Dam termite nest three-dimensional imaging method and system based on ground penetrating radar

Abstract

The invention discloses a dam termite nest three-dimensional imaging method and system based on a ground penetrating radar, and the method comprises the steps: obtaining the three-dimensional ground penetrating radar data of a target region, and carrying out the data preprocessing, and obtaining the preprocessed three-dimensional ground penetrating radar data; performing feature extraction processing and encryption processing on the preprocessed three-dimensional ground penetrating radar data to obtain a multi-channel radar data body; based on the U-shaped network structure, constructing an ant nest imaging three-dimensional deep learning network model; and performing data imaging processing on the multi-channel radar data volume based on the ant nest imaging three-dimensional deep learning network model to obtain an ant nest imaging result based on the three-dimensional ground penetrating radar. According to the embodiment of the invention, the complex features of the ant nest and the dam structure on the three-dimensional ground penetrating radar image can be more accurately captured, and then the accuracy of ant nest imaging is improved. The method can be widely applied to the technical field of termite nest detection.

Description

Technical Field

[0001] The present application relates to the technical field of termite nest detection, and in particular to a three-dimensional imaging method and system for dam termite nests based on ground penetrating radar. Background Art

[0002] Due to the concealment and structural complexity of termite nests, the detection of termite nests on dams has always been a difficult problem in water conservancy projects. At present, termite nest detection is mainly divided into three types: manual exploration, geophysical detection, and biological characteristic detection. Among them, three-dimensional radar detection generates high-resolution underground structure images through the coordinated work of multiple transmitting antennas and multiple receiving antennas, which has certain advantages in termite nest detection. Through multi-frequency scanning, three-dimensional radar can obtain three-dimensional data inside the dam. However, the current data processing method of three-dimensional radar has certain limitations. First, the amount of three-dimensional data is large, so data processing takes a lot of time and the processing steps are relatively complicated. In addition, three-dimensional radar technology obtains three-dimensional electromagnetic wave reflection signals, which appear as waveform features rather than object shapes in the image. Therefore, the three-dimensional imaging of ant nests obtained by the current three-dimensional radar imaging method based on clustering or thresholding is inaccurate.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the invention

[0004] The main purpose of the embodiments of the present application is to propose a three-dimensional imaging method and system for termite nests on dams based on ground penetrating radar, which can more accurately determine the location and shape of the ant nest and capture the complex features of the ant nest and dam structure on the three-dimensional ground penetrating radar image, thereby improving the efficiency and accuracy of ant nest imaging.

[0005] To achieve the above-mentioned purpose, one aspect of the embodiment of the present application proposes a three-dimensional imaging method of termite nests on dams based on ground penetrating radar, the method comprising:

[0006] Acquire three-dimensional ground penetrating radar data of a target area and perform data preprocessing to obtain preprocessed three-dimensional ground penetrating radar data;

[0007] Performing feature extraction and encryption processing on the preprocessed three-dimensional ground penetrating radar data to obtain a multi-channel radar data volume;

[0008] Based on the U-shaped network structure, a three-dimensional deep learning network model for ant nest imaging is constructed;

[0009] The multi-channel radar data volume is subjected to data imaging processing based on the ant nest imaging three-dimensional deep learning network model to obtain a three-dimensional ground penetrating radar ant nest imaging result.

[0010] In some embodiments, the step of acquiring the three-dimensional ground penetrating radar data of the target area and performing data preprocessing to obtain the preprocessed three-dimensional ground penetrating radar data includes:

[0011] Determine a target area, and scan the target area using a three-dimensional radar to obtain a two-dimensional radar image of the target area;

[0012] Combining the two-dimensional radar images of the target area in a direction perpendicular to the survey line to obtain three-dimensional ground penetrating radar data of the target area;

[0013] The three-dimensional ground penetrating radar data of the target area is preprocessed to obtain the preprocessed three-dimensional ground penetrating radar data.

[0014] In some embodiments, the preprocessing of the three-dimensional ground penetrating radar data of the target area to obtain the preprocessed three-dimensional ground penetrating radar data includes:

[0015] Performing zero-time correction processing on the three-dimensional ground penetrating radar data of the target area to obtain corrected three-dimensional ground penetrating radar data;

[0016] Performing signal gain processing on the corrected three-dimensional ground penetrating radar data to obtain three-dimensional ground penetrating radar data after gain;

[0017] The three-dimensional ground penetrating radar data after gain is filtered and background noise is removed by a bandpass filter to obtain the pre-processed three-dimensional ground penetrating radar data.

[0018] In some embodiments, the step of performing feature extraction and encryption processing on the preprocessed three-dimensional ground penetrating radar data to obtain a multi-channel radar data volume includes:

[0019] Performing Hilbert transform analysis on the preprocessed three-dimensional ground penetrating radar data to obtain radar instantaneous frequency;

[0020] Performing clustering processing on the preprocessed three-dimensional ground penetrating radar data by using a k-means clustering method to obtain radar clustering data;

[0021] Performing offset processing on the preprocessed three-dimensional ground penetrating radar data to obtain radar offset data;

[0022] The preprocessed three-dimensional ground penetrating radar data, the radar instantaneous frequency, the radar clustering data and the radar offset data are combined and interpolated and encrypted to obtain the multi-channel radar data volume.

[0023] In some embodiments, performing offset processing on the preprocessed three-dimensional ground penetrating radar data to obtain radar offset data includes:

[0024] Convert the time-space domain of the preprocessed three-dimensional ground penetrating radar data into a frequency wavenumber domain to obtain converted three-dimensional ground penetrating radar data;

[0025] Determining frequency wave number information of the converted three-dimensional ground penetrating radar data according to the converted three-dimensional ground penetrating radar data;

[0026] Performing offset compensation on the frequency wavenumber information to obtain corrected frequency wavenumber domain data;

[0027] The corrected frequency wavenumber domain data is subjected to inverse two-dimensional Fourier transform and converted into time-space domain processing to obtain the radar deviation data.

[0028] In some embodiments, the ant nest imaging three-dimensional deep learning network model includes several three-dimensional convolution modules and a deep neural network structure, and the several three-dimensional convolution modules and the deep neural network structure constitute a U-shaped network structure, wherein the three-dimensional convolution module includes a three-dimensional convolution layer, a batch normalization layer and an activation function, and the deep neural network structure includes a data input module, an encoder, a decoder, a jump connection and a convolution output module.

[0029] In some embodiments, the loss function of the ant nest imaging three-dimensional deep learning network model is specifically as follows:

[0030] Loss=BCE+Focal

[0031] BCE(p,y)=-[y·log(p)+(1-y)·log(1-p)]

[0032] Focal t )=-α t (1-p t ) 2 log(p t )

[0033] In the above formula, Loss represents the ant nest imaging 3D deep learning network model, BCE(·) represents the cross entropy loss function, Focal(·) represents the focal loss function, y represents the label, p represents the probability of predicting the positive class, and p t represents the modulation factor, α t Represents a constant.

[0034] In some embodiments, the performing data imaging processing on the multi-channel radar data volume based on the ant nest imaging three-dimensional deep learning network model to obtain a three-dimensional ground penetrating radar ant nest imaging result includes:

[0035] Inputting the multi-channel radar data volume into the ant nest imaging three-dimensional deep learning network model;

[0036] Based on the three-dimensional convolution module of the ant nest imaging three-dimensional deep learning network model, the multi-channel radar data volume is automatically subjected to feature extraction processing to obtain multi-channel radar feature data;

[0037] Based on the deep neural network structure of the ant nest imaging three-dimensional deep learning network model, data imaging processing is performed on the multi-channel radar feature data to obtain the three-dimensional ground penetrating radar ant nest imaging result.

[0038] In some embodiments, the deep neural network structure based on the ant nest imaging three-dimensional deep learning network model performs data imaging processing on the multi-channel radar feature data to obtain the three-dimensional ground penetrating radar ant nest imaging result, including:

[0039] Inputting the multi-channel radar feature data into the deep neural network structure of the ant nest imaging three-dimensional deep learning network model;

[0040] Acquire the multi-channel radar feature data based on the data input module of the deep neural network structure;

[0041] Based on the encoder of the deep neural network structure, the multi-channel radar feature data is subjected to feature space resolution reduction processing to obtain reduced multi-channel radar feature data;

[0042] Based on the decoder of the deep neural network structure, performing spatial resolution recovery processing on the reduced multi-channel radar feature data to obtain restored multi-channel radar feature data;

[0043] Based on the skip connection of the deep neural network structure, the reduced multi-channel radar feature data and the restored multi-channel radar feature data are fused and down-sampled to obtain fused multi-channel radar feature data;

[0044] Based on the convolution output module of the deep neural network structure, the fused multi-channel radar feature data is classified into two categories to obtain the three-dimensional ground penetrating radar ant nest imaging result.

[0045] To achieve the above object, another aspect of the embodiment of the present application provides a three-dimensional imaging system of termite nests on dams based on ground penetrating radar, the system comprising:

[0046] The first module is used to acquire the three-dimensional ground penetrating radar data of the target area and perform data preprocessing to obtain the preprocessed three-dimensional ground penetrating radar data;

[0047] The second module is used to perform feature extraction and encryption processing on the pre-processed three-dimensional ground penetrating radar data to obtain a multi-channel radar data volume;

[0048] The third module is used to build a three-dimensional deep learning network model for ant nest imaging based on a U-shaped network structure;

[0049] The fourth module is used to perform data imaging processing on the multi-channel radar data volume based on the ant nest imaging three-dimensional deep learning network model to obtain a three-dimensional ground penetrating radar ant nest imaging result.

[0050] The embodiments of the present application include at least the following beneficial effects: The present application provides a method and system for three-dimensional imaging of termite nests on dams based on ground penetrating radar. The scheme obtains three-dimensional ground penetrating radar data of the target area and performs data preprocessing, and then performs feature extraction and encryption processing on the preprocessed three-dimensional ground penetrating radar data to obtain a multi-channel radar data body, which can better identify the positions of different medium interfaces and more accurately determine the position and shape of the ant nest, and then construct a three-dimensional deep learning network model for ant nest imaging based on a U-shaped network structure. Finally, data imaging processing is performed on the multi-channel radar data body, which can process the spatial and depth information of the ground penetrating radar data and capture the complex features of the ant nest and dam structure in the three-dimensional ground penetrating radar image, thereby improving the efficiency and accuracy of ant nest imaging. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a flow chart of a method for three-dimensional imaging of termite nests on dams based on ground penetrating radar provided in an embodiment of the present application;

[0052] Figure 2 It is a structural schematic diagram of a three-dimensional imaging system of dam termite nests based on ground penetrating radar provided in an embodiment of the present application;

[0053] Figure 3 is a schematic diagram of three-dimensional radar data collection provided by an embodiment of the present application;

[0054] Figure 4 is a schematic diagram of three-dimensional radar data processing provided by an embodiment of the present application;

[0055] Figure 5 It is a schematic diagram of the structure of a three-dimensional convolution operation provided in an embodiment of the present application;

[0056] Figure 6 It is a structural schematic diagram of the ant nest imaging three-dimensional deep learning network model provided in an embodiment of the present application;

[0057] Figure 7 It is a schematic diagram of the ant nest imaging effect of the three-dimensional ground penetrating radar provided in an embodiment of the present application. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is further described in detail below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the embodiments of the present application. They are only examples of systems and methods consistent with some aspects of the embodiments of the present application as detailed in the attached claims.

[0059] It is understood that the terms "first", "second", etc. used in this application can be used to describe various concepts in this article, but unless otherwise specified, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another concept. For example, without departing from the scope of the embodiment of the present application, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the words "if" and "if" as used herein can be interpreted as "at the time of" or "when" or "in response to determination".

[0060] The terms "at least one", "multiple", "each", "any", etc. used in this application, at least one includes one, two or more, multiple includes two or more, each refers to each of the corresponding multiple, and any refers to any one of the multiple.

[0061] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application and are not intended to limit this application.

[0062] Reference Figure 1 , Figure 1 A flowchart of a method for three-dimensional imaging of termite nests on dams based on ground penetrating radar provided by an embodiment of the present invention, referring to Figure 1 , the method comprises the following steps:

[0063] S100, acquiring three-dimensional ground penetrating radar data of a target area and performing data preprocessing to obtain preprocessed three-dimensional ground penetrating radar data;

[0064] It should be noted that, in some embodiments, step S100 may include:

[0065] S110, determining a target area, scanning the target area by a three-dimensional radar, and obtaining a two-dimensional radar image of the target area;

[0066] S120, combining the two-dimensional radar images of the target area in a direction perpendicular to the survey line to obtain three-dimensional ground penetrating radar data of the target area;

[0067] In steps S110 to S120 of some embodiments, a three-dimensional ground penetrating radar is used to scan the survey area, and the collected data is preprocessed and initially imaged.

[0068] Specifically, the center frequency of the three-dimensional radar is 470MHz. When detecting, the main controller controls the antennas in the antenna array to excite each transmitting antenna at different times, and the receiving antenna receives the reflected signal within the corresponding time window. The transmitting antenna of the three-dimensional radar emits electromagnetic waves, which will be reflected and refracted in different media. The reflection coefficient of the electromagnetic wave at the interface of different media is:

[0069]

[0070] In the above formula, R represents the reflection coefficient of electromagnetic waves at the interface of different media, ε1 represents the dielectric constant of the upper medium, and ε2 represents the dielectric constant of the lower medium.

[0071] like Figure 3 As shown in the figure, the dielectric constant of the ant nest in the dam is quite different from that of the dam building materials, and the electromagnetic waves will be reflected. Therefore, when the 3D radar scans above the dam, it will receive the electromagnetic waves reflected by the ant nest. The 2D radar images formed by each transmitting-receiving antenna pair are combined in the direction perpendicular to the survey line to obtain the 3D ground penetrating radar data.

[0072] S130 , performing data preprocessing on the three-dimensional ground penetrating radar data of the target area to obtain preprocessed three-dimensional ground penetrating radar data.

[0073] It should be noted that in this embodiment, the collected radar data is preprocessed, and the purpose of the preprocessing is to remove noise in the radar data, improve signal quality, and highlight the reflected wave of the target body.

[0074] Wherein, step S130 further includes:

[0075] S131, performing zero-time correction processing on the three-dimensional ground penetrating radar data of the target area to obtain corrected three-dimensional ground penetrating radar data;

[0076] In this embodiment S131, the electromagnetic waves emitted by the transmitting antenna need a certain distance to propagate to the ground, so the position where the data time window takes 0 is not the ground, so zero time correction is required to improve the accuracy of depth calculation. The calculation method is to discard the data above the ground phase axis.

[0077] S132, performing signal gain processing on the corrected three-dimensional ground penetrating radar data to obtain three-dimensional ground penetrating radar data after gain;

[0078] In this embodiment S132, the electromagnetic waves emitted by the radar will attenuate during the propagation process in the underground medium, resulting in weak signals in the deep, so it is necessary to perform gain processing on the radar data. Assuming that the signal received by the radar receiving antenna is s(t), a gain window T is specified, and the root mean square value of the signal at each time point t is:

[0079]

[0080] In the above formula, E(t) represents the RMS value of the signal at each time point t, T represents the gain window, t represents the time point, s represents the signal received by the receiving antenna, and τ represents the integral independent variable.

[0081] The root mean square of the signal at each time point t is the local energy of the signal. The gain factor is calculated based on the local energy of the signal, and its expression is:

[0082]

[0083] In the above formula, G(t) represents the gain factor, and ε is a very small positive number to avoid the denominator being zero.

[0084] S133, filtering and background noise removal are performed on the three-dimensional ground penetrating radar data after gain through a bandpass filter to obtain pre-processed three-dimensional ground penetrating radar data.

[0085] In this embodiment S133, a bandpass filter is used to filter out interference waves received by the receiving antenna. Then, all measured values ​​are horizontally added according to the overall average technology to obtain an average value, and then the average value is subtracted from the initial data to achieve the processing purpose of eliminating background interference data.

[0086] S200, performing feature extraction and encryption processing on the pre-processed three-dimensional ground penetrating radar data to obtain a multi-channel radar data volume;

[0087] It should be noted that, in some embodiments, step S200 may include:

[0088] S210, performing Hilbert transform analysis on the pre-processed three-dimensional ground penetrating radar data to obtain the radar instantaneous frequency;

[0089] In this embodiment S210, the instantaneous frequency of the ground penetrating radar refers to the frequency of the signal at a certain moment in the process of the signal changing with time, reflecting the dynamic change of the signal frequency with time. In the transition layer or reflection interface of different media, the instantaneous frequency will change significantly. By analyzing the instantaneous frequency, the layered structure or reflection interface of the underground medium can be located. Assuming that the received radar signal is s(t), its analytical signal is obtained by Hilbert transform, and its expression is:

[0090] z(t)=s(t)+j·H[s(t)]

[0091] In the above formula, H[s(t)] is the Hilbert transform of the radar signal s(t), and z(t) is the complex representation of the analytical signal.

[0092] Furthermore, the instantaneous frequency f(t) is the derivative of the analytical signal phase φ(t) with respect to time, and its expression is:

[0093]

[0094] In the above formula, φ(t) = arg(z(t)) is the phase of the analytical signal. The instantaneous frequency of the two-dimensional radar data formed by each transmitting-receiving antenna pair of the three-dimensional radar is calculated, and the following is obtained: Figure 4 Data shown.

[0095] S220, performing clustering processing on the preprocessed three-dimensional ground penetrating radar data by using a k-means clustering method to obtain radar clustering data;

[0096] In this embodiment S220, the k-means clustering method is used to process the three-dimensional radar data. The role of k-means clustering is to perform data pattern recognition and target classification on the radar data, which helps to extract the target echo in the radar signal. K-means clustering is performed on the two-dimensional radar data formed by each transmitting-receiving antenna pair of the three-dimensional radar to obtain the following Figure 4 Data shown.

[0097] S230, performing offset processing on the preprocessed three-dimensional ground penetrating radar data to obtain radar offset data;

[0098] Wherein, step S230 further includes:

[0099] S231, converting the time-space domain of the preprocessed three-dimensional ground penetrating radar data into a frequency wavenumber domain to obtain converted three-dimensional ground penetrating radar data;

[0100] S232, determining frequency wave number information of the converted three-dimensional ground penetrating radar data according to the converted three-dimensional ground penetrating radar data;

[0101] S233, performing offset compensation on the frequency wave number information to obtain corrected frequency wave number domain data;

[0102] S234, performing inverse two-dimensional Fourier transform and conversion into time and space domain processing on the corrected frequency wavenumber domain data to obtain radar offset data.

[0103] In this embodiment S230, the three-dimensional radar data is subjected to an offset process. The purpose of the offset process is to correct the reflected echo of the target to the real position of the target. When performing the offset process, the radar data is first converted from the time-space domain (tx domain) to the frequency wave number domain (fk domain),

[0104]

[0105] In the above formula, d(x, t) is the radar data in the time and space domain, and D(k, f) represents the radar data in the wave number k and frequency f domain.

[0106] In the frequency-wavenumber domain, the relationship between frequency and wavenumber is expressed as:

[0107]

[0108] In the above formula, v is the speed of electromagnetic waves in the medium.

[0109] In the migration process, the offset wave is corrected to the correct position by using the known wave velocity information. This process involves compensating the frequency corresponding to each wave number accordingly. The compensation process is to calculate a new wave number k based on the given wave velocity v. new , whose expression is:

[0110]

[0111] The new wave number k new That is, the corrected wave number, the new frequency-wave number domain data D(k new ,f) Perform an inverse two-dimensional Fourier transform and convert it back to the time-space domain to obtain the offset data d shift (x,t).

[0112]

[0113] In the above formula, d shifted (x, t) represents the offset data, D(k new ,f) represents the new frequency-wavenumber domain data.

[0114] S240, combining the preprocessed three-dimensional ground penetrating radar data, radar instantaneous frequency, radar clustering data and radar offset data and performing interpolation encryption to obtain a multi-channel radar data volume.

[0115] In S240 of this embodiment, the instantaneous frequency, clustering data and offset data of the calculated three-dimensional radar data are interpolated and encrypted to obtain a three-dimensional radar data volume of size 256*256*64 (track*time*line), such as Figure 4 shown.

[0116] In summary, the embodiment of the present invention not only processes three-dimensional radar reflection data, but also combines multi-dimensional information such as instantaneous frequency, clustering data, and offset data. Data of different dimensions have different functions.

[0117] 1) Raw radar data: By preprocessing the data (denoising, gain correction), noise interference can be removed and the signal quality can be enhanced. This is the reflection signal directly obtained by the ground penetrating radar, which contains the propagation and reflection information of electromagnetic waves in the underground medium. It reflects the electromagnetic reflection characteristics of different underground structures.

[0118] 2) Instantaneous frequency data: The instantaneous frequency reflects the frequency distribution of radar signals over time and explains the propagation characteristics of electromagnetic waves in different media. At the interface of materials, the instantaneous frequency will change significantly, so the location of the interface between different media can be better identified by the change of instantaneous frequency.

[0119] 3) Clustering data: K-means clustering is used to perform pattern recognition and classification on radar data, and different areas in the reflected signal are classified according to data characteristics, thereby separating the target (such as an ant nest) from the background signal.

[0120] 4) Offset data: By offsetting the position of the corrected echo and adjusting the reflected signal from the recorded position to the actual position of the target, the position and shape of the ant nest can be determined more accurately, avoiding position errors caused by the wave propagation path.

[0121] The above data are encrypted through interpolation to form a multi-channel data body, and the different information of the above data are fused in the network input, providing a more comprehensive ant nest feature representation, which helps to improve the accuracy of imaging.

[0122] S300, based on the U-shaped network structure, build a 3D deep learning network model for ant nest imaging;

[0123] It should be noted that in some embodiments, the three-dimensional ground penetrating radar data is imaged by a three-dimensional deep neural network. The deep neural network is composed of Figure 5 The three-dimensional convolution module is composed of the following three-dimensional convolution modules: Figure 6 The specific hierarchical network structure parameters of the “U”-shaped network shown are shown in Table 1.

[0124] The ant nest imaging three-dimensional deep learning network model includes several three-dimensional convolution modules and deep neural network structures. Several three-dimensional convolution modules and the deep neural network structure constitute a U-shaped network structure, wherein the three-dimensional convolution module includes a three-dimensional convolution layer, a batch normalization layer and an activation function, and the deep neural network structure includes a data input module, an encoder, a decoder, a jump connection and a convolution output module.

[0125] In addition, it should be noted that the traditional 3D GPR data processing method relies on clustering or thresholding techniques, and the imaging results are not accurate enough. The embodiment of the present invention uses a 3D convolution module to process the spatial and depth information of GPR data. The 3D convolution network can capture the complex features of ant nests and dam structures on 3D GPR images, which is more suitable for 3D data processing than the traditional 2D convolution. The 3D convolution neural network structure in the embodiment of the present invention, the 3D convolution module forms a "U"-shaped network structure, and the 3D convolution module improves the accuracy and resolution of imaging by automatically extracting and modeling multi-channel 3D GPR data. The skip connection is combined to retain detail information, solving the contradiction between high-level feature extraction and reduced spatial resolution.

[0126] Table 1 Network structure hierarchy data table

[0127]

[0128]

[0129] S400, performing data imaging processing on a multi-channel radar data volume based on an ant nest imaging three-dimensional deep learning network model to obtain a three-dimensional ground penetrating radar ant nest imaging result;

[0130] It should be noted that, in some embodiments, step S400 may include:

[0131] S410, inputting the multi-channel radar data volume into the ant nest imaging three-dimensional deep learning network model;

[0132] S420, performing automatic feature extraction processing on the multi-channel radar data volume using a three-dimensional convolution module based on the ant nest imaging three-dimensional deep learning network model to obtain multi-channel radar feature data;

[0133] In the present embodiment S420, the three-dimensional convolution module is used to process three-dimensional data with space and depth. The input of the three-dimensional convolution operation is (C, H, W, D), where C is the number of channels, H is the height, W is the width, and D represents the depth. The size of the three-dimensional convolution kernel is 3*3*3. When calculating, the convolution kernel slides in the depth, height and width directions of the input, and at each position, the convolution kernel and the corresponding part of the input are element-wise multiplied to obtain a numerical output. Each value of the output feature map is obtained by multiplying a three-dimensional neighborhood area of ​​the input data with the convolution kernel. Batch normalization and activation function operations are performed after each three-dimensional convolution operation. The batch normalization operation adjusts the output mean of each small batch of data to 0 and the variance to 1, thereby reducing the offset of the input feature. The activation function layer is composed of an activation function, which is defined as f(x)=max(0,x), that is, all negative values ​​are set to 0 and positive values ​​remain unchanged. By introducing the activation function, the linear transformation is converted into a more complex nonlinear mapping, enhancing the expression ability of the network. Therefore, the composition of the 3D convolution module is "3D convolution + batch normalization + activation function".

[0134] S430, based on the deep neural network structure of the ant nest imaging three-dimensional deep learning network model, perform data imaging processing on the multi-channel radar feature data to obtain the three-dimensional ground penetrating radar ant nest imaging result.

[0135] Wherein, step S430 further includes:

[0136] S431, inputting the multi-channel radar feature data into the deep neural network structure of the ant nest imaging three-dimensional deep learning network model;

[0137] S432, a data input module based on a deep neural network structure acquires multi-channel radar feature data;

[0138] In this embodiment S432, the network input is the original radar data volume, instantaneous frequency volume, clustering data volume and offset data volume after interpolation encryption, and multiple data volumes form a multi-channel radar data volume with a shape of (B, 4, 256, 256, 64), where B is the batch size. The above data volumes first pass through a convolution layer to extract preliminary low-level features and convert the data from the input channel to a higher feature channel.

[0139] S433, based on the encoder of the deep neural network structure, performing feature space resolution reduction processing on the multi-channel radar feature data to obtain reduced multi-channel radar feature data;

[0140] In this embodiment S433, the encoder is used to gradually extract the features of the input data and compress the original data into a feature representation with a higher dimension. The encoder reduces the spatial resolution of the features through a downsampling operation (implemented by maximum pooling) and extracts more complex and abstract features through multi-layer convolution.

[0141] The feature extraction function of the encoder is realized through three-dimensional convolution. The data is passed through the convolution layer to extract high-level features, and the number of channels of the data will gradually increase. The data after three-dimensional convolution undergoes a downsampling operation. Downsampling reduces the spatial dimension of the data through maximum pooling, with the aim of reducing the resolution of the data and retaining important global information.

[0142] A data body of size (B, C, H, W, D) (where B is the batch size, C is the number of channels, H is the height, W is the width, and D is the data depth) becomes (B, 2C, H, W, D) after the convolution operation, and becomes (B, 2C, H / 2, W / 2, D / 2) after the convolution operation and the downsampling operation.

[0143] S434, based on a decoder of a deep neural network structure, performing spatial resolution recovery processing on the reduced multi-channel radar feature data to obtain recovered multi-channel radar feature data;

[0144] In this embodiment S434, the role of the decoder is to gradually restore the spatial resolution, map the high-level features output by the encoder to the size of the original input data, and generate accurate predictions. The decoder gradually increases the spatial dimensions of the feature map through upsampling operations (implemented by bilinear interpolation), and combines the skip connection information from the encoder through convolutional layers to fuse local details and global features.

[0145] The decoding function of the decoder is implemented through 3D convolution. The input of the 3D convolution layer is the features after upsampling and the features spliced ​​from the corresponding level of the encoder through jump connection. The data after 3D convolution undergoes an upsampling operation, which restores the spatial size of the feature map to the size of the original input through bilinear interpolation.

[0146] A data volume of size (B, C, H, W, D) becomes (B, C / 2, H, W, D) after convolution operation, and becomes (B, C / 2, 2*H, 2*W, 2*D) after convolution operation and upsampling operation.

[0147] S435, based on the skip connection of the deep neural network structure, fusing and down-sampling the reduced multi-channel radar feature data and the restored multi-channel radar feature data to obtain fused multi-channel radar feature data;

[0148] In this embodiment S435, the role of the jump connection is to fuse the low-level information in the encoder with the high-level information in the decoder. After multiple downsampling operations in the encoder, although high-level semantic features are extracted, the spatial resolution is gradually reduced, resulting in the loss of low-level information. These low-level features contain details such as boundaries and textures. The jump connection directly transfers the high-resolution low-level features in the encoder to the corresponding level of the decoder, splicing them with the features in the decoder to help restore the detailed information. These low-level features contain details such as boundaries and textures.

[0149] S436, based on the convolution output module of the deep neural network structure, the fused multi-channel radar feature data is classified into two categories to obtain the three-dimensional ground penetrating radar ant nest imaging results.

[0150] In this embodiment S436, after all the upsampling and convolution processing in the decoder, the network passes through a final convolution layer to compress the multi-channel feature map to the required number of channels. Subsequently, the features output by the convolution layer are processed by an activation function, which converts the final output result into a binary classification result, where the ant nest target body is 1 and the disease-free dam body is 0, such as Figure 7 shown.

[0151] The activation function is:

[0152]

[0153] In the above formula, f(x) represents the activation function.

[0154] The input of the network is a multi-channel data volume with a shape of (B, 4, 256, 256, 64). After the final convolution processing, the output data volume of the network has a shape of (B, 1, 256, 256, 64), which is a single-channel data representing the spatial characteristics of the target body.

[0155] In addition, it should be noted that the input data of the network are the original radar data volume, the instantaneous frequency volume, the clustering data volume and the offset data volume. The input is denoted as input, the output of the network is denoted as output, and the three-dimensional deep neural network is denoted as F. Then the output of the network is:

[0156] output=F(input)

[0157] The output of the network is binary classification data. The label of each voxel in the real model is y, where y takes the value of 0 or 1, representing the dam background and the ant nest target respectively. The probability of predicting the positive class in the output of the binary classification data body output by the network is p. The output of the network is compared with the actual model to calculate the cross entropy error, which is expressed as:

[0158] BCE(p,y)=-[y·log(p)+(1-y)·log(1-p)]

[0159] At the same time, most of the network output is the area without any target, and only a small area is the area with ant nest targets, so there is a large category imbalance problem. Focus loss is used to reduce this imbalance problem.

[0160] The formula of focal loss is:

[0161] Focal t )=-α t (1-p t ) 2 log(p t )

[0162] In the above formula (1-p t ) 2 is the modulation factor, where p t for:

[0163]

[0164] When p t The closer it is to 1, the closer the modulation factor will be to 0, and the loss of easy-to-classify samples will be greatly reduced; when p t The closer it is to 0, the closer the modulation factor will be to 1, and the model will pay more attention to difficult-to-classify samples. t The expression is:

[0165]

[0166] In the above formula, α is set to 0.75 to reduce the dominant effect of negative samples on the overall loss. The final network loss is a combination of cross entropy loss and focal loss, and its expression is:

[0167] Loss=BCE+Focal

[0168] In the above formula, Loss represents the ant nest imaging 3D deep learning network model, BCE(·) represents the cross entropy loss function, and Focal(·) represents the focal loss function.

[0169] Network training optimizes the above loss function to improve the network's prediction performance.

[0170] In summary, the embodiments of the present invention perform three-dimensional radar detection on the dam and collect three-dimensional ground-penetrating radar data; perform data preprocessing on the collected three-dimensional ground-penetrating radar data; perform further data processing on the preprocessed three-dimensional ground-penetrating radar data to extract its instantaneous frequency, clustering and offset data; perform interpolation encryption on the above data to form a multi-channel radar data body; construct a three-dimensional deep learning network and perform network training optimization, and optimize the network parameters by optimizing the loss function; the measured data is input into the neural network and imaged, which can process the spatial and depth information of the ground-penetrating radar data and capture the complex features of ant nests and dam structures in the three-dimensional ground-penetrating radar images, thereby improving the efficiency and accuracy of ant nest imaging.

[0171] See also Figure 2 The embodiment of the present application further provides a 3D imaging system of dam termite nests based on ground penetrating radar, which can implement the above-mentioned 3D imaging method of dam termite nests based on ground penetrating radar. The system includes:

[0172] The first module 201 is used to acquire the three-dimensional ground penetrating radar data of the target area and perform data preprocessing to obtain the preprocessed three-dimensional ground penetrating radar data;

[0173] The second module 202 is used to perform feature extraction and encryption processing on the pre-processed three-dimensional ground penetrating radar data to obtain a multi-channel radar data volume;

[0174] The third module 203 is used to construct a three-dimensional deep learning network model for ant nest imaging based on a U-shaped network structure;

[0175] The fourth module 204 is used to perform data imaging processing on the multi-channel radar data volume based on the ant nest imaging three-dimensional deep learning network model to obtain the three-dimensional ground penetrating radar ant nest imaging results.

[0176] It can be understood that the contents of the above method embodiments are all applicable to the present system embodiments, the functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0177] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by a person skilled in the art without departing from the scope and essence of the present invention should be within the scope of the rights of the present invention.

Claims

1. A three-dimensional imaging method of termite nests on dams based on ground penetrating radar, characterized in that: The method comprises the following steps: Acquire three-dimensional ground penetrating radar data of a target area and perform data preprocessing to obtain preprocessed three-dimensional ground penetrating radar data; Performing feature extraction and encryption processing on the preprocessed three-dimensional ground penetrating radar data to obtain a multi-channel radar data volume; Based on the U-shaped network structure, a three-dimensional deep learning network model for ant nest imaging is constructed; The multi-channel radar data volume is subjected to data imaging processing based on the ant nest imaging three-dimensional deep learning network model to obtain a three-dimensional ground penetrating radar ant nest imaging result.

2. The method according to claim 1, characterized in that The method of acquiring the three-dimensional ground penetrating radar data of the target area and performing data preprocessing to obtain the preprocessed three-dimensional ground penetrating radar data includes: Determine a target area, and scan the target area using a three-dimensional radar to obtain a two-dimensional radar image of the target area; Combining the two-dimensional radar images of the target area in a direction perpendicular to the survey line to obtain three-dimensional ground penetrating radar data of the target area; The three-dimensional ground penetrating radar data of the target area is preprocessed to obtain the preprocessed three-dimensional ground penetrating radar data.

3. The method according to claim 2, characterized in that The step of preprocessing the three-dimensional ground penetrating radar data of the target area to obtain the preprocessed three-dimensional ground penetrating radar data includes: Performing zero-time correction processing on the three-dimensional ground penetrating radar data of the target area to obtain corrected three-dimensional ground penetrating radar data; Performing signal gain processing on the corrected three-dimensional ground penetrating radar data to obtain three-dimensional ground penetrating radar data after gain; The three-dimensional ground penetrating radar data after gain is filtered and background noise is removed by a bandpass filter to obtain the pre-processed three-dimensional ground penetrating radar data.

4. The method according to claim 1, characterized in that: The step of performing feature extraction and encryption processing on the preprocessed three-dimensional ground penetrating radar data to obtain a multi-channel radar data volume includes: Performing Hilbert transform analysis on the preprocessed three-dimensional ground penetrating radar data to obtain radar instantaneous frequency; Performing clustering processing on the preprocessed three-dimensional ground penetrating radar data by using a k-means clustering method to obtain radar clustering data; Performing offset processing on the preprocessed three-dimensional ground penetrating radar data to obtain radar offset data; The preprocessed three-dimensional ground penetrating radar data, the radar instantaneous frequency, the radar clustering data and the radar offset data are combined and interpolated and encrypted to obtain the multi-channel radar data volume.

5. The method according to claim 4, characterized in that The performing offset processing on the preprocessed three-dimensional ground penetrating radar data to obtain radar offset data includes: Convert the time-space domain of the preprocessed three-dimensional ground penetrating radar data into a frequency wavenumber domain to obtain converted three-dimensional ground penetrating radar data; Determining frequency wave number information of the converted three-dimensional ground penetrating radar data according to the converted three-dimensional ground penetrating radar data; Performing offset compensation on the frequency wavenumber information to obtain corrected frequency wavenumber domain data; The corrected frequency wavenumber domain data is subjected to inverse two-dimensional Fourier transform and converted into time-space domain processing to obtain the radar deviation data.

6. The method according to claim 1, characterized in that The ant nest imaging three-dimensional deep learning network model includes several three-dimensional convolution modules and a deep neural network structure. The three-dimensional convolution modules and the deep neural network structure form a U-shaped network structure, wherein the three-dimensional convolution module includes a three-dimensional convolution layer, a batch normalization layer and an activation function, and the deep neural network structure includes a data input module, an encoder, a decoder, a jump connection and a convolution output module.

7. The method according to claim 6, characterized in that The loss function of the ant nest imaging three-dimensional deep learning network model is specifically as follows: Loss=BCE+Focal BCE(p,y)=-[y·log(p)+(1-y)·log(1-p)] Focal(p t )−α t (1-p t ) 2 log(p t ) In the above formula, Loss represents the ant nest imaging 3D deep learning network model, BCE(·) represents the cross entropy loss function, Focal(·) represents the focal loss function, y represents the label, p represents the probability of predicting the positive class, and p t represents the modulation factor, α t Represents a constant.

8. The method according to claim 6, characterized in that The performing data imaging processing on the multi-channel radar data volume based on the ant nest imaging three-dimensional deep learning network model to obtain a three-dimensional ground penetrating radar ant nest imaging result includes: Inputting the multi-channel radar data volume into the ant nest imaging three-dimensional deep learning network model; Based on the three-dimensional convolution module of the ant nest imaging three-dimensional deep learning network model, the multi-channel radar data volume is automatically subjected to feature extraction processing to obtain multi-channel radar feature data; Based on the deep neural network structure of the ant nest imaging three-dimensional deep learning network model, data imaging processing is performed on the multi-channel radar feature data to obtain the three-dimensional ground penetrating radar ant nest imaging result.

9. The method according to claim 8, characterized in that The deep neural network structure based on the ant nest imaging three-dimensional deep learning network model performs data imaging processing on the multi-channel radar feature data to obtain the three-dimensional ground penetrating radar ant nest imaging result, including: Inputting the multi-channel radar feature data into the deep neural network structure of the ant nest imaging three-dimensional deep learning network model; Acquire the multi-channel radar feature data based on the data input module of the deep neural network structure; Based on the encoder of the deep neural network structure, the multi-channel radar feature data is subjected to feature space resolution reduction processing to obtain reduced multi-channel radar feature data; Based on the decoder of the deep neural network structure, performing spatial resolution recovery processing on the reduced multi-channel radar feature data to obtain restored multi-channel radar feature data; Based on the skip connection of the deep neural network structure, the reduced multi-channel radar feature data and the restored multi-channel radar feature data are fused and down-sampled to obtain fused multi-channel radar feature data; Based on the convolution output module of the deep neural network structure, the fused multi-channel radar feature data is classified into two categories to obtain the three-dimensional ground penetrating radar ant nest imaging result.

10. A three-dimensional imaging system of termite nests on dams based on ground penetrating radar, characterized in that: The system comprises: The first module is used to acquire the three-dimensional ground penetrating radar data of the target area and perform data preprocessing to obtain the preprocessed three-dimensional ground penetrating radar data; The second module is used to perform feature extraction and encryption processing on the pre-processed three-dimensional ground penetrating radar data to obtain a multi-channel radar data volume; The third module is used to build a three-dimensional deep learning network model for ant nest imaging based on a U-shaped network structure; The fourth module is used to perform data imaging processing on the multi-channel radar data volume based on the ant nest imaging three-dimensional deep learning network model to obtain a three-dimensional ground penetrating radar ant nest imaging result.

Images