A dam surface particle component area intelligent identification method

Abstract

The application discloses a kind of dam body surface particle group component area intelligent identification method, comprising the following steps: S1, collection dam body surface image obtains dam body surface image training set, it includes first- third training subset;S2, first- second training subset is input YOLO5 network, obtains dam body surface image parameter;S3, according to dam body surface image parameter setting YOLO5 network, the outline identification of particle is carried out to the third training subset of input by YOLO5 network, obtains the gradation curve of dam body surface image, completes dam body surface particle group component area identification.The application establishes YOLO5 network when using multi-scale and global context information, designs feature pyramid module and boundary optimization module, to meet the precision requirement of engineering practical application to the extraction identification of dam body material structure, and simplifies the process of dam body surface particle identification, improves the speed and efficiency of dam body surface particle grouping.

Description

Technical Field

[0001] This invention belongs to the field of landslide dam management technology, specifically relating to an intelligent identification method for particle zoning on the surface of a landslide dam. Background Technology

[0002] The particle size and gradation of soil and rock materials are the most important indicators for in-depth analysis of landslide dams. However, traditionally, this analysis method mainly involves laboratory screening and experiments, requiring long-distance material transportation and lengthy experiments. For landslide dam rescue operations, the lack of on-site equipment, transportation difficulties, and time-consuming processes all restrict the application of experimental techniques to material identification in landslide dam rescue. In recent years, deep learning technology has made significant progress, greatly improving the segmentation accuracy and recognition rate of images. Therefore, this project attempts to use an improved YOLO5 deep learning model, adding dilated convolution and improving the loss function model to expand the recognition field of small targets in images, thereby improving recognition accuracy. This model can provide rapid on-site material photographs in landslide dam rescue operations. Then, using pre-trained neural network parameters, it can quickly achieve surface material partitioning and material particle gradation calculation, ultimately obtaining accurate gradation curves. This serves as an important method in landslide dam rescue and disaster relief, rapidly obtaining the material properties of the landslide dam material. Summary of the Invention

[0003] To address the aforementioned shortcomings in existing technologies, this invention provides an intelligent identification method for particle grouping on the surface of landslide dams, which solves the problems of complexity and slow speed in existing methods for particle grouping and identification on the surface of landslide dams.

[0004] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for intelligent identification of particle zoning on the surface of a landslide dam, comprising the following steps:

[0005] S1. Collect images of the surface of the landslide dam to obtain a training set of images of the surface of the landslide dam, which includes the first to third training subsets;

[0006] S2. Input the first and second training subsets into the YOLO5 network to obtain the surface image parameters of the landslide dam.

[0007] S3. Set up a YOLO5 network based on the surface image parameters of the landslide dam, and use the YOLO5 network to perform particle contour recognition on the input third training subset to obtain the gradation curve of the landslide dam surface image, thus completing the particle group zoning recognition of the landslide dam surface.

[0008] Further: S1 includes the following sub-steps:

[0009] S11. Obtain the minimum value l1, maximum value l2, and number of pixels n covered by the minimum particle size for the current surface particle size identification of the landslide dam.

[0010] S12. Calculate the relative altitude H of the UAV flight based on the image resolution and focal length;

[0011] H = f * n * r / a

[0012] In the formula, f is the focal length of the camera lens, r is the ground resolution of the image, and a is the size of the pixel.

[0013] S13. Control the drone to fly to a relative altitude and collect images of the surface of the landslide dam. Each image of the landslide dam surface should have a resolution of 1024*1024 or higher and be equal in length and width.

[0014] S14. Perform image enhancement and image calibration processing on the surface image of the landslide dam to obtain the training set of the surface image of the landslide dam, and input it into the YOLO5 network.

[0015] The training set of images of the landslide dam surface includes a first training subset, a second training subset, and a third training subset;

[0016] The first training subset includes image-enhanced images of the landslide dam surface, the second training subset includes image-calibrated images of the landslide dam surface, and the third training subset includes images of the landslide dam surface that have undergone both image enhancement and image calibration processing.

[0017] Furthermore, the YOLO5 network includes interconnected feature pyramid modules and boundary optimization modules;

[0018] The feature pyramid module has several parallel dilated convolutional layers and global average pooling layers, and the dilated convolutional layers are all connected to the global average pooling layer.

[0019] The dilation rate of each of the dilated convolutional layers can be set to 1, 6, 12 and 18 respectively, and each of the dilated convolutional layers may include 512 filters, batch normalization, ReLU function and dropout function;

[0020] The boundary optimization module is specifically a conditional random field.

[0021] The beneficial effects of the above-mentioned further scheme are: the global average pooling layer can extract image-level features, which helps to capture long-range information beyond the capabilities of dilated convolutional layers, and the boundary optimization module helps to refine object boundaries.

[0022] Furthermore, the YOLO5 network also includes an image classification module, which is connected to the boundary optimization module.

[0023] The image classification module is equipped with a normalization exponential function and a loss function.

[0024] The normalized exponential function p(y) i =j|xi The specific expression for ) is:

[0025]

[0026] In the formula, x ij The i-th sample is grouped into the j-th class, where j is the number of classes, x i Let y be the vector of unnormalized scores at image position i. i For x i The tag, x ic The i-th sample is grouped into the c-th class;

[0027] The specific expression for the loss function J(Θ) is as follows:

[0028]

[0029] In the formula, M is the number of samples, C is the number of classes, Θ is the network parameter, 1 / {·} is the indicator function, λ is the balance coefficient, and y i For x i The tag.

[0030] Further: In S2, the surface image parameters of the landslide dam include the surface image partitioning parameters of the landslide dam and the particle identification parameters of the surface of the landslide dam;

[0031] Specifically, S2 is:

[0032] (1) Set the training parameters of the YOLO5 network according to the surface images of the landslide dam in the first training subset, input the first training subset into the YOLO5 network, and perform multi-core parallel acceleration training through GPU to obtain the partition parameters of the surface images of the landslide dam.

[0033] (2) Set the training parameters of the YOLO5 network based on the surface images of the landslide dam in the second training subset, input the second training subset into YOLO5 for training, and accelerate the training through multi-core parallel processing using GPU to obtain the particle recognition parameters of the landslide dam surface.

[0034] Further: S3 includes the following sub-steps:

[0035] S31. Set the training parameters of the YOLO5 network according to the partition parameters of the surface image of the landslide dam, input the third training subset into the YOLO5 network to identify the partition type, obtain the contour of the partition of the surface image of the landslide dam in the third training subset, and use it as the fourth training subset.

[0036] S32. Set the training parameters of the YOLO5 network according to the particle recognition parameters on the surface of the landslide dam, input the fourth training subset into the YOLO5 network to identify the partition type, and obtain the outline of the particles on the surface of the landslide dam in the fourth training subset.

[0037] S33. Perform triangulation on the contours of particles in the surface image of the landslide dam in the fourth training subset, and calculate the relative area S of each particle in the surface image of the landslide dam in the fourth training subset using the following formula. i :

[0038]

[0039] In the formula, (x i, y i ) is the coordinate of the i-th point of a particle outline polygon, and n is the total number of coordinate points;

[0040] S34. Calculate the actual area of ​​each particle based on the relative area of ​​each particle in the surface image of the landslide dam in the fourth training subset, and obtain the longest major axis and the shortest minor axis of each particle based on the actual area.

[0041] The specific expression for calculating the actual area S′1 of each particle is as follows:

[0042] S′1=H×a×S i

[0043] S35. Calculate the length, volume and mass of each particle in the depth direction based on the actual area, longest major axis and shortest minor axis of each particle in the surface image of the landslide dam in the fourth training subset.

[0044] S36. Based on the length, volume, and mass of each particle in the depth direction of the surface image of the landslide dam in the fourth training subset, obtain the mass distribution table of all particles in the surface image of the landslide dam, and obtain the gradation curve of the surface image of the landslide dam according to the drawing specifications of the gradation curve, thus completing the particle group zoning identification of the surface of the landslide dam.

[0045] Further: In S32, when the surface image of the landslide dam in the fourth training subset has several partitions, the training parameters of the YOLO5 network are set according to the particle recognition parameters of the landslide dam surface corresponding to the several partitions. The partition type is identified in the fourth training subset according to the training parameters of the YOLO5 network, and the contours of the particles in the landslide dam surface image obtained by all training parameters are merged to obtain the contours of the particles in the landslide dam surface image.

[0046] Further: In step S35, the length c of each particle in the depth direction is calculated. i Volume V i and quality M i The specific expression is as follows:

[0047] c i =0.5×α×(a i +b i )

[0048]

[0049] M i =V i ×ρ

[0050] In the formula, α is the length conversion factor, ρ is the average material density of the current partition type, and a i For the longest major axis of each particle, b i The shortest axis for each particle.

[0051] The beneficial effects of this invention are as follows:

[0052] (1) The present invention provides a method for intelligent identification of surface particle groups of landslide dams. When establishing a YOLO5 network, it utilizes multi-scale and global context information, resolution preservation and boundary refinement to achieve better habitat element extraction. It designs a feature pyramid module and a boundary optimization module to meet the habitat element extraction of remote sensing images and to meet the accuracy requirements for environmental element extraction and identification in practical engineering applications.

[0053] (2) Compared with traditional indoor experimental methods, the present invention is more efficient, simplifies the process of identifying particles on the surface of the landslide dam, and improves the speed of grouping particles on the surface of the landslide dam. Attached Figure Description

[0054] Figure 1 This is a flowchart of the method of the present invention.

[0055] Figure 2 These are the original images taken in Example 2.

[0056] Figure 3 This is a diagram showing the results of particle group identification on the surface of the dam in Example 2. Detailed Implementation

[0057] 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.

[0058] like Figure 1 As shown, in one embodiment of the present invention, a method for intelligent identification of particle zoning on the surface of a landslide dam includes the following steps:

[0059] S1. Collect images of the surface of the landslide dam to obtain a training set of images of the surface of the landslide dam, which includes the first to third training subsets;

[0060] S2. Input the first and second training subsets into the YOLO5 network to obtain the surface image parameters of the landslide dam.

[0061] S3. Set up a YOLO5 network based on the surface image parameters of the landslide dam, and use the YOLO5 network to perform particle contour recognition on the input third training subset to obtain the gradation curve of the landslide dam surface image, thus completing the particle group zoning recognition of the landslide dam surface.

[0062] S11. Obtain the minimum value l1, maximum value l2, and number of pixels n covered by the minimum particle size for the current surface particle size identification of the landslide dam.

[0063] S12. Calculate the relative altitude H of the UAV flight based on the image resolution and focal length;

[0064] H = f * n * r / a

[0065] In the formula, f is the focal length of the camera lens, r is the ground resolution of the image, and a is the size of the pixel.

[0066] S13. Control the drone to fly to a relative altitude and collect images of the surface of the landslide dam. Each image of the landslide dam surface should have a resolution of 1024*1024 or higher and be equal in length and width.

[0067] S14. Perform image enhancement and image calibration processing on the surface image of the landslide dam to obtain the training set of the surface image of the landslide dam, and input it into the YOLO5 network.

[0068] The training set of images of the landslide dam surface includes a first training subset, a second training subset, and a third training subset;

[0069] The first training subset includes image-enhanced images of the landslide dam surface, the second training subset includes image-calibrated images of the landslide dam surface, and the third training subset includes images of the landslide dam surface that have undergone both image enhancement and image calibration processing.

[0070] In S2, the YOLO5 network includes interconnected feature pyramid modules and boundary optimization modules;

[0071] The feature pyramid module has several parallel dilated convolutional layers and global average pooling layers, and the dilated convolutional layers are all connected to the global average pooling layer.

[0072] The dilation rate of each of the dilated convolutional layers can be set to 1, 6, 12 and 18 respectively, and each of the dilated convolutional layers may include 512 filters, batch normalization, ReLU function and dropout function;

[0073] The boundary optimization module is specifically a conditional random field.

[0074] The dilated convolutional layer expands the field of view of the filter and effectively incorporates multi-scale environments. The global average pooling layer can extract image-level features, which helps to capture long-range information beyond the capabilities of the dilated convolutional layer. The boundary optimization module helps to refine object boundaries. The above YOLO5 network settings can make the segmentation results more accurate and reliable.

[0075] The YOLO5 network structure incorporates dilated convolutions in the middle to improve the accuracy of small feature extraction. A designed feature pyramid module is used at different stages of ResNet-50 to extract multi-scale information. The extracted information is then processed by 3×3 convolutions (including 512 filters, batch normalization, and non-linear layers) to reduce the number of channels. Finally, a 1×1 convolution (without batch normalization and non-linear layers) is used to generate the semantic segmentation score map.

[0076] The boundary refinement module can use a Conditional Random Field (CRF) as a post-processing step to improve the segmentation results. For example, DeeplabV2+CRF can be designed to have residual structure.

[0077] Dilated convolutions, compared to regular convolutions, have an additional parameter besides the kernel size: the dilation rate. This parameter represents the magnitude of the dilation. Dilated convolutions and regular convolutions share the same kernel size (i.e., the number of parameters remains constant in neural networks). The difference lies in the larger receptive field of the dilated convolution. The receptive field is the size that the convolution kernel perceives in the image; for example, a 5×5 convolution kernel has a receptive field of 25.

[0078] The YOLO5 network also includes an image classification module, which is connected to the boundary optimization module.

[0079] The image classification module is equipped with a normalization exponential function and a loss function.

[0080] The normalized exponential function p(y) i =j|x i The specific expression for ) is:

[0081]

[0082] In the formula, x ij The i-th sample is grouped into the j-th class, where j is the number of classes, x i Let y be the vector of unnormalized scores at image position i. i For x i The tag, x ic The i-th sample is grouped into the c-th class;

[0083] The specific expression for the loss function J(Θ) is as follows:

[0084]

[0085] In the formula, M is the number of samples, C is the number of classes, Θ is the network parameter, 1 / {·} is the indicator function, λ is the balance coefficient, and y i For x i The tag.

[0086] In S2, the surface image parameters of the landslide dam include the surface image partitioning parameters of the landslide dam and the particle identification parameters of the surface of the landslide dam.

[0087] Specifically, S2 is:

[0088] (1) Set the training parameters of the YOLO5 network according to the surface images of the landslide dam in the first training subset, input the first training subset into the YOLO5 network, and perform multi-core parallel acceleration training through GPU to obtain the partition parameters of the surface images of the landslide dam.

[0089] (2) Set the training parameters of the YOLO5 network based on the surface images of the landslide dam in the second training subset, input the second training subset into YOLO5 for training, and accelerate the training through multi-core parallel processing using GPU to obtain the particle recognition parameters of the landslide dam surface.

[0090] S3 includes the following steps:

[0091] S31. Set the training parameters of the YOLO5 network according to the partition parameters of the surface image of the landslide dam, input the third training subset into the YOLO5 network to identify the partition type, obtain the contour of the partition of the surface image of the landslide dam in the third training subset, and use it as the fourth training subset.

[0092] S32. Set the training parameters of the YOLO5 network according to the particle recognition parameters on the surface of the landslide dam, input the fourth training subset into the YOLO5 network to identify the partition type, and obtain the outline of the particles on the surface of the landslide dam in the fourth training subset.

[0093] S33. Perform triangulation on the contours of particles in the surface image of the landslide dam in the fourth training subset, and calculate the relative area S of each particle in the surface image of the landslide dam in the fourth training subset using the following formula. i :

[0094]

[0095] In the formula, (x i, y i ) is the coordinate of the i-th point of a particle outline polygon, and n is the total number of coordinate points;

[0096] S34. Calculate the actual area of ​​each particle based on the relative area of ​​each particle in the surface image of the landslide dam in the fourth training subset, and obtain the longest major axis and the shortest minor axis of each particle based on the actual area.

[0097] The specific expression for calculating the actual area S′1 of each particle is as follows:

[0098] S′1=H×a×S i

[0099] S35. Calculate the length, volume and mass of each particle in the depth direction based on the actual area, longest major axis and shortest minor axis of each particle in the surface image of the landslide dam in the fourth training subset.

[0100] S36. Based on the length, volume, and mass of each particle in the depth direction of the surface image of the landslide dam in the fourth training subset, obtain the mass distribution table of all particles in the surface image of the landslide dam, and obtain the gradation curve of the surface image of the landslide dam according to the drawing specifications of the gradation curve, thus completing the particle group zoning identification of the surface of the landslide dam.

[0101] In step S32, when the surface image of the landslide dam in the fourth training subset has several partitions, the training parameters of the YOLO5 network are set according to the particle recognition parameters of the landslide dam surface corresponding to the several partitions. The partition type is identified in the fourth training subset according to the training parameters of the YOLO5 network, and the contours of the particles in the landslide dam surface image obtained by all training parameters are merged to obtain the contours of the particles in the landslide dam surface image.

[0102] In step S35, the length c of each particle in the depth direction is calculated. i Volume V i and quality M i The specific expression is as follows:

[0103] c i =0.5×α×(a i +b i )

[0104]

[0105] M i =V i ×ρ

[0106] In the formula, α is the length conversion factor, ρ is the average material density of the current partition type, and a i For the longest major axis of each particle, b i The shortest axis for each particle.

[0107] Example 2:

[0108] This embodiment is a specific experiment on a method for intelligent identification of particle zoning on the surface of a landslide dam.

[0109] In this embodiment, training was performed using multiple soil and rock aggregate image calibration data. First, aerial photography of the site was conducted using a drone, and the original images are shown below. Figure 2 As shown, model training was then carried out. Taking the landslide dammed lake of Baige as an example, identification and accuracy verification were performed. Key steps in the implementation process, apart from the model-related parts, included aerial photography of on-site images, image calibration, and accuracy verification.

[0110] Multiple representative landslide dam areas were selected, including unexcavated deposits, excavated landslide dams, and images from laboratory sampling, among other types of raw data. A typical local area was selected, and the flight altitude and spacing design standards used for UAV aerial photography are shown in Table 1. The distance is based on the highest point in the area as the starting elevation.

[0111] Table 1

[0112]

[0113]

[0114] The obtained photos are first reviewed and verified, and photos that are obstructed, outside the research scope, or do not meet the requirements are removed. Image enhancement is then performed through steps such as contrast enhancement and thresholding. The pixel size of each image is guaranteed to be consistent at 1080*1080.

[0115] Images taken at a flight altitude of 100 meters are used as the zonal calibration images, while images taken at flight altitudes of 100, 60, 40, and 20 meters are used for particle size identification. For particle sizes smaller than those that can be identified by the image, calibration is performed directly according to the designated area. Both types of images are calibrated separately, and the calibration parameters are used to differentiate between particle types.

[0116] The calibrated photos and results are input into the improved YOLO5 model according to a fixed format for training. The required GPU hardware for training is a professional computing graphics card; the higher the photo resolution, the more video memory is needed, ensuring sufficient video memory for computation. After training, the data will be saved for further training, recognition, and computation. This example uses a Quadro P6000 graphics card with 24GB of video memory.

[0117] Based on the principle of cross-validation, a new orthophoto of the unknown region is taken and input into the parameters of two models for image partitioning and particle size identification, respectively. Then, particle size distribution curve identification is performed.

[0118] Different length conversion coefficients α are used for verification during the identification process. By selecting appropriate parameters and following the proposed method, a high-precision gradation curve can be obtained.

[0119] The final value is the average of the three data points, selected from a small subset of known, closely related data. The specific method for determining this value is shown in Table 2.

[0120] Table 2

[0121]

[0122] Recognition results as follows Figure 3 As shown.

[0123] The beneficial effects of this invention are as follows: The intelligent identification method for surface particle division of landslide dams provided by this invention utilizes multi-scale and global contextual information, resolution preservation and boundary refinement when establishing a YOLO5 network to achieve better habitat element extraction. The design of feature pyramid module and boundary optimization module makes it conform to the habitat element extraction of remote sensing images, so as to meet the accuracy requirements for environmental element extraction and identification in practical engineering applications.

[0124] Compared with traditional indoor experimental methods, this invention is more efficient, simplifies the process of identifying particles on the surface of landslide dams, and increases the speed of grouping particles on the surface of landslide dams.

[0125] In the description of this invention, it should be understood that the terms "center," "thickness," "upper," "lower," "horizontal," "top," "bottom," "inner," "outer," and "radial," etc., indicating orientation or positional relationships based on the orientation or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying the relative importance or the number of technical features implicitly specified. Therefore, a feature defined by "first," "second," and "third" may explicitly or implicitly include one or more of that feature.

Claims

1. A method for intelligent identification of particle zoning on the surface of a landslide dam, characterized in that, Includes the following steps: S1. Collect images of the surface of the landslide dam to obtain a training set of images of the landslide dam surface, which includes the first to third training subsets; the first training subset includes image-enhanced images of the landslide dam surface, the second training subset includes image-calibrated images of the landslide dam surface, and the third training subset includes images of the landslide dam surface that have undergone image enhancement and image calibration processing. S2. Input the first and second training subsets into the YOLO5 network to obtain the surface image parameters of the landslide dam; S2 specifically involves: (1) Set the training parameters of the YOLO5 network according to the surface images of the landslide dam in the first training subset, input the first training subset into the YOLO5 network, and perform multi-core parallel acceleration training through GPU to obtain the partition parameters of the surface images of the landslide dam. (2) Set the training parameters of the YOLO5 network according to the surface images of the landslide dam in the second training subset, input the second training subset into YOLO5 for training, and accelerate the training through multi-core parallel processing using GPU to obtain the particle recognition parameters of the landslide dam surface. In S2, the YOLO5 network includes interconnected feature pyramid modules and boundary optimization modules; The feature pyramid module has several parallel dilated convolutional layers and global average pooling layers, and the dilated convolutional layers are all connected to the global average pooling layer. The dilation rates of each of the dilated convolutional layers are set to 1, 6, 12 and 18, respectively, and each of the dilated convolutional layers includes 512 filters, batch normalization, ReLU function and dropout function; The boundary optimization module is specifically a conditional random field; S3. Set up a YOLO5 network based on the surface image parameters of the landslide dam, and use the YOLO5 network to perform particle contour recognition on the input third training subset to obtain the gradation curve of the landslide dam surface image, thus completing the particle group zoning identification of the landslide dam surface; S3 includes the following sub-steps: S31. Set the training parameters of the YOLO5 network according to the partition parameters of the surface image of the landslide dam, input the third training subset into the YOLO5 network to identify the partition type, obtain the contour of the partition of the surface image of the landslide dam in the third training subset, and use it as the fourth training subset. S32. Set the training parameters of the YOLO5 network according to the particle recognition parameters on the surface of the landslide dam, input the fourth training subset into the YOLO5 network to identify the partition type, and obtain the outline of the particles on the surface of the landslide dam in the fourth training subset. S33. Perform triangulation on the contours of particles in the surface images of the landslide dam in the fourth training subset, and calculate the relative area of ​​each particle in the surface images of the landslide dam in the fourth training subset using the following formula. S i : In the formula, It is the first of the particle outline polygons i The coordinates of the points n This represents the total number of coordinate points. S34. Calculate the actual area of ​​each particle based on the relative area of ​​each particle in the surface image of the landslide dam in the fourth training subset, and obtain the longest major axis and the shortest minor axis of each particle based on the actual area. This involves calculating the actual area of ​​each particle. The specific expression is: S35. Calculate the length, volume and mass of each particle in the depth direction based on the actual area, longest major axis and shortest minor axis of each particle in the surface image of the landslide dam in the fourth training subset. S36. Based on the length, volume, and mass of each particle in the depth direction of the surface image of the landslide dam in the fourth training subset, obtain the mass distribution table of all particles in the surface image of the landslide dam, and obtain the gradation curve of the surface image of the landslide dam according to the drawing specifications of the gradation curve, thus completing the particle group zoning identification of the surface of the landslide dam.

2. The intelligent identification method for particle zoning on the surface of a landslide dam according to claim 1, characterized in that, S1 includes the following steps: S11. Obtain the minimum value of the current surface particle size identification of the landslide dam. l 1. Maximum value l 2 and the number of pixels covered by the smallest grain size n ; S12. Calculate the relative altitude H of the UAV flight based on the image resolution and focal length; H = f * n * r / a In the formula, f The focal length of the camera lens. r The ground resolution of the image. a The size of the pixel; S13. Control the drone to fly to a relative altitude and collect images of the surface of the landslide dam. Each image of the landslide dam surface should have a resolution of 1024*1024 or higher and be equal in length and width. S14. Perform image enhancement and image calibration processing on the surface image of the landslide dam to obtain the training set of the surface image of the landslide dam, and input it into the YOLO5 network. The training set of images of the landslide dam surface includes a first training subset, a second training subset, and a third training subset; The first training subset includes image-enhanced images of the landslide dam surface, the second training subset includes image-calibrated images of the landslide dam surface, and the third training subset includes images of the landslide dam surface that have undergone both image enhancement and image calibration processing.

3. The intelligent identification method for particle zoning on the surface of a landslide dam according to claim 1, characterized in that, The YOLO5 network also includes an image classification module, which is connected to the boundary optimization module. The image classification module is equipped with a normalization exponential function and a loss function. The normalized exponential function The specific expression is: In the formula, For the first i The sample was grouped into the first... j kind, j For the number of categories, Image location i The vector of unstandardized scores, y i for x i The tag, x ic For the first i The samples were grouped into class c; The loss function The specific expression is: In the formula, M For the sample size, C For the number of categories, Θ Here are the network parameters, and {·} is the indicator function. λ For balance coefficient, y i for x i The tag.

4. The intelligent identification method for particle zoning on the surface of a landslide dam according to claim 1, characterized in that, In step S32, when the surface image of the landslide dam in the fourth training subset has several partitions, the training parameters of the YOLO5 network are set according to the particle recognition parameters of the landslide dam surface corresponding to the several partitions. The partition type is identified in the fourth training subset according to the training parameters of the YOLO5 network, and the contours of the particles in the landslide dam surface image obtained by all training parameters are merged to obtain the contours of the particles in the landslide dam surface image.

5. The intelligent identification method for particle zoning on the surface of a landslide dam according to claim 1, characterized in that, In step S35, the length of each particle in the depth direction is calculated. c i ,volume and quality The specific expression is as follows: In the formula, This is the length conversion factor. The average density of the material for the current partition type. The longest major axis of each particle, The shortest axis for each particle, The length of each particle in the depth direction, For volume.

Images