Tree root system and soil dielectric constant ground penetrating radar inversion method

Abstract

The invention provides a tree root system and soil dielectric constant ground penetrating radar inversion method, and belongs to the field of electromagnetic inverse scattering imaging and geophysical exploration. The method comprises the following steps: acquiring an electrical parameter distribution image; acquiring a ground penetrating radar simulation echo image; constructing a medium edge extraction auxiliary task according to the preprocessed ground penetrating radar simulation echo image and a medium edge extraction auxiliary task training label, and constructing a dielectric constant inversion main task according to the preprocessed ground penetrating radar simulation echo image and a dielectric constant inversion main task training label; and carrying out joint training on the dielectric constant inversion main task and the medium edge extraction auxiliary task, and constructing a tree root system and layered non-uniform soil dielectric constant inversion model. The boundary of the electrical parameter distribution image is identified through the auxiliary task, the attention of the model to the non-uniform structure details in the ground penetrating radar simulation echo image is increased, and the capacity of the model for outputting the underground root system and the soil dielectric constant is improved.

Description

Technical Field

[0001] The present application relates to the field of electromagnetic backscattering imaging and geophysical exploration, and provides a ground penetrating radar inversion method for tree roots and soil dielectric constants. Background Art

[0002] The root system of trees plays an important role in the plant ecosystem. Accurate detection of tree roots helps to analyze and evaluate tree health risks and achieve reasonable management and control of fertilizers, water, and medicines in tree planting. Ground Penetrating Radar (GPR), as a non-invasive and nondestructive testing technology, is widely used to detect and image the characteristics of underground structures, strata, and buried objects. By emitting high-frequency electromagnetic pulses and receiving signals reflected from the interfaces of different materials, GPR can detect tree roots without destroying the underground medium. Ground penetrating radar inversion is an imaging technology that infers the position, morphology, and electrical parameters of the target medium or scatterer based on the quantitative echo information received by the receiver.

[0003] However, in the field of GPR, existing methods are unable to distinguish structural details within stratified and non-uniform soil layers when inverting underground tree roots and surrounding soil. Summary of the invention

[0004] In order to solve the above technical problems, the present application provides a method for ground penetrating radar inversion of tree roots and soil dielectric constants, the method comprising:

[0005] Constructing a layered non-uniform soil structure simulation model with varying texture, temperature and humidity, and obtaining an electrical parameter distribution image through the layered non-uniform soil structure simulation model with varying texture, temperature and humidity;

[0006] Constructing a physical model of electrical parameters of layered non-uniform soil, and obtaining a ground penetrating radar simulation echo image through the physical model of electrical parameters of layered non-uniform soil and the electrical parameter distribution image;

[0007] Performing a first preprocessing on the ground penetrating radar simulated echo image to obtain a preprocessed ground penetrating radar simulated echo image;

[0008] Performing a second preprocessing on the electrical parameter distribution image to obtain a training label for the main task of dielectric constant inversion, and annotating image boundaries and image non-boundaries on the electrical parameter distribution image after the second preprocessing to obtain a training label for the auxiliary task of medium edge extraction;

[0009] Constructing a dielectric edge extraction auxiliary task according to the preprocessed ground penetrating radar simulation echo image and the dielectric edge extraction auxiliary task training label, and constructing a dielectric constant inversion main task according to the preprocessed ground penetrating radar simulation echo image and the dielectric constant inversion main task training label;

[0010] The dielectric constant inversion main task and the medium edge extraction auxiliary task are jointly trained, and the medium edge extraction auxiliary task is used to identify the medium interface of the electrical parameter distribution image to improve the inversion effect of the dielectric constant inversion main task on the dielectric constant distribution, and to construct a dielectric constant inversion model for tree roots and layered non-uniform soil.

[0011] In one embodiment, the first preprocessing of the ground penetrating radar simulated echo image includes:

[0012] The ground penetrating radar simulated echo image is subjected to background removal, gain correction and normalization processing in sequence;

[0013] Scaling the normalized ground penetrating radar simulation echo image to a preset size to obtain a preprocessed ground penetrating radar simulation echo image;

[0014] The performing a second preprocessing on the electrical parameter distribution image comprises:

[0015] The electrical parameter distribution image is normalized, and the normalized electrical parameter distribution image is scaled to the preset size, where the preset size is 128×128.

[0016] In one embodiment, the constructing of the medium edge extraction auxiliary task according to the preprocessed ground penetrating radar simulated echo image and the medium edge extraction auxiliary task training label includes:

[0017] Constructing a medium edge extraction auxiliary task feature extraction module and a medium edge extraction auxiliary task loss function, wherein the medium edge extraction auxiliary task feature extraction module includes a first preset number of pyramid convolution feature extraction modules, and the first preset number is 5;

[0018] Output a first coding vector from the preprocessed ground penetrating radar simulation echo image through an encoder, input the coding vector into the medium edge extraction auxiliary task feature extraction module to obtain a first auxiliary task feature vector, and input the first auxiliary task feature vector into a decoder to obtain an output result of the medium edge extraction auxiliary task;

[0019] The output result of the medium edge extraction auxiliary task and the training label of the medium edge extraction auxiliary task are fitted based on the medium edge extraction auxiliary task loss function to construct the medium edge extraction auxiliary task.

[0020] In one embodiment, the step of constructing the dielectric constant inversion main task according to the preprocessed ground penetrating radar simulated echo image and the dielectric constant inversion main task training label includes:

[0021] Constructing a dielectric constant inversion main task feature extraction module and a dielectric constant inversion main task loss function, wherein the dielectric constant inversion main task feature extraction module includes a second preset number of pyramid convolution feature extraction modules and visual transformer feature extraction modules, and the second preset number is 4;

[0022] The preprocessed ground penetrating radar simulation echo image outputs a second encoding vector through the encoder, and the encoding vector obtains a second main task feature vector through the dielectric constant inversion main task feature extraction module, and the second main task feature vector is fused with the first auxiliary task feature vector, and the fused second main task feature vector is input into the decoder to obtain the dielectric constant inversion main task output result;

[0023] The output result of the dielectric constant inversion main task and the training label of the dielectric constant inversion main task are fitted based on the dielectric constant inversion main task loss function to construct the dielectric constant inversion main task.

[0024] In one embodiment, the joint training of the dielectric constant inversion main task and the medium edge extraction auxiliary task includes:

[0025] The first weight and the second weight corresponding to the dielectric constant inversion main task loss and the medium edge extraction auxiliary task loss are set respectively, wherein the first weight is 1 and the second weight is 0.1;

[0026] The dielectric constant inversion main task loss and the first weight are weighted summed with the medium edge extraction auxiliary task loss and the second weight to obtain the total loss of the tree root system and layered non-uniform soil dielectric constant inversion model.

[0027] In one implementation, the pyramid convolution feature extraction module includes:

[0028] Convolutions with different convolution kernel sizes are set to extract feature vectors, and the extracted feature vectors are spliced ​​through channels;

[0029] The feature vector after channel splicing is convolved, normalized and activated to obtain the first preliminary feature vector;

[0030] Adding an attention mechanism to the first preliminary feature vector to obtain a first intermediate feature vector;

[0031] The first preliminary feature vector is added to the first intermediate feature vector to obtain a first target feature vector.

[0032] In one embodiment, the visual transformer feature extraction module comprises:

[0033] Construct feature extraction submodules based on convolutions with different convolution kernel sizes;

[0034] The feature calculation submodule is composed of multi-head self-attention and multi-layer perceptron:

[0035] Extracting a basic feature vector through the feature extraction submodule, and inputting the basic feature vector into the feature calculation submodule to obtain a second preliminary feature vector;

[0036] The second preliminary feature vector is convolved and then an attention mechanism is performed to obtain a second intermediate feature vector, and the second preliminary feature vector and the second intermediate feature vector are added to obtain a second target feature vector.

[0037] The tree root system and soil dielectric constant ground penetrating radar inversion method provided by the present application above constructs a layered non-uniform soil structure simulation model with changes in texture, temperature and humidity, and obtains an electrical parameter distribution image through the layered non-uniform soil structure simulation model with changes in texture, temperature and humidity; constructs a layered non-uniform soil electrical parameter physical model, and obtains a ground penetrating radar simulation echo image through the layered non-uniform soil electrical parameter physical model and the electrical parameter distribution image; performs a first preprocessing on the ground penetrating radar simulation echo image to obtain a preprocessed ground penetrating radar simulation echo image; performs a second preprocessing on the electrical parameter distribution image to obtain a dielectric constant inversion main task training label, and performs a second preprocessing on the electrical parameter distribution image to obtain a dielectric constant inversion main task training label, and performs a second preprocessing on the electrical parameter distribution image to obtain a dielectric constant inversion main task training label, and a dielectric constant inversion main task training label. The image is labeled with image boundaries and image non-boundaries to obtain training labels for the medium edge extraction auxiliary task; the medium edge extraction auxiliary task is constructed according to the preprocessed ground penetrating radar simulation echo image and the medium edge extraction auxiliary task training labels, and the dielectric constant inversion main task is constructed according to the preprocessed ground penetrating radar simulation echo image and the dielectric constant inversion main task training labels; the dielectric constant inversion main task and the medium edge extraction auxiliary task are jointly trained, and the dielectric interface of the electrical parameter distribution image is identified through the medium edge extraction auxiliary task to improve the inversion effect of the dielectric constant inversion main task in the dielectric constant distribution, and to construct a tree root system and layered non-uniform soil dielectric constant inversion model. By identifying the boundaries of the electrical parameter distribution image through the auxiliary task, the model's attention to the non-uniform structural details in the ground penetrating radar simulation echo image is increased, and the model's ability to output the underground root system and soil dielectric constant is improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solution of the present application, the following is a brief introduction to the drawings required for use in the embodiments. It should be understood that the following drawings only illustrate certain embodiments of the present application and should not be regarded as limiting the scope of protection of the present application. In each of the drawings, similar components are numbered similarly.

[0039] Figure 1 A schematic diagram of a process of a ground penetrating radar inversion method for tree roots and soil dielectric constants provided by an embodiment of the present application is shown;

[0040] Figure 2 A schematic diagram of the structure of a physical model of electrical parameters of layered non-uniform soil provided in an embodiment of the present application is shown;

[0041] Figure 3 Another schematic diagram of the process of the ground penetrating radar inversion method for the tree root system and soil dielectric constant provided by the embodiment of the present application is shown;

[0042] Figure 4 A schematic diagram of the structure of the dielectric constant inversion model of tree roots and layered non-uniform soil provided in an embodiment of the present application is shown;

[0043] Figure 5 Another schematic diagram of a process of a method for inverting the dielectric constant of a tree root system and soil by ground penetrating radar provided in an embodiment of the present application is shown;

[0044] Figure 6 A schematic diagram showing the inversion visualization comparison results of the tree root system and layered non-uniform soil dielectric constant inversion model and other models under a simulation data set provided by an embodiment of the present application;

[0045] Figure 7 A schematic diagram of the structure of a pyramid convolution feature extraction module provided in an embodiment of the present application is shown;

[0046] Figure 8 A schematic diagram of the structure of a feature extraction module of a visual transformer provided in an embodiment of the present application is shown;

[0047] Fig. 9 A schematic diagram of a map of Xishan, Urumqi, the source of soil in the measured data provided in the embodiments of the present application;

[0048] Fig.10 A schematic diagram of the measured data acquisition configuration provided in an embodiment of the present application is shown;

[0049] Fig.11 A schematic diagram of the inversion visualization comparison results of the tree root system and layered non-uniform soil dielectric constant inversion model provided by an embodiment of the present application under a measured data set and other models is shown. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments.

[0051] The components of the embodiments of the present application generally described and shown in the drawings herein may be arranged and designed in various configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application.

[0052] Hereinafter, the terms "including", "having" and their cognates, which may be used in various embodiments of the present application, are intended only to indicate specific features, numbers, steps, operations, elements, components, or combinations of the foregoing items, and should not be understood as first excluding the existence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing items or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations of the foregoing items.

[0053] Furthermore, the terms “first”, “second”, “third”, etc. are merely used for distinguishing descriptions and are not to be understood as indicating or implying relative importance.

[0054] Unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meanings as those generally understood by those skilled in the art to which the various embodiments of the present application belong. The terms (such as those defined in generally used dictionaries) will be interpreted as having the same meanings as the contextual meanings in the relevant technical field and will not be interpreted as having idealized meanings or overly formal meanings unless clearly defined in the various embodiments of the present application.

[0055] Example 1

[0056] The embodiment of the present application provides a ground penetrating radar inversion method for tree roots and soil dielectric constants.

[0057] See also Figure 1 The ground penetrating radar inversion method for tree roots and soil dielectric constants includes steps S101-S106:

[0058] S101: constructing a layered non-uniform soil structure simulation model with varying texture, temperature and humidity, and acquiring an electrical parameter distribution image through the layered non-uniform soil structure simulation model with varying texture, temperature and humidity.

[0059] In this embodiment, soil, as a medium, has its own unique physical properties, which have a significant impact on the propagation of electromagnetic waves. Specifically, the dielectric constant of soil (i.e., the degree of influence of soil on the propagation speed of electromagnetic waves) is closely related to factors such as its composition, humidity, temperature, and porosity. Different types of soil (such as sand, loam, clay, etc.) have different dielectric constants due to differences in their composition and structure. In order to obtain enough dielectric constant distribution data, the simulation software gprMax is used to create a virtual underground scene model. The virtual underground scene includes different strata, objects and structures, that is, a layered non-uniform soil structure simulation model with changes in texture, temperature and humidity.

[0060] For example, in order to better adapt to complex soil conditions, a soil structure simulation model is constructed based on the root growth characteristics of walnut trees in Kashgar, Xinjiang and the specific conditions of the surrounding soil. First, the soil structure is divided into three levels. Each layer is designed according to a specific humidity range and temperature conditions, thus forming a non-uniform soil layering model, as follows: the first layer has a volumetric moisture content of 5% to 10% and a temperature of 20°C, the second layer has a volumetric moisture content of 10% to 13% and a temperature of 17°C, and the third layer has a volumetric moisture content of 13% to 15% and a temperature of 15°C. In addition to temperature and humidity, each layer will also select different soil textures to increase the diversity of data. Soil texture is one of the inherent basic properties of soil. It reflects the characteristics of soil particle composition and is the basic input parameter for conducting relevant numerical simulation studies. The international soil community has developed a set of standardized soil texture classification systems, of which the most widely adopted is the classification system proposed by the International Society of Soil Science (ISSS), referred to as the "International System". The specific classification rules are as follows: three types of soil particles are divided according to the effective diameter, namely clay particles (<0.002mm), silt particles (0.002mm~0.02mm), and sand particles (0.02mm~2mm). The soil is divided into 12 soil texture categories according to the proportion of the three types of soil particles. The three layers of soil in the simulation data are loamy sand, sandy loam, loam, silt loam, sandy clay loam, clay fill, silt clay loam, sandy clay, loamy clay, silt clay, clay, and heavy clay. A soil texture category is randomly selected for each layer, and in this texture category, 5 proportions of soil are randomly selected as the soil to fill this layer.

[0061] For example, after determining the particle ratio, temperature and humidity of the soil, it is necessary to further use the specific gravity empirical value of the soil particles and the soil bulk density, and calculate the electrical characteristic parameters of the soil according to the particle ratio, temperature and humidity of the soil and the specific gravity empirical value of the soil particles and the soil bulk density. These parameters include the dielectric constant and conductivity of the soil, which are input into the soil structure simulation model to accurately simulate the electromagnetic characteristics of the soil.

[0062] Exemplarily, the soil structure simulation model will also randomly select circular, semicircular, square and triangular objects in the second layer of the soil, and embed them after rotating them at any angle from 0 to 360 degrees. The relative dielectric constant of each object will be randomly assigned in the range of 2 to 32. This operation achieves the purpose of simulating the distribution of underground tree roots. By integrating the dielectric constant of the soil with the dielectric constant of the tree roots, an intuitive electrical parameter distribution image is generated, which can also be called a dielectric constant distribution image. The horizontal axis of the electrical parameter distribution image represents the horizontal distance along the surface, and the vertical axis directly represents the depth of the soil.

[0063] S102: constructing a physical model of electrical parameters of layered inhomogeneous soil, and acquiring a ground penetrating radar simulation echo image through the physical model of electrical parameters of layered inhomogeneous soil and the electrical parameter distribution image.

[0064] In this embodiment, the physical model of the electrical parameters of layered non-uniform soil is as follows: Figure 2 As shown in the figure, TX represents the transmitting antenna, whose function is to transmit high-frequency electromagnetic pulses to the underground medium in a direction; RX is the receiving antenna, which is responsible for capturing and recording the electromagnetic signals reflected back to the ground from the underground target (object or tree root system). The underground soil presents a layered structure due to different texture classifications, and there is a non-uniform detail structure because the different water content and soil particle ratio in the layer vary within a certain range. The layered non-uniform soil structure simulation model with changing texture, temperature and humidity includes a non-uniform detail structure, and the layered non-uniform soil electrical parameter physical model collects it to construct an image of the soil underground structure, that is, the ground penetrating radar simulation echo image.

[0065] Exemplarily, the way of scanning the electrical parameter distribution image of the physical model of the electrical parameters of the layered non-uniform soil is B-scan, which uses a Ricker wavelet (Ricker Wavelet) with an amplitude of 1A and a center frequency of 900MHz, i.e., a Ricker waveform. The time window is set to 20ns. The transmitting antenna and the receiving antenna are both located at a height of 0.005 meters from the ground, and the horizontal spacing between the two is fixed at 0.14 meters. At the beginning of the experiment, the transmitting antenna moves to the right from a position of 0.025 meters and transmits a signal, while the receiving antenna is at a position of 0.165 meters. The transmitting antenna and the receiving antenna slide synchronously. When the transmitting antenna moves to the rightmost position of 1.375 meters, the receiving antenna reaches a position of 1.515 meters accordingly. During the entire sliding process, an electromagnetic signal was emitted every 0.015 meters of advance, and a total of 90 channels of data were collected. The collected data are the reflection and scattering information of the electromagnetic signal by different underground soil geology and tree root simulations. Each data is mapped to a pixel value in the image to obtain a ground penetrating radar simulated echo image. The horizontal axis of the ground penetrating radar simulated echo image represents the horizontal distance along the surface, and the vertical axis represents time.

[0066] S103: performing a first preprocessing on the ground penetrating radar simulated echo image to obtain a preprocessed ground penetrating radar simulated echo image.

[0067] In this embodiment, after obtaining the ground penetrating radar simulated echo image of the soil, that is, the B-Scan image data, in order to improve the recognizability of the reflected wave, three operations of background removal, gain and normalization are used, and the normalized image is scaled to a size of 128×128 as the input of the subsequent deep learning model.

[0068] In one embodiment, the first preprocessing of the ground penetrating radar simulated echo image includes: performing background removal, gain correction and normalization on the ground penetrating radar simulated echo image in sequence; scaling the normalized ground penetrating radar simulated echo image to a preset size to obtain a preprocessed ground penetrating radar simulated echo image.

[0069] In this embodiment, since the direct wave usually generates almost the same signal on all channels, these signals include the direct wave and certain types of environmental noise. In order to eliminate the direct wave component in the echo signal, the following formula (1) is used: Perform background removal, where i represents the time step, j represents the data channel number, and x ij is the original B-Scan data, N is the total number of channels (i.e., the signals received by the receiving antenna at different positions), x i ′ j This is the result after background removal.

[0070] For example, when the ground penetrating radar detects underground, the electromagnetic signal will gradually attenuate as it propagates downward, resulting in a very small amplitude of the deep echo signal. In order to ensure that the echo signals at different depths can be clearly displayed, it is necessary to compensate for these rapidly decaying deep signals in the B-Scan image data. This process is called time gain, which aims to compensate for the energy loss caused by the radar wave re-propagation process, thereby enhancing the amplitude of the deep reflection wave. Correct gain adjustment can improve the ground penetrating radar's ability to identify underground targets and reduce the risk of misjudgment and missed judgment. The calculation method of time gain is formula (2): Among them, the sampling interval of the signal is Δt = 5.89664 × 10 -12 Seconds, x″ ij is the signal after gain correction.

[0071] Exemplarily, the signal data after gain correction is normalized by formula (3): Among them, x″′ ij is the normalized B-Scan data, min(x″ ij ) and max(x″ ij ) represent the minimum and maximum values ​​of the B-Scan signal data after gain correction. The normalized image is scaled to 128×128, and the preprocessed ground penetrating radar simulation echo image is used as the input of the subsequent deep learning model.

[0072] S104: performing a second preprocessing on the electrical parameter distribution image to obtain a training label for the main task of dielectric constant inversion, and annotating the image boundary and image non-boundary of the electrical parameter distribution image after the second preprocessing to obtain a training label for the auxiliary task of medium edge extraction.

[0073] In one embodiment, the second preprocessing of the electrical parameter distribution image includes:

[0074] The electrical parameter distribution image is normalized, and the normalized electrical parameter distribution image is scaled to the preset size, where the preset size is 128×128.

[0075] In this embodiment, the second preprocessing includes: normalizing the electrical parameter distribution image as well, and setting the width and height of the electrical parameter distribution image to 128 pixels, i.e. scaling it to 128×128 size, as the training label for the dielectric constant inversion main task. The image boundary and image non-boundary labeling of the electrical parameter distribution image includes: extracting the edge of the preprocessed electrical parameter distribution image through an edge detection algorithm, marking the detected boundary value as 1, and marking the other parts of the electrical parameter distribution image as 0, to obtain the medium edge extraction auxiliary task training label.

[0076] S105: constructing a dielectric edge extraction auxiliary task according to the preprocessed ground penetrating radar simulation echo image and the dielectric edge extraction auxiliary task training label, and constructing a dielectric constant inversion main task according to the preprocessed ground penetrating radar simulation echo image and the dielectric constant inversion main task training label.

[0077] In this embodiment, in the ground penetrating radar inversion method for the dielectric constant of tree roots and soil, in addition to the dielectric constant inversion main task constructed by preprocessing the ground penetrating radar simulation echo image and the dielectric constant inversion main task training label, a dielectric edge extraction auxiliary inversion task constructed by preprocessing the ground penetrating radar simulation echo image and the dielectric edge extraction auxiliary task training label is added. The dielectric edge extraction auxiliary task enhances the model's ability to recognize finer non-uniform structures inside the soil. At the same time, during model training, the feature vector output by the feature extraction module of the dielectric edge extraction auxiliary task is fused with the feature vector output by the feature extraction module of the dielectric constant inversion main task, thereby improving the inversion effect of the dielectric constant inversion main task on the dielectric constant distribution.

[0078] See also Figure 3 In one embodiment, the medium edge extraction auxiliary task is constructed according to the preprocessed ground penetrating radar simulated echo image and the medium edge extraction auxiliary task training label, including steps S301-S303:

[0079] S301: Construct a medium edge extraction auxiliary task feature extraction module and a medium edge extraction auxiliary task loss function, wherein the medium edge extraction auxiliary task feature extraction module includes a first preset number of pyramid convolution feature extraction modules, and the first preset number is 5.

[0080] In this embodiment, see Figure 4 The auxiliary task feature extraction module in the medium edge extraction auxiliary task consists of 5 pyramidal convolution feature extraction blocks (PyConvFEB). The medium edge extraction auxiliary task loss function uses the mean square error (MSE) to calculate the accuracy of the model inversion. The specific definition of the medium edge extraction auxiliary task loss function is as follows: Where B is the size of the training batch, Y is the boundary value and non-boundary value of the real electrical parameter distribution image, is the boundary value and non-boundary value of the predicted electrical parameter distribution image, and b is the number of the data in each batch.

[0081] S302: Output a first coding vector of the preprocessed ground penetrating radar simulation echo image through an encoder, input the coding vector into the medium edge extraction auxiliary task feature extraction module to obtain a first auxiliary task feature vector, and input the first auxiliary task feature vector into a decoder to obtain an output result of the medium edge extraction auxiliary task.

[0082] In this embodiment, the encoder in the medium edge extraction auxiliary task uses a 3×3 small convolution kernel plus batch normalization (BatchNorm, BN) and ReLU activation function (Rectified Linear Unit, ReLU), and the number of output channels is 64. The input feature map is encoded to obtain an encoding vector; the medium edge extraction auxiliary task feature extraction module is composed of 5 PyConvFEBs, and the output channels of the 5 modules are 64, 128, 128, 64, and 32 respectively. The dielectric constant inversion main task is given more information about the boundary of the electrical parameter distribution image by identifying the medium interface of the electrical parameter distribution image through the medium edge extraction auxiliary task. Finally, the structure of the decoder is a 3×3 small convolution kernel plus a Sigmoid activation function, which merges the channels into 1 channel, and maps the result between 0 and 1 through the activation function as the final output of the medium edge extraction auxiliary task, and compares it with the medium edge extraction auxiliary task training label to obtain the medium edge extraction auxiliary task loss.

[0083] S303: Fitting the output result of the medium edge extraction auxiliary task and the training label of the medium edge extraction auxiliary task based on the medium edge extraction auxiliary task loss function to construct a medium edge extraction auxiliary task.

[0084] In this embodiment, the output result of the medium edge extraction auxiliary task is between 0 and 1. The auxiliary task loss is calculated based on the constructed medium edge extraction auxiliary task training label and the medium edge extraction auxiliary task loss function calculation formula. The gradient of the medium edge extraction auxiliary task loss function is propagated from the output layer back to the input layer through the back propagation algorithm, and the medium edge extraction auxiliary task parameters are adjusted layer by layer to minimize the medium edge extraction auxiliary task loss, thereby constructing a medium edge extraction auxiliary task model.

[0085] See also Figure 5 In one embodiment, the process of constructing the dielectric constant inversion main task according to the preprocessed ground penetrating radar simulated echo image and the dielectric constant inversion main task training label includes S501-S503:

[0086] S501: Construct a dielectric constant inversion main task feature extraction module and a dielectric constant inversion main task loss function, wherein the dielectric constant inversion main task feature extraction module includes a second preset number of pyramid convolution feature extraction modules and visual transformer feature extraction modules, and the second preset number is 4.

[0087] In this example, please continue to refer to Figure 4 The dielectric constant inversion main task feature extraction module consists of 4 pyramid convolution feature extraction modules PyConvFEB and 1 visual transformer feature extraction module (ViT Feature ExtractionBlock, ViTFEB). The introduction of pyramid convolution (PyConv) and vision transformer (ViT) aims to improve the diversity of image feature extraction. The number of output channels of the dielectric constant inversion main task feature extraction module is 128, 256, 256, 128 and 64 respectively. The dielectric constant inversion main task loss function also uses the mean square error to calculate the accuracy of the model inversion. The specific definition of the dielectric constant inversion main task loss function is as follows: Where B is the size of the training batch, Y is the pixel value of the real electrical parameter distribution image, is the pixel value of the predicted electrical parameter distribution image, and b is the number of the data in each batch.

[0088] S502: The preprocessed ground penetrating radar simulation echo image outputs a second encoding vector through the encoder, and the encoding vector obtains a second main task feature vector through the dielectric constant inversion main task feature extraction module, and the second main task feature vector is fused with the first auxiliary task feature vector, and the fused second main task feature vector is input into the decoder to obtain the dielectric constant inversion main task output result.

[0089] In this embodiment, the encoder of the dielectric constant inversion main task is the same as the encoder of the medium edge extraction auxiliary task, both of which use a 3×3 small convolution kernel plus batch normalization and ReLU activation function, and the number of output channels is 64. After completing the encoding of the preprocessed ground penetrating radar simulation echo image, the encoded feature vector is input into the dielectric constant inversion main task feature extraction module for feature extraction, and the feature maps of the first, second, fourth, and fifth PyConvFEBs of the dielectric edge extraction auxiliary task feature extraction module are sequentially connected and fused with the first, second, fourth, and fifth PyConvFEBs of the corresponding dielectric constant inversion main task feature extraction module. The second main task feature vector obtained by the final fusion is input into the decoder. The structure of the decoder of the inversion part of the dielectric constant inversion main task is a 3×3 small convolution kernel, which merges the channels into 1 channel, adds a Sigmoid activation function, maps the result between 0 and 1, and serves as the final output result of the dielectric constant inversion main task.

[0090] S503: Fitting the output result of the dielectric constant inversion main task and the training label of the dielectric constant inversion main task based on the dielectric constant inversion main task loss function to construct the dielectric constant inversion main task.

[0091] In this embodiment, the edge features of the electrical parameter distribution image are integrated into the feature vector output by the dielectric constant inversion main task feature extraction module, and the fused second main task feature vector is input into the decoder to obtain the output result of the dielectric constant inversion main task, which is combined with the constructed dielectric constant inversion main task training label, and the sum of the pixel losses of each electrical parameter distribution image is calculated based on the dielectric constant inversion main task loss function calculation formula, that is, the dielectric constant inversion main task loss. Through the back propagation algorithm, the gradient of the dielectric constant inversion main task loss function is propagated from the output layer back to the input layer, and the auxiliary task parameters are adjusted layer by layer to minimize the dielectric constant inversion main task loss, thereby constructing the dielectric constant inversion main task.

[0092] S106: The dielectric constant inversion main task and the medium edge extraction auxiliary task are jointly trained, and the medium interface of the electrical parameter distribution image is identified through the medium edge extraction auxiliary task to improve the inversion effect of the dielectric constant inversion main task on the dielectric constant distribution, and to construct a dielectric constant inversion model of tree roots and layered non-uniform soil.

[0093] In this embodiment, by setting the weights of the loss functions of the dielectric constant inversion main task and the medium edge extraction auxiliary task, the importance of the two tasks is balanced, and finally the parameters are continuously adjusted to achieve the best simulation effect, that is, to minimize the overall loss function, and to construct a dielectric constant inversion model of tree roots and layered non-uniform soil.

[0094] For example, see Figure 6 , which is a comparison chart of the inversion results of various methods after experiments on the simulated data set of ground penetrating radar simulated echo images. In the figure, the traditional inversion method FWI is based on a given initial solution. This method takes a long time and has lower inversion accuracy than other deep learning methods. For deep learning methods, since the step size and convolution kernel size of the PINet deconvolution layer cannot be divided evenly, some parts of the output feature map are covered more times than other parts, resulting in black noise, making the evaluation index of PINet lower than that of other deep learning models. GPRInvNet and DMRF-UNet have poor inversion results in this complex simulation data set. They can neither accurately invert the relative dielectric constant value (i.e., the electrical parameter distribution image) nor accurately predict the scatterer and soil layer shape. While UNet, TransUNet, and EDMFEBs can roughly and accurately invert the shape, dielectric constant, and soil stratification of the scatterer, the tree root system and layered inhomogeneous soil dielectric constant inversion model (PyViNet) in this application can better invert the shape and dielectric constant of the scatterer. With the help of the auxiliary task of medium edge extraction, it can identify the soil stratification boundary and the structural details of the inhomogeneous medium more clearly than other methods.

[0095] In one embodiment, the dielectric constant inversion main task and the medium edge extraction auxiliary task are jointly trained, including: setting a first weight and a second weight corresponding to the dielectric constant inversion main task loss and the medium edge extraction auxiliary task loss, respectively, the first weight is 1, and the second weight is 0.1; weighted summing the dielectric constant inversion main task loss and the first weight with the medium edge extraction auxiliary task loss and the second weight to obtain the total loss of the tree root system and layered non-uniform soil dielectric constant inversion model.

[0096] In this embodiment, the formula Loss total =a×Loss1 mse +b×Loss2 mse Calculate the total loss of tree roots and layered inhomogeneous soil dielectric constant inversion model total , where a represents the dielectric constant inversion main task loss Loss1 mse , b represents the medium edge extraction auxiliary task loss Loss2 mse weights, preferably, the first weight is set to 1 and the second weight is set to 0.1.

[0097] In one embodiment, the pyramid convolution feature extraction module includes: setting convolutions of different convolution kernel sizes to extract feature vectors, and performing channel splicing on the extracted feature vectors; performing convolution, normalization and activation operations on the feature vectors after channel splicing to obtain a first preliminary feature vector; adding an attention mechanism to the first preliminary feature vector to obtain a first intermediate feature vector; adding the first preliminary feature vector to the first intermediate feature vector to obtain a first target feature vector.

[0098] In this embodiment, since the vertical axes of the GPR simulation echo image and the electrical parameter distribution image represent time and depth respectively, and the horizontal axes, i.e., image width, represent distance, the two cannot correspond directly on the vertical axis, which requires the model to extract regular information from a larger scale receptive field. Therefore, this application defines a convolution module named PyConvFEB that uses pyramid convolution PyConv to perform multi-scale feature extraction. Its structure is as follows Figure 7 As shown in the figure. In the figure, k represents the convolution kernel size, s represents the step size, p represents padding, and g represents grouping. PyConv can process input signals on convolution kernels of multiple scales. PyConv contains a kernel pyramid, in which each level contains different types of convolution kernels of different sizes and depths. These convolution kernels discover details at different levels in the feature map and can improve the recognition ability of the model. The information extracted by different types of convolution kernels is complementary. Kernels with smaller receptive fields can focus on details and capture partial information about smaller objects and objects, while larger receptive fields can extract more reliable details about contextual information. It can effectively capture more detailed and detailed information by parsing input data across multiple scales while keeping the computational cost unchanged.

[0099] Exemplarily, the PyConvFEB module specifically includes three parallel convolution operations: convolution using a 3×3 kernel, with a convolution group number of 1; convolution using a 5×5 kernel, with a convolution group number of 4; and convolution using a 7×7 kernel, with a convolution group number of 8. The input channels are evenly divided by the number of groups. This design can extract image feature vectors of different scales and different levels of detail through multi-scale convolution kernels and different group convolution configurations. By splicing these feature vectors of different scales together, the number of feature channels is adjusted to the target output channel number through a 3×3 convolution layer, and batch normalization (BN) and ReLU activation operations are performed to obtain the first preliminary feature vector. At the end of the module, the Convolutional Block Attention Module (CBAM) attention mechanism is introduced, the first preliminary feature vector is input, and the output first intermediate feature vector is added to the first preliminary feature vector to obtain the first target feature vector, further enhancing the feature representation.

[0100] In one embodiment, the visual transformer feature extraction module includes: constructing a feature extraction submodule based on convolutions of different convolution kernel sizes; forming a feature calculation submodule based on multi-head self-attention and a multi-layer perceptron: extracting a basic feature vector through the feature extraction submodule, and inputting the basic feature vector into the feature calculation submodule to obtain a second preliminary feature vector; convolving the second preliminary feature vector and then performing an attention mechanism to obtain a second intermediate feature vector, and approximately adding the second preliminary feature vector and the second intermediate feature vector to obtain a second target feature vector.

[0101] In this embodiment, ViTFEB is composed of a serial feature extraction submodule and a feature calculation submodule (ie, ViTBlock), wherein the feature extraction submodule performs a multi-scale feature extraction operation. The specific network structure is as follows: Figure 8 As shown in the figure. In the figure, k represents the convolution kernel size, s represents the step size, and p represents padding. ViTFEB uses a 7×7 convolution kernel for initial feature extraction, and uses two 3×3 convolution kernels in the subsequent layers to refine the features. Then, a 5×5 convolution kernel is used, and three 3×3 convolution layers are used to further extract features. The above serial multi-scale feature extraction transforms the 256-channel 128×128 feature map into a 256-channel 32×32 basic feature vector, which improves the computational efficiency of the subsequent ViTBlock.

[0102] Exemplarily, ViTBlock consists of 2 Transformer encoders, which contain Multihead Self-Attention (MSA) and Multilayer Perceptron (MLP) blocks, where the number of MSA heads is 16 and the MLP mapping dimension is 512. LayerNorm is normalized on the feature dimension of each sample. The input basic feature vector is divided into a series of 1×1 patches, and these patches are converted into vectors of fixed dimension. These vectors are then sent to the Transformer for processing, and these layers capture global context information during the process of processing the input sequence. The result of the calculation is the addition of the second preliminary feature vector and the basic feature vector, followed by a 6×6 deconvolution, three 3×3 convolutions and a 4×4 deconvolution to restore the feature map size, and the feature representation is further enhanced by the CBAM module to obtain the second intermediate feature vector, and the second preliminary feature vector and the second intermediate feature vector are approximately added to obtain the second target feature vector, which is output to the next pyramid convolution feature extraction module.

[0103] For example, the inversion experiment was conducted again using the measured data set measured in the laboratory. A GER900A ground penetrating radar was used with a center frequency of 900 MHz, a length of 30 cm, and a distance of 14 cm between the transmitting antenna and the receiving antenna. The ground penetrating radar was measured in an experimental box with absorbing materials. Fig. 9 This is a schematic diagram of the map of Urumqi West Mountain, the source of the soil. The soil is dry silt sand in Urumqi West Mountain (87°31'10.452" east longitude, 43°48'14.731" north latitude), with a soil depth of 0.6m, a length of 1m and a width of 0.8m. Fig.10 This is a schematic diagram of the actual data collection configuration in the experimental site. The measurement line track is 0.7m long, with a total of 50 data, 512 time sampling points for each data, and a sampling window of 10ns. There are three kinds of wooden objects buried underground in the experimental site. The diameters of the cylindrical wood blocks are 6 and 7cm, 15cm and 10cm long, and the material is soaked fir. The rectangular wooden board is 18cm×18cm×3cm and the material is basswood. The data of the measured wooden board includes 60 soaked and dry data, and the wooden board is placed horizontally and vertically. One is selected for a single measurement, and the burial depth is 10cm, 15cm, and 20cm. The data set contains a total of 190 data, which are divided into training set and test set at a ratio of 9:1.

[0104] Exemplarily, the simulation experiment gives the initial values ​​of the learnable parameters in the deep learning model, which can be extracted, saved as a pth weight file, and migrated to subsequent training. Due to the differences between the simulation experiment of this application and the actual measured site conditions, it is necessary to add 2000 groups of pre-training data consistent with the various parameters of the experimental site for pre-training, and load the weight file obtained by the first simulation experiment at the beginning of pre-training. After completing the pre-training, a new weight file is obtained, and the new weight file is loaded again on the measured data for training, and the learnable parameters in the deep learning model are fine-tuned so that the model can transfer the information learned on the simulation data to the field of measured data. Fig.11 The prediction results of some test set data are shown. The first and second columns are cylindrical wooden blocks, and the third, fourth, and fifth columns are rectangular wooden boards. The fourth column is relatively dry and close to the soil dielectric constant, so the inversion effect of each method is not ideal. In addition, PyViTNet more accurately inverts the shape, position, and relative dielectric constant of the wetted object than other methods, proving the versatility of the model.

[0105] The tree root system and soil dielectric constant ground penetrating radar inversion method provided in this embodiment constructs a layered non-uniform soil structure simulation model with texture, temperature and humidity changes, and obtains an electrical parameter distribution image through the layered non-uniform soil structure simulation model with texture, temperature and humidity changes; constructs a layered non-uniform soil electrical parameter physical model, and obtains a ground penetrating radar simulation echo image through the layered non-uniform soil electrical parameter physical model and the electrical parameter distribution image; performs a first preprocessing on the ground penetrating radar simulation echo image to obtain a preprocessed ground penetrating radar simulation echo image; performs a second preprocessing on the electrical parameter distribution image to obtain a dielectric constant inversion main task training label, and performs a second preprocessing on the electrical parameter distribution image to obtain a dielectric constant inversion main task training label, and performs a second preprocessing on the electrical parameter distribution image to obtain a dielectric constant inversion main task training label, and performs a second preprocessing on the electrical parameter distribution image to obtain a dielectric constant inversion main task training label, and obtains ... The image is labeled with image boundaries and image non-boundaries to obtain training labels for the medium edge extraction auxiliary task; the medium edge extraction auxiliary task is constructed according to the preprocessed ground penetrating radar simulation echo image and the medium edge extraction auxiliary task training labels, and the dielectric constant inversion main task is constructed according to the preprocessed ground penetrating radar simulation echo image and the dielectric constant inversion main task training labels; the dielectric constant inversion main task and the medium edge extraction auxiliary task are jointly trained, and the dielectric interface of the electrical parameter distribution image is identified through the medium edge extraction auxiliary task to improve the inversion effect of the dielectric constant inversion main task in the dielectric constant distribution, and to construct a tree root system and layered non-uniform soil dielectric constant inversion model. By identifying the boundaries of the electrical parameter distribution image through the auxiliary task, the model's attention to the non-uniform structural details in the ground penetrating radar simulation echo image is increased, and the model's ability to output the underground root system and soil dielectric constant is improved.

[0106] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or terminal including the element.

[0107] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for a terminal (which can be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0108] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present application, ordinary technicians in this field can also make many forms without departing from the purpose of the present application and the scope of protection of the claims, all of which are within the protection of the present application.

Claims

1. A method for inverting the dielectric constant of tree roots and soil by ground penetrating radar, characterized in that: The method comprises: Constructing a layered non-uniform soil structure simulation model with varying texture, temperature and humidity, and obtaining an electrical parameter distribution image through the layered non-uniform soil structure simulation model with varying texture, temperature and humidity; Constructing a physical model of electrical parameters of layered non-uniform soil, and obtaining a ground penetrating radar simulation echo image through the physical model of electrical parameters of layered non-uniform soil and the electrical parameter distribution image; Performing a first preprocessing on the ground penetrating radar simulated echo image to obtain a preprocessed ground penetrating radar simulated echo image; Performing a second preprocessing on the electrical parameter distribution image to obtain a training label for the main task of dielectric constant inversion, and annotating image boundaries and image non-boundaries on the electrical parameter distribution image after the second preprocessing to obtain a training label for the auxiliary task of medium edge extraction; Constructing a dielectric edge extraction auxiliary task according to the preprocessed ground penetrating radar simulation echo image and the dielectric edge extraction auxiliary task training label, and constructing a dielectric constant inversion main task according to the preprocessed ground penetrating radar simulation echo image and the dielectric constant inversion main task training label; The dielectric constant inversion main task and the medium edge extraction auxiliary task are jointly trained, and the medium edge extraction auxiliary task is used to identify the medium interface of the electrical parameter distribution image to improve the inversion effect of the dielectric constant inversion main task on the dielectric constant distribution, and to construct a dielectric constant inversion model for tree roots and layered non-uniform soil.

2. The method according to claim 1, characterized in that The first preprocessing of the ground penetrating radar simulated echo image includes: The ground penetrating radar simulated echo image is subjected to background removal, gain correction and normalization processing in sequence; Scaling the normalized ground penetrating radar simulation echo image to a preset size to obtain a preprocessed ground penetrating radar simulation echo image; The performing a second preprocessing on the electrical parameter distribution image comprises: The electrical parameter distribution image is normalized, and the normalized electrical parameter distribution image is scaled to the preset size, where the preset size is 128×128.

3. The method according to claim 1, characterized in that: The step of constructing the medium edge extraction auxiliary task according to the preprocessed ground penetrating radar simulated echo image and the medium edge extraction auxiliary task training label comprises: Constructing a medium edge extraction auxiliary task feature extraction module and a medium edge extraction auxiliary task loss function, wherein the medium edge extraction auxiliary task feature extraction module includes a first preset number of pyramid convolution feature extraction modules, and the first preset number is 5; Output a first coding vector from the preprocessed ground penetrating radar simulation echo image through an encoder, input the coding vector into the medium edge extraction auxiliary task feature extraction module to obtain a first auxiliary task feature vector, and input the first auxiliary task feature vector into a decoder to obtain an output result of the medium edge extraction auxiliary task; The output result of the medium edge extraction auxiliary task and the training label of the medium edge extraction auxiliary task are fitted based on the medium edge extraction auxiliary task loss function to construct the medium edge extraction auxiliary task.

4. The method according to claim 3, characterized in that The step of constructing the dielectric constant inversion main task according to the preprocessed ground penetrating radar simulated echo image and the dielectric constant inversion main task training label comprises: Constructing a dielectric constant inversion main task feature extraction module and a dielectric constant inversion main task loss function, wherein the dielectric constant inversion main task feature extraction module includes a second preset number of pyramid convolution feature extraction modules and visual transformer feature extraction modules, and the second preset number is 4; The preprocessed ground penetrating radar simulation echo image outputs a second encoding vector through the encoder, and the encoding vector obtains a second main task feature vector through the dielectric constant inversion main task feature extraction module, and the second main task feature vector is fused with the first auxiliary task feature vector, and the fused second main task feature vector is input into the decoder to obtain the dielectric constant inversion main task output result; The output result of the dielectric constant inversion main task and the training label of the dielectric constant inversion main task are fitted based on the dielectric constant inversion main task loss function to construct the dielectric constant inversion main task.

5. The method according to claim 1, characterized in that The joint training of the dielectric constant inversion main task and the medium edge extraction auxiliary task includes: The first weight and the second weight corresponding to the dielectric constant inversion main task loss and the medium edge extraction auxiliary task loss are set respectively, wherein the first weight is 1 and the second weight is 0.1; The dielectric constant inversion main task loss and the first weight are weighted summed with the medium edge extraction auxiliary task loss and the second weight to obtain the total loss of the tree root system and layered non-uniform soil dielectric constant inversion model.

6. The method according to claim 3, characterized in that The pyramid convolution feature extraction module includes: Convolutions with different convolution kernel sizes are set to extract feature vectors, and the extracted feature vectors are spliced ​​through channels; The feature vector after channel splicing is convolved, normalized and activated to obtain the first preliminary feature vector; Adding an attention mechanism to the first preliminary feature vector to obtain a first intermediate feature vector; The first preliminary feature vector is added to the first intermediate feature vector to obtain a first target feature vector.

7. The method according to claim 4, characterized in that The visual transformer feature extraction module comprises: Construct feature extraction submodules based on convolutions with different convolution kernel sizes; The feature calculation submodule is composed of multi-head self-attention and multi-layer perceptron: Extracting a basic feature vector through the feature extraction submodule, and inputting the basic feature vector into the feature calculation submodule to obtain a second preliminary feature vector; The second preliminary feature vector is convolved and then an attention mechanism is performed to obtain a second intermediate feature vector, and the second preliminary feature vector and the second intermediate feature vector are added to obtain a second target feature vector.

Images