convolutional layer is connected to the input of the sigmoid layer. The second input of multiplier M3 is connected to the output of the sigmoid layer, and its output serves as the output of the first feature extraction unit and the second feature extraction unit.

[0037] A debris flow risk assessment system includes: a coefficient extraction subsystem, a suspected debris flow pixel acquisition subsystem, a mobility weight matrix construction subsystem, an image construction subsystem, and a classification subsystem;

[0038] The coefficient extraction subsystem is used to collect multispectral images of the monitoring area after rain and extract the wet and loose system number and moisture rock debris coefficient for each pixel.

[0039] The suspected debris flow pixel acquisition subsystem is used to obtain suspected debris flow pixels by taking the comparison value of wet loose body and water-rock debris at the current time and the start time of the rain at the same pixel.

[0040] The liquidity weight matrix construction subsystem is used to count the number of consecutive suspected debris flow pixels in the row and column of each suspected debris flow pixel in the multispectral image at the time of change and the current time, and to construct the liquidity weight matrix.

[0041] The image construction subsystem is used to construct a wet loose body comparison image by comparing the wet loose body comparison values ​​of each pixel at the current time, and to construct a water-rock debris comparison image by comparing the water-rock debris comparison values ​​of each pixel at the current time.

[0042] The classification subsystem is used to process the contrast images of moist loose material and water-bearing debris using a risk assessment neural network. Based on the weights applied to the features by the mobility weight matrix, the debris flow risk value is obtained.

[0043] The beneficial effects of this invention are as follows:

[0044] 1. This invention introduces both the "wet loose system number" and the "moisture-rock debris coefficient" to comprehensively consider the impact of loose body structure changes and rock debris-water interaction on debris flows, thereby improving the accuracy of debris flow risk assessment.

[0045] 2. This invention accurately locates suspected debris flow pixels by comparing the coefficient ratio between the current time and the time when the rain started, reducing the evaluation deviation caused by misjudgment of data at a single moment.

[0046] 3. This invention constructs a flow weight matrix by statistically analyzing the number of consecutive rows and columns of suspected debris flow pixels in multispectral images at different times and the current time. This matrix intuitively reflects the diffusion of loose material and provides feature weights for the risk assessment neural network, enabling the network to more accurately capture the development state of debris flows.

[0047] 4. This invention uses a risk assessment neural network to process comparative images of moist loose material with significant moist loose material characteristics and comparative images of water and rock debris with significant water and rock debris characteristics. Then, based on the weights applied to the features by the fluidity weight matrix, the accuracy of debris flow risk assessment is improved. Attached Figure Description

[0048] Figure 1 A flowchart of a debris flow risk assessment method;

[0049] Figure 2 This is a schematic diagram of the structure of a risk assessment neural network;

[0050] Figure 3 This is a schematic diagram of the structure of the first feature extraction unit and the second feature extraction unit. Detailed Implementation

[0051] 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. Example

[0052] like Figure 1 As shown, a debris flow risk assessment method includes the following steps:

[0053] Collect multispectral images of the monitoring area after rainfall, and extract the wet and loose system number and moisture-rock coefficient for each pixel;

[0054] At the same pixel, take the comparison value of the wet loose body and the comparison value of the water and rock debris at the current time and the starting time of the rain to obtain the suspected debris flow pixel.

[0055] The number of consecutive suspected debris flow pixels in the row and column of each suspected debris flow pixel in the multispectral image at the time of change and the current time is counted to construct a mobility weight matrix;

[0056] The wet loose body comparison value of each pixel at the current time is used to form a wet loose body comparison image, and the water and rock debris comparison value of each pixel at the current time is used to form a water and rock debris comparison image.

[0057] A risk assessment neural network is used to process the current moment's contrast images of moist loose material and water-bearing debris. Based on the current moment's fluidity weight matrix, the weights applied to the features are used to obtain the debris flow risk value.

[0058] Replace the original pixels of the pixels with the contrast values ​​of the moist loose material to obtain the contrast image of the moist loose material, and replace the original pixels of the pixels with the contrast values ​​of the water-bearing rock fragments to obtain the contrast image of the water-bearing rock fragments.

[0059] In this embodiment, the process of obtaining the number of moist and loose systems includes: subtracting the vegetation index NDVI of each pixel from 1 to obtain the non-vegetation index NDVI; adding the MNDWI water index of each pixel to 1 to obtain the enhanced MNDWI water index; multiplying the enhanced MNDWI water index of the same pixel with the non-vegetation index NDVI; and normalizing the multiplication result to obtain the number of moist and loose systems for each pixel.

[0060] In this embodiment, the formula for calculating the wetted loose system number is: , where θ ML,i MNDWI is the wet / loose system number of the i-th pixel. i NDVI is the MNDWI (Modified Normalized Difference Water Index) of the i-th pixel. i NDVI (Normalized Difference Vegetation Index) is the vegetation index of the i-th pixel.

[0061] The water body index (MNDWI) is sensitive to shallow surface water (such as pore water and surface water in debris flows), and its value ranges from [-1, 1]. Adding 1 maps the range to [0, 2] to avoid the cancellation of negative values ​​in the product. The vegetation index (NDVI) reflects vegetation cover. In debris flow areas, due to vegetation destruction, the NDVI value is low (close to 0 or negative). 1-NDVI can highlight loose areas without vegetation cover (value range [0, 2]).

[0062] After multiplying the two and dividing by 4, the result is normalized to [0,1]. Areas with high values ​​(such as the number of moist and loose materials > 0.6) simultaneously satisfy the characteristics of "high moisture + low vegetation (i.e., high loose material)," which is highly consistent with debris flow prone areas and can effectively suppress the interference of densely vegetated areas and dry bare land.

[0063] In this embodiment, the process of obtaining the water-rock debris coefficient includes: subtracting the near-infrared (NIR) reflectance of each pixel from the short-wave infrared (SWIR) 2-band reflectance, and normalizing the subtraction result to obtain the rock debris factor; subtracting the red (Green) reflectance of each pixel from the green (Red) band reflectance, and normalizing the subtraction result to obtain the soil moisture factor; and multiplying the soil moisture factor by the rock debris factor to obtain the water-rock debris coefficient.

[0064] The formula for calculating the moisture content coefficient of rock cuttings is: , where θ MD,i ρ is the moisture and rock debris coefficient of the i-th pixel. SWIR2,i Let ρ be the reflectance of the shortwave infrared 2 (SWIR2) band of the i-th pixel. NIR,i Let ρ be the near-infrared (NIR) reflectance of the i-th pixel. Red,i Let ρ be the red band reflectance of the i-th pixel. Green,i Let be the green reflectance of the i-th pixel.

[0065] Shortwave infrared band 2 (SWIR2, 2.0-2.5μm) is sensitive to the reflection of rocks / coarse-grained minerals, while near-infrared band (NIR) is sensitive to the reflection of vegetation. The normalized difference between the two can highlight the area of ​​"high coarse-grained rock debris + low vegetation" (more coarse-grained rocks result in higher SWIR2 reflection).

[0066] The red band is sensitive to soil moisture absorption (higher moisture content results in lower reflectance), while the green band is sensitive to moisture reflection (higher moisture content results in higher reflectance). The product of these two bands can enhance the coupling characteristics of "high coarse particles (material source) + high moisture (water source)". Areas with high values ​​(moisture-rock debris coefficient > 0.4) are highly matched with debris flow areas (coarse particle deposition + surface runoff) and can suppress interference from pure vegetation areas and pure water bodies.

[0067] In this invention, the moist loose system index uses a combination of MNDWI (sensitive to shallow surface moisture) and NDVI (reflecting vegetation cover) to accurately pinpoint areas that are "sufficiently moist, severely damaged by vegetation, and with exposed loose material," which are the conditions for debris flows.

[0068] In this invention, the moisture-rock fragment coefficient highlights the rock fragment distribution through the combination of SWIR2 (sensitive to coarse-grained minerals) and NIR (sensitive to vegetation), and highlights the moisture content through the combination of Green (sensitive to water reflection) and Red (sensitive to water absorption), accurately identifying areas where "coarse particles are carried by water" (such as debris flow faucets and floodplain deposits), with a greater emphasis on areas where debris flows have already occurred.

[0069] In this embodiment, the process of acquiring suspected debris flow pixels for each image includes:

[0070] At the same pixel, the ratio of the number of wet and loose systems at the current pixel to the number of wet and loose systems at the start of the rain is taken as the wet and loose body comparison value.

[0071] At the same pixel, the ratio of the moisture and rock debris coefficient of the pixel at the current time to the moisture and rock debris coefficient of the pixel at the start of the rain is taken as the moisture and rock debris comparison value.

[0072] In each multispectral image A, the average value of the wet loose matter contrast at each pixel is taken to obtain the average value of the wet loose matter contrast.

[0073] In each multispectral image A, the average value of the water-rock-chip contrast at each pixel is taken to obtain the average water-rock-chip contrast value.

[0074] In the same image A, pixels with a contrast value of moist loose mass greater than the mean contrast value of moist loose mass and a contrast value of water and rock debris greater than the mean contrast value of water and rock debris are suspected debris flow pixels.

[0075] This invention extracts a moist loose mass comparison value to obtain the increase factor of the moist loose mass index at a given pixel location after rainfall, and extracts a moisture-rock debris comparison value to obtain the increase factor of the moisture-rock debris coefficient at the same pixel location after rainfall, highlighting the "changes in the wetting of loose material" and "changes in water-sand mixing" during rainfall. This invention filters out pixels whose moist loose mass comparison value is greater than the average moist loose mass comparison value, and whose moisture-rock debris comparison value is also greater than the average moisture-rock debris comparison value, thus simultaneously obtaining pixels that meet both conditions and improving the accuracy of filtering suspected debris flow pixels.

[0076] In this embodiment, the process of constructing the liquidity weight matrix includes:

[0077] The moment when each pixel is marked as a suspected debris flow pixel is the change moment;

[0078] Taking each suspected debris flow pixel in the multispectral image at the time of change and the current time as the center, count the number of consecutive suspected debris flow pixels in the row (i.e. count the number of consecutive suspected debris flow pixels on the left and right sides starting from the center), and get the row suspected number. Count the number of consecutive suspected debris flow pixels in the column (i.e. count the number of consecutive suspected debris flow pixels on the top and bottom sides starting from the center), and get the column suspected number.

[0079] Based on the number of suspected rows and columns of suspected debris flow pixels at the time of change and the current time, obtain the mobility weight of the suspected debris flow pixel.

[0080] Reset the liquidity weights of other pixels to 0, and construct a liquidity weight matrix based on the liquidity weights of each suspected debris flow pixel.

[0081] The mobility weight matrix is ​​the same size as the multispectral image. The mobility weight of each pixel is replaced with the original pixel of that pixel to obtain the mobility weight matrix.

[0082] In this embodiment, the process of obtaining liquidity weights includes:

[0083] The lateral flow weight is obtained based on the difference in the number of suspected debris flow pixels at the current time and at the time of change.

[0084] The vertical flow weight is obtained based on the difference in the number of suspected debris flow pixels in the column between the current time and the time of change.

[0085] The average of the lateral and longitudinal flow weights for the same pixel is used to obtain the flow weight of the suspected debris flow pixel.

[0086] In this embodiment, the process of obtaining the lateral flow weight includes: at the same suspected debris flow pixel position, subtracting the suspected row number at the time of change from the suspected row number at the current time to obtain the row suspected number difference; when the row suspected number difference is less than 0, assigning the lateral flow weight to 0; when the row suspected number difference is greater than 0, normalizing the row suspected number difference to obtain the lateral flow weight.

[0087] The process of obtaining the longitudinal flow weight includes: at the same suspected debris flow pixel location, subtracting the suspected column number at the time of change from the suspected column number at the current time to obtain the column suspected number difference. When the column suspected number difference is less than 0, the longitudinal flow weight is assigned a value of 0. When the column suspected number difference is greater than 0, the column suspected number difference is normalized to obtain the longitudinal flow weight.

[0088] This invention compares the difference in the number of consecutive suspected pixels in rows / columns at a time of change with the current time, and uses horizontal and vertical flow weights to characterize the changes in the diffusion range of debris flows in the horizontal and vertical directions, respectively. When the difference in the number of suspected pixels in rows / columns is positive, the normalized weight value intuitively reflects the degree of expansion of the flow range. When the difference in the number of suspected pixels in rows / columns is less than 0, the weight is assigned to 0, which effectively excludes areas of "flow range contraction", focuses on the truly expanding flow area, and avoids misjudging stable accumulation areas as active flow areas.

[0089] The formula for normalizing the difference in the number of suspected rows is: , where γ L,i N represents the lateral flow weight of the i-th pixel. L,t,i N represents the number of rows of suspected debris flow pixels at the i-th pixel at the current time. L,τ,i Let t be the number of rows of suspected debris flow pixels at the i-th pixel at the time of change, τ be the current time, M be the time of change, and M be the normalization constant.

[0090] The formula for normalizing the difference in the number of suspected cases is: , where γ A,i N represents the vertical flow weight of the i-th pixel. A,t,i N represents the column count of suspected debris flow pixels at the i-th pixel at the current time. A,τ,i The column count of suspected debris flow pixels at the i-th pixel at the time of change.

[0091] In this embodiment, when collecting multispectral images of the monitoring area after rain, the shooting angle and shooting height are kept consistent, that is, the camera parameters for capturing multiple images remain unchanged.

[0092] like Figure 2As shown, the risk assessment neural network includes: a first stacked convolutional unit, a second stacked convolutional unit, a multiplier M1, a multiplier M2, a first feature extraction unit, a second feature extraction unit, an adder A1, and a fully connected layer;

[0093] The input of the first stacked convolutional unit is used to input a contrast image of moist loose material; the input of the second stacked convolutional unit is used to input a contrast image of water-bearing rock debris.

[0094] The first input of multiplier M1 is used to input the fluidity weights, its second input is connected to the output of the first stacked convolutional unit, and its output is connected to the input of the first feature extraction unit; the first input of multiplier M2 is used to input the fluidity weights, its second input is connected to the output of the second stacked convolutional unit, and its output is connected to the input of the second feature extraction unit.

[0095] The input of adder A1 is connected to the output of the first feature extraction unit and the output of the second feature extraction unit, and its output is connected to the input of the fully connected layer; the output of the fully connected layer serves as the output of the risk assessment neural network.

[0096] This invention processes the image of wet loose material (focusing on the characteristics of loose material after rainfall) through a first stacked convolutional unit, and processes the image of water and rock debris (focusing on the characteristics of water and sand mixing and flow) through a second stacked convolutional unit. Then, a flow weight matrix is ​​introduced through multipliers M1 and M2 to weight the feature maps of wet loose material and water and rock debris respectively, so that the risk assessment neural network focuses on key features. The first feature extraction unit and the second feature extraction unit extract features respectively. At the adder A1, the two feature paths are fused and the debris flow risk value is output through a fully connected layer.

[0097] like Figure 3 As shown, both the first feature extraction unit and the second feature extraction unit include: a first convolutional layer, a max pooling layer, an average pooling layer, an adder A2, a second convolutional layer, a sigmoid layer, and a multiplier M3;

[0098] The input of the first convolutional layer serves as the input of the first feature extraction unit and the second feature extraction unit, and its output is connected to the input of the max pooling layer and the input of the average pooling layer, respectively.

[0099] The input of adder A2 is connected to the output of max pooling layer and average pooling layer, respectively, and its output is connected to the first input of multiplier M3 and the input of the second convolutional layer, respectively. The output of the second convolutional layer is connected to the input of the sigmoid layer. The second input of multiplier M3 is connected to the output of the sigmoid layer, and its output serves as the output of the first feature extraction unit and the second feature extraction unit.

[0100] The outputs of the first convolutional layer are fed into a max pooling layer and an average pooling layer to extract salient and global features, respectively. After being fused by adder A2, multi-scale feature complementarity is achieved. The second convolutional layer, together with the sigmoid layer and multiplier M3, constitutes an attention mechanism. The second convolutional layer further mines features, and the sigmoid layer outputs weights between 0 and 1. These weights are then applied to the fused features by multiplier M3, which automatically highlights features strongly correlated with debris flow risk, suppresses irrelevant background noise, and increases the proportion of key information in risk assessment.

[0101] In this embodiment, the first stacked convolutional unit and the second stacked convolutional unit include a third convolutional layer, a fourth convolutional layer and a fifth convolutional layer connected in sequence. The kernel size of the third convolutional layer is 1×1, the padding is 0, and the stride is set to 1. The kernel size of the fourth convolutional layer and the fifth convolutional layer are both 3×3, the padding is 1, and the stride is 1.

[0102] The output feature maps of the first and second stacked convolutional units have the same size as the input image.

[0103] The kernel size of the first convolutional layer is 5×5, the padding is 0, and the stride is 1; the kernel size of the second convolutional layer is 1×1, the padding is 0, and the stride is 1.

[0104] Example 2:

[0105] A debris flow risk assessment system includes: a coefficient extraction subsystem, a suspected debris flow pixel acquisition subsystem, a mobility weight matrix construction subsystem, an image construction subsystem, and a classification subsystem;

[0106] The coefficient extraction subsystem is used to collect multispectral images of the monitoring area after rain and extract the wet and loose system number and moisture rock debris coefficient for each pixel.

[0107] The suspected debris flow pixel acquisition subsystem is used to obtain suspected debris flow pixels by taking the comparison value of wet loose body and water-rock debris at the current time and the start time of the rain at the same pixel.

[0108] The liquidity weight matrix construction subsystem is used to count the number of consecutive suspected debris flow pixels in the row and column of each suspected debris flow pixel in the multispectral image at the time of change and the current time, and to construct the liquidity weight matrix.

[0109] The image construction subsystem is used to construct a wet loose body comparison image by comparing the wet loose body comparison values ​​of each pixel at the current time, and to construct a water-rock debris comparison image by comparing the water-rock debris comparison values ​​of each pixel at the current time.

[0110] The classification subsystem is used to process the contrast images of moist loose material and water-bearing debris using a risk assessment neural network. Based on the weights applied to the features by the mobility weight matrix, the debris flow risk value is obtained.

[0111] The specific implementation process of Example 1 and Example 2 is the same.

[0112] This invention introduces both the "wet loose system number" and the "moisture-rock debris coefficient" to comprehensively consider the impact of loose body structure changes and rock debris-water interaction on debris flows, thereby improving the accuracy of debris flow risk assessment.

[0113] This invention accurately identifies suspected debris flow pixels by comparing the coefficient ratio between the current moment and the moment when the rain started, reducing evaluation bias caused by misjudgment of data from a single moment.

[0114] This invention constructs a flow weight matrix by statistically analyzing the number of consecutive rows and columns of suspected debris flow pixels in multispectral images at different times and the current time. This matrix intuitively reflects the diffusion of loose material and provides feature weights for the risk assessment neural network, enabling the network to more accurately capture the development status of debris flows.

[0115] This invention employs a risk assessment neural network to process comparative images of moist loose bodies with significant moist features and comparative images of water and rock debris with significant water and rock debris features. Then, based on the fluidity weight matrix, weights are applied to the features, thereby improving the accuracy of debris flow risk assessment.

[0116] 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 debris flow risk assessment method, characterized in that, Includes the following steps: Collect multispectral images of the monitoring area after rainfall, and extract the wet and loose system number and moisture-rock coefficient for each pixel; At the same pixel, take the comparison value of the wet loose body and the comparison value of the water and rock debris at the current time and the starting time of the rain to obtain the suspected debris flow pixel. The moment when each pixel is marked as a suspected debris flow pixel is the change moment. The number of consecutive suspected debris flow pixels in the row and column of each suspected debris flow pixel in the multispectral image at the change moment and the current moment is counted to construct a mobility weight matrix. The wet loose body comparison value of each pixel at the current time is used to form a wet loose body comparison image, and the water and rock debris comparison value of each pixel at the current time is used to form a water and rock debris comparison image. A risk assessment neural network is used to process the current moment's contrast images of moist loose material and water-bearing debris. Based on the current moment's fluidity weight matrix, the weights applied to the features are used to obtain the debris flow risk value.

2. The debris flow risk assessment method according to claim 1, characterized in that, The process of obtaining the number of moist loose systems includes: subtracting the vegetation index NDVI of each pixel from 1 to obtain the non-vegetation index NDVI; adding the MNDWI water index of each pixel to 1 to obtain the enhanced MNDWI water index; multiplying the enhanced MNDWI water index of the same pixel with the non-vegetation index NDVI; and normalizing the multiplication result to obtain the number of moist loose systems for each pixel.

3. The debris flow risk assessment method according to claim 1, characterized in that, The process of obtaining the water-rock debris coefficient includes: subtracting the near-infrared (NIR) reflectance of each pixel from the short-wave infrared (SWIR) 2-band reflectance, and normalizing the subtraction result to obtain the rock debris factor; subtracting the red (Green) reflectance of each pixel from the green (Red) band reflectance, and normalizing the subtraction result to obtain the soil moisture factor; and multiplying the soil moisture factor by the rock debris factor to obtain the water-rock debris coefficient.

4. The debris flow risk assessment method according to claim 1, characterized in that, The process of obtaining suspected debris flow pixels includes: At the same pixel, the ratio of the number of wet and loose systems at the current pixel to the number of wet and loose systems at the start of the rain is taken as the wet and loose body comparison value. At the same pixel, the ratio of the moisture and rock debris coefficient of the pixel at the current time to the moisture and rock debris coefficient of the pixel at the start of the rain is taken as the moisture and rock debris comparison value. The average of the wet loose body contrast values ​​of each pixel is taken to obtain the average wet loose body contrast value; The average value of the water content and rock debris contrast for each pixel is taken to obtain the average water content and rock debris contrast. Pixels with a contrast value of moist loose mass greater than the mean contrast value of moist loose mass and a contrast value of water-rock debris greater than the mean contrast value of water-rock debris are suspected debris flow pixels.

5. The debris flow risk assessment method according to claim 1, characterized in that, The process of constructing the liquidity weight matrix includes: Taking each suspected debris flow pixel in the multispectral image at the time of change and the current time as the center, count the number of consecutive suspected debris flow pixels in the row to obtain the row suspected number, and count the number of consecutive suspected debris flow pixels in the column to obtain the column suspected number. Based on the number of suspected rows and columns of suspected debris flow pixels at the time of change and the current time, obtain the mobility weight of the suspected debris flow pixel. Reset the liquidity weights of other pixels to 0, and construct a liquidity weight matrix based on the liquidity weights of each suspected debris flow pixel.

6. The debris flow risk assessment method according to claim 5, characterized in that, The process of obtaining liquidity weights includes: The lateral flow weight is obtained based on the difference in the number of suspected debris flow pixels at the current time and at the time of change. The vertical flow weight is obtained based on the difference in the number of suspected debris flow pixels in the column between the current time and the time of change. The average of the lateral and longitudinal flow weights for the same pixel is used to obtain the flow weight of the suspected debris flow pixel.

7. The debris flow risk assessment method according to claim 6, characterized in that, The process of obtaining the lateral flow weight includes: at the same suspected debris flow pixel position, subtract the suspected row number at the current time from the suspected row number at the time of change to obtain the row suspected number difference. When the row suspected number difference is less than 0, the lateral flow weight is assigned a value of 0. When the row suspected number difference is greater than 0, the row suspected number difference is normalized to obtain the lateral flow weight. The process of obtaining the longitudinal flow weight includes: at the same suspected debris flow pixel location, subtracting the suspected column number at the time of change from the suspected column number at the current time to obtain the column suspected number difference. When the column suspected number difference is less than 0, the longitudinal flow weight is assigned a value of 0. When the column suspected number difference is greater than 0, the column suspected number difference is normalized to obtain the longitudinal flow weight.

8. The debris flow risk assessment method according to claim 1, characterized in that, The risk assessment neural network includes: a first stacked convolutional unit, a second stacked convolutional unit, multipliers M1 and M2, a first feature extraction unit, a second feature extraction unit, an adder A1, and a fully connected layer; The input of the first stacked convolutional unit is used to input a contrast image of moist loose material; the input of the second stacked convolutional unit is used to input a contrast image of water-bearing rock debris. The first input of multiplier M1 is used to input the fluidity weights, its second input is connected to the output of the first stacked convolutional unit, and its output is connected to the input of the first feature extraction unit; the first input of multiplier M2 is used to input the fluidity weights, its second input is connected to the output of the second stacked convolutional unit, and its output is connected to the input of the second feature extraction unit. The input of adder A1 is connected to the output of the first feature extraction unit and the output of the second feature extraction unit, and its output is connected to the input of the fully connected layer; the output of the fully connected layer serves as the output of the risk assessment neural network.

9. The debris flow risk assessment method according to claim 8, characterized in that, Both the first feature extraction unit and the second feature extraction unit include: a first convolutional layer, a max pooling layer, an average pooling layer, an adder A2, a second convolutional layer, a sigmoid layer, and a multiplier M3; The input of the first convolutional layer serves as the input of the first feature extraction unit and the second feature extraction unit, and its output is connected to the input of the max pooling layer and the input of the average pooling layer, respectively. The input of adder A2 is connected to the output of max pooling layer and average pooling layer, respectively, and its output is connected to the first input of multiplier M3 and the input of the second convolutional layer, respectively. The output of the second convolutional layer is connected to the input of the sigmoid layer. The second input of multiplier M3 is connected to the output of the sigmoid layer, and its output serves as the output of the first feature extraction unit and the second feature extraction unit.

10. A debris flow risk assessment system, implemented based on the debris flow risk assessment method according to any one of claims 1 to 9, characterized in that, include: The system includes a coefficient extraction subsystem, a suspected debris flow pixel acquisition subsystem, a mobility weight matrix construction subsystem, an image construction subsystem, and a classification subsystem. The coefficient extraction subsystem is used to collect multispectral images of the monitoring area after rain and extract the wet and loose system number and moisture rock debris coefficient for each pixel. The suspected debris flow pixel acquisition subsystem is used to obtain suspected debris flow pixels by taking the comparison value of wet loose body and water-rock debris at the current time and the start time of the rain at the same pixel. The moment when each pixel is marked as a suspected debris flow pixel is the change moment. The mobility weight matrix construction subsystem is used to count the number of consecutive suspected debris flow pixels in the row and column of each suspected debris flow pixel in the multispectral image at the change moment and the current moment, and construct the mobility weight matrix. The image construction subsystem is used to construct a wet loose body comparison image by comparing the wet loose body comparison values ​​of each pixel at the current time, and to construct a water-rock debris comparison image by comparing the water-rock debris comparison values ​​of each pixel at the current time. The classification subsystem is used to process the contrast images of moist loose material and water-bearing debris using a risk assessment neural network. Based on the weights applied to the features by the mobility weight matrix, the debris flow risk value is obtained.

Images