Ground penetrating radar based metal reflection decoupling and three-dimensional inversion enhancement method

Abstract

The application discloses a metal reflection decoupling and three-dimensional inversion enhancement method based on a ground penetrating radar, first, original radar echo data collected by the ground penetrating radar is acquired, and preprocessing is performed on the original radar echo data to obtain a standardized C-scan image; then the standardized C-scan image is input into a pre-constructed metal reflection decoupling module; the metal reflection decoupling module decouples metal reflection and non-metal reflection in the standardized C-scan image according to a pre-constructed metal reflection physical characteristic library by utilizing the electromagnetic response characteristic difference between the metal target and the non-metal target, and outputs a metal reflection mask, a pure radar image and a metal area uncertainty image; finally, the metal reflection mask, the pure radar image and the metal area uncertainty image are input into a pre-constructed three-dimensional inversion network, and through multi-scale feature aggregation and voxel-level prediction, a three-dimensional distribution of underground dielectric constant is output. The application realizes accurate separation and effective suppression of the metal reflection, and significantly improves the three-dimensional inversion imaging quality of the underground structure.

Description

Technical Field

[0001] This invention relates to the field of ground-penetrating radar signal processing technology, and in particular to a method for decoupling and three-dimensional inversion enhancement of metal reflections based on ground-penetrating radar. Background Technology

[0002] Ground-penetrating radar (GPR), as a highly efficient non-destructive testing technology, detects and images underground structures by emitting high-frequency electromagnetic waves and receiving reflected signals. It has wide applications in road engineering, municipal pipelines, and tunnel construction. In actual engineering environments, underground structures typically include reinforced concrete, metal pipes, manholes, and other metal structures. Due to the extremely high dielectric constant and conductivity of metals, they strongly reflect electromagnetic waves, forming typical interference characteristics such as saturated strong echoes, multiple wave tails, and shadow areas. These metal reflections not only obscure effective information from the vicinity of the metal and disrupt the overall time-frequency structure of the radar signal, but also severely interfere with the accurate identification and inversion imaging of weakly reflective targets such as underground cavities and water-rich bodies, thus limiting the detection accuracy and application effectiveness of GPR.

[0003] Existing technologies for attenuating metallic reflections mainly fall into two categories: physical methods and signal processing methods. Physical methods include adding absorbing materials to the antenna front end and adding antenna shielding slats. While these can reduce the impact of metallic reflections to some extent, they cannot distinguish between metallic reflections and weak non-metallic reflections, resulting in the attenuation of weakly reflecting target information and the inability to recover the signal from the shadow area formed by the metallic obstruction. Traditional signal processing methods include high-pass or band-pass filtering, wavelet thresholding denoising, and direct wave removal. Although these methods can filter or decompose signals through mathematical transformations, their signal discrimination ability is limited. They can only achieve overall noise reduction and are unable to accurately separate the metallic reflection component. Furthermore, while attenuating metallic reflections, they can easily destroy effective target information. Therefore, how to achieve accurate separation of metallic reflections and effective reconstruction of shadow area information, while maintaining the signal integrity of weakly reflecting targets, has become a pressing technical problem to be solved in this field. Summary of the Invention

[0004] The purpose of this invention is to provide a method for decoupling and enhancing three-dimensional inversion of metal reflections based on ground-penetrating radar. By constructing a metal reflection decoupling module guided by physical priors and a three-dimensional inversion network for enhancing metal reflections, and combining a multi-task joint optimization strategy, the method achieves selective separation and weakening of metal reflections, and improves the quality of three-dimensional inversion imaging of underground structures. This solves the technical problems of existing technologies, such as the inability to accurately separate metal reflections from weak reflection signals, the difficulty in maintaining the integrity of weak reflection information while suppressing metal interference, and the inability to effectively reconstruct information of shadow areas in metal-obscured regions.

[0005] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: This invention provides a method for decoupling and three-dimensional inversion enhancement of metal reflection based on ground penetrating radar, which includes: acquiring raw radar echo data collected by ground penetrating radar in the field, and preprocessing the raw radar echo data to obtain a standardized C-scan image. The standardized C-scan image is input into a pre-built metal reflection decoupling module. The metal reflection decoupling module decouples the metal reflection and non-metal reflection in the standardized C-scan image based on the standardized C-scan image and the pre-built metal reflection physical feature library, and utilizes the difference in electromagnetic response characteristics between metal targets and non-metal targets to output a metal reflection mask, a clean radar image, and a metal region uncertainty map. The metal reflection mask, the pure radar image, and the metal region uncertainty map are input into a pre-constructed three-dimensional inversion network. Through multi-scale feature aggregation and voxel-level prediction, the three-dimensional distribution of the underground dielectric constant is output.

[0006] In one embodiment of the present invention, the step of acquiring raw radar echo data collected on-site by ground-penetrating radar and preprocessing the raw radar echo data to obtain a standardized C-scan image includes: Acquire A-scan data generated by ground-penetrating radar scanning point by point along the survey line; Multiple A-scan data are spliced ​​together along the survey line direction to obtain B-scan data; Multiple B-scan data are spatially combined and interpolated to obtain C-scan data, which is a three-dimensional data volume. The C-scan data is subjected to depth slicing, amplitude normalization, and spatial size adjustment to obtain the standardized C-scan image.

[0007] In one embodiment of the present invention, the metal reflection physical feature library is constructed in the following manner: Multiple radar echo samples are acquired, and each radar echo sample corresponds to preset metal parameters and environmental parameters; Each radar echo sample is subjected to feature analysis to extract the corresponding set of physical features of metal reflection. The set of physical features of metal reflection includes at least the peak amplitude, the proportion of high-frequency energy, the number of multiple peak values, the trail length, and the amplitude of local phase perturbation. The physical feature set of metal reflection for each radar echo sample is associated with and stored with the corresponding metal parameters and environmental parameters to form the physical feature library of metal reflection.

[0008] In one embodiment of the present invention, the radar echo sample is generated by an electromagnetic forward modeling tool or obtained by ground penetrating radar in-situ detection.

[0009] In one embodiment of the present invention, the metal reflection decoupling module includes a physical prior guidance attention unit, a metal reflection mask prediction unit, and a pure reflection recovery unit; The physical prior-guided attention unit is used to determine attention weights based on the physical feature library of metal reflection features, and to perform weighted enhancement on the input standardized C-scan image through the synergistic effect of channel attention and spatial attention to generate an enhanced feature map; The metal reflection mask prediction unit is used to perform pixel-level metal reflection probability prediction on the enhanced feature map output by the physical prior guided attention unit to obtain a metal reflection probability map, and to perform threshold segmentation on the metal reflection probability map to generate the metal reflection mask. The pure reflection recovery unit is used to perform channel stitching of the standardized C-scan image and the metal reflection mask output by the metal reflection mask prediction unit, and to infer and complete the weak reflection structure of the metal-shaded area through residual learning and spatial correlation modeling to generate the pure radar image.

[0010] In one embodiment of the present invention, the metal reflection decoupling module further includes an uncertainty map generation unit, which is used to evaluate the uncertainty of the metal reflection probability map output by the metal reflection mask prediction unit to generate the metal region uncertainty map.

[0011] In one embodiment of the present invention, the uncertainty assessment includes local statistical feature analysis, uncertainty estimation based on Monte Carlo random inactivation inference, or uncertainty assessment based on multi-model ensemble learning.

[0012] In one embodiment of the present invention, the step of inputting the metal reflection mask, the pure radar image, and the metal region uncertainty map into a pre-constructed three-dimensional inversion network, and outputting the three-dimensional distribution of the underground dielectric constant through multi-scale feature aggregation and voxel-level prediction, includes: The metal reflective mask, the pure radar image, and the metal region uncertainty map are stitched together along the channel dimension to form a multi-channel three-dimensional feature tensor. The multi-channel three-dimensional feature tensor is input into a pre-constructed three-dimensional inversion network, and through multi-scale feature aggregation and voxel-level prediction, the three-dimensional distribution of the underground dielectric constant is output.

[0013] In one embodiment of the present invention, the three-dimensional inversion network includes a multi-scale feature aggregation module, an inversion weight adjustment module, and a voxel-level prediction head; The multi-scale feature aggregation module is used to extract multi-scale features from the multi-channel three-dimensional feature tensor using multiple three-dimensional convolutional kernels of different sizes, and to aggregate the extracted multi-scale features using a feature fusion gating unit to generate aggregated features. The inversion weight adjustment module is used to dynamically adjust the loss weight of the metal reflection region according to the metal reflection mask, and to assign feature enhancement weights to non-metal reflection regions with uncertainty values ​​below a preset threshold according to the metal region uncertainty map; The voxel-level prediction head is used to output the predicted value of the underground dielectric constant on a voxel-by-voxel basis in order to construct a three-dimensional distribution of the underground dielectric constant.

[0014] In one embodiment of the present invention, the metal reflection decoupling module and the three-dimensional inversion network are optimized by joint training. The network parameters are iteratively updated by the backpropagation algorithm to minimize the multi-task loss function, which includes metal reflection mask prediction loss, denoising recovery loss, three-dimensional inversion reconstruction loss and physical consistency loss.

[0015] As described above, the present invention provides a method for decoupling and enhancing metal reflection based on ground-penetrating radar (GPR). First, raw radar echo data collected in-situ by GPR is acquired and preprocessed to obtain a standardized C-scan image. Then, the standardized C-scan image is input into a pre-built metal reflection decoupling module. Based on the standardized C-scan image and a pre-built metal reflection physical feature library, the module decouples metal and non-metal reflections in the standardized C-scan image using the differences in electromagnetic response characteristics between metal and non-metal targets, outputting a metal reflection mask, a clean radar image, and a metal region uncertainty map. Finally, the metal reflection mask, the clean radar image, and the metal region uncertainty map are input into a pre-built three-dimensional inversion network. Through multi-scale feature aggregation and voxel-level prediction, the three-dimensional distribution of the underground dielectric constant is output. This invention, based on the physical characteristics of metal reflection, precisely separates the metal reflection component from the original radar echo signal through a metal reflection decoupling module. This fundamentally solves the problem of confusion between strong metal reflection and effective signal, achieving selective attenuation of metal reflection rather than simple filtering. This preserves the signal characteristics of weakly reflective targets such as underground cavities, water-rich bodies, and soil layer variations, significantly improving the detection rate of small targets. Simultaneously, through residual learning and spatial correlation modeling of the pure reflection recovery unit, the weak reflection structure in metal-obscured areas is inferred and completed, filling detection blind spots that traditional methods cannot cover, thus improving the integrity of the inversion imaging. Furthermore, diverse metal reflection datasets are generated through electromagnetic forward modeling, enabling the model to adapt to complex metal environments. Reflection decoupling and 3D inversion are integrated into the same network framework for multi-task joint optimization, achieving synergistic effects and ultimately reducing the inversion error of underground dielectric constant, making target boundaries clearer and significantly improving overall imaging quality. Of course, any product implementing this invention does not necessarily need to achieve all the advantages described above simultaneously. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of a method for decoupling metal reflection and enhancing three-dimensional inversion based on ground-penetrating radar, provided as an exemplary embodiment of this application.

[0018] Figure 2This is a system block diagram of a ground-penetrating radar-based metal reflection decoupling and three-dimensional inversion enhancement method provided for an exemplary embodiment of this application.

[0019] Figure 3 A schematic diagram of the structure of a physically priori guided metal reflection decoupling module provided for an exemplary embodiment of this application.

[0020] Figure 4 A flowchart illustrating multi-task joint optimization provided for an exemplary embodiment of this application. Detailed Implementation

[0021] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0022] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0023] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, publicly known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0024] To address the technical challenges of accurately separating metallic reflections from weak reflections, maintaining the integrity of weak reflection information while suppressing metallic interference, and effectively reconstructing information from shadowed areas obscured by metal in existing technologies, this invention proposes a ground-penetrating radar-based method for decoupling metallic reflections and enhancing 3D inversion. This method is applicable to various engineering scenarios, including road void assessment, underground pipeline surveys, tunnel lining defect identification, and geotechnical investigation. By constructing a physically prior-guided metallic reflection decoupling module and a demetallization-enhanced 3D inversion network, combined with a multi-task joint optimization strategy, this method achieves accurate separation and effective suppression of metallic reflections, thereby significantly improving the 3D inversion imaging quality of underground structures.

[0025] Please see Figure 1and Figure 2 As shown in an exemplary embodiment of this application, the metal reflection decoupling and three-dimensional inversion enhancement method based on ground-penetrating radar includes the following steps: S100: Acquire raw radar echo data collected by ground penetrating radar on site, and preprocess the raw radar echo data to obtain a standardized C-scan image. S200: The standardized C-scan image is input to a pre-built metal reflection decoupling module. The metal reflection decoupling module decouples the metal reflection and non-metal reflection in the standardized C-scan image based on the standardized C-scan image and the pre-built metal reflection physical feature library, and utilizes the difference in electromagnetic response characteristics between metal targets and non-metal targets to output a metal reflection mask, a clean radar image, and a metal region uncertainty map. S300: The metal reflection mask, the pure radar image, and the metal region uncertainty map are input into a pre-constructed three-dimensional inversion network. Through multi-scale feature aggregation and voxel-level prediction, the three-dimensional distribution of the underground dielectric constant is output.

[0026] The steps in the above-mentioned method for decoupling metal reflection and three-dimensional inversion enhancement based on ground-penetrating radar will be discussed in detail below.

[0027] It should be noted that before formally executing step S100, the metal reflection physical feature library, the metal reflection decoupling module, and the three-dimensional inversion network need to be designed and constructed in advance.

[0028] In an exemplary embodiment of this application, the metal reflection physical feature library is constructed as follows: Multiple radar echo samples are acquired, each corresponding to preset metal parameters and environmental parameters; feature analysis is performed on each radar echo sample to extract a corresponding metal reflection physical feature set, which includes at least amplitude peak value, high-frequency energy ratio, number of multiple peak values, tail length, and local phase perturbation amplitude; the metal reflection physical feature set of each radar echo sample is associated with and stored with the corresponding metal parameters and environmental parameters to form the metal reflection physical feature library. The metal parameters include metal material, metal geometry, metal size specifications, and metal burial depth; the environmental parameters include soil moisture content, soil dry density, and soil mineral composition. Furthermore, the radar echo samples are generated using electromagnetic simulation tools or obtained through on-site detection and collection using ground-penetrating radar.

[0029] Specifically, in this embodiment, an electromagnetic forward modeling simulation tool based on the finite-difference time-domain method is used to generate radar echo samples. By setting simulation parameters such as spatial discrete interval, time step, and center frequency of the transmitting antenna, radar echo samples covering various metal materials, geometries, sizes, burial depths, and soil dielectric properties are generated. The metal materials include, but are not limited to, high-conductivity metals such as steel, cast iron, and stainless steel. The geometries include, but are not limited to, tubular structures, plate structures, block structures, and disk structures. The burial depth and soil dielectric properties are combined and set according to actual engineering application scenarios. The soil dielectric properties are parametrically modeled based on soil moisture content, soil dry density, and soil mineral composition to simulate the propagation and attenuation characteristics of electromagnetic waves under different soil environments, thereby constructing a simulated radar echo sample library.

[0030] It is understood that, in another embodiment, radar echo samples can also be obtained through on-site measurements of existing road engineering projects. Detection targets include, but are not limited to, underground steel mesh, metal pipes, manhole covers, and other metal structures, while artificial cavities and water-rich bodies are also deployed as verification objects. During data acquisition, ground-penetrating radar is used to acquire raw time-domain A-scan echo signals, and continuous measurements are taken along the survey line to obtain two-dimensional B-scan data. In the data preprocessing stage, mean filtering is first used to remove background noise and inherent system reflections to enhance the reflection response of underground targets. Then, time zero-point correction is used to eliminate interference from direct surface waves and near-surface reflections. Finally, interpolation is used to map the discrete sampled data to a regular grid to generate C-scan images. Based on manual annotation and borehole verification results, the spatial location and attributes of metal targets are determined, thereby constructing a measured radar echo sample library.

[0031] Next, feature extraction is performed on the radar echo samples in the simulated radar echo sample library or the measured radar echo sample library. The radar echo samples are represented in the form of C-scan images, that is, each spatial location corresponds to one time series signal. The feature extraction process includes: extracting the maximum amplitude value of the time series signal at each spatial location to obtain the amplitude peak value feature; using wavelet transform to decompose the time series signal at each spatial location into different frequency bands, and calculating the proportion of high-frequency component energy to the total energy to obtain the high-frequency energy proportion feature; detecting the number of multiple peaks after a strong reflection signal using the threshold method to obtain the multiple peak number feature; calculating the time length from the moment of the strong reflection peak until the amplitude decays to 10% of the peak value as the tail length to obtain the tail length feature; using Hilbert transform to calculate the instantaneous phase of the time series signal at each spatial location, and calculating the phase standard deviation through a sliding window to obtain the local phase perturbation amplitude feature. The amplitude peak value, high-frequency energy ratio, multiple peak value, tail length, and local phase perturbation amplitude features extracted from each radar echo sample are associated and stored with corresponding metal parameters and environmental parameters to construct a metal reflection physical feature library. It should be noted that in this embodiment, each radar echo sample in the metal reflection physical feature library is stored in the form of a multi-channel feature tensor with a feature dimension of 256×256×5. Here, 256×256 corresponds to the spatial grid size of the C-scan image, and 5 corresponds to the five feature channels extracted: amplitude peak value, high-frequency energy ratio, multiple peak value, tail length, and local phase perturbation amplitude. Metal parameters and environmental parameters serve as label information and index identifiers for the radar echo samples and do not participate in the construction of the multi-channel feature tensor. An indexing mechanism is established based on these parameters to facilitate subsequent access to the metal reflection physical feature library by the metal reflection decoupling module.

[0032] First, step S100 is executed, which involves acquiring the raw radar echo data collected on-site by the ground penetrating radar and preprocessing the raw radar echo data to obtain a standardized C-scan image.

[0033] In an exemplary embodiment of this application, step S100 further includes the following steps: S110: Acquire A-scan data generated by ground-penetrating radar scanning point by point along the survey line; S120: The multiple A-scan data are spliced ​​along the survey line direction to obtain B-scan data; S130: Spatial combination and interpolation processing are performed on multiple B-scan data to obtain C-scan data, wherein the C-scan data is a three-dimensional data volume; S140: Perform depth slicing, amplitude normalization, and spatial size adjustment on the C-scan data to obtain the standardized C-scan image.

[0034] Specifically, in step S110, the ground-penetrating radar transmitting antenna transmits high-frequency electromagnetic pulses into the ground, and the receiving antenna receives the echo signals reflected from the underground medium interface and the target body. After analog-to-digital conversion, discrete A-scan data is formed. In step S120, multiple A-scan data collected at equal intervals along the survey line are arranged in the order of the measurement points to form B-scan data. In step S130, the B-scan data obtained from multiple parallel survey lines are three-dimensionally gridded according to their spatial position relationship, and interpolation is performed on the missing positions to form a regular grid C-scan three-dimensional data volume. In step S140, slice data within the corresponding depth range is extracted according to the target detection depth, the amplitude is normalized to eliminate amplitude differences between different measurement points, and the spatial size is unified to a preset size through interpolation or cropping, finally obtaining a standardized C-scan image that meets the input requirements of subsequent modules.

[0035] Next, step S200 is executed, which involves inputting the standardized C-scan image into a pre-built metal reflection decoupling module. The metal reflection decoupling module uses the difference in electromagnetic response characteristics between metal and non-metal targets based on the standardized C-scan image and the pre-built metal reflection physical feature library to decouple the metal reflection and non-metal reflection in the standardized C-scan image, so as to output a metal reflection mask, a clean radar image, and a metal region uncertainty map.

[0036] It should be noted that, in this embodiment, the metal reflection decoupling module is a neural network model pre-built and trained based on a deep learning framework, used to achieve adaptive decoupling of metal reflection and non-metal reflection.

[0037] Please see Figure 3 As shown, in an exemplary embodiment of this application, the metal reflection decoupling module includes a physical prior guided attention unit, a metal reflection mask prediction unit, and a pure reflection recovery unit; The physical prior-guided attention unit is used to determine attention weights based on the physical feature library of metal reflection features, and to perform weighted enhancement on the input standardized C-scan image through the synergistic effect of channel attention and spatial attention to generate an enhanced feature map; The metal reflection mask prediction unit is used to perform pixel-level metal reflection probability prediction on the enhanced feature map output by the physical prior guided attention unit to obtain a metal reflection probability map, and to perform threshold segmentation on the metal reflection probability map to generate the metal reflection mask. The pure reflection recovery unit is used to perform channel stitching of the standardized C-scan image and the metal reflection mask output by the metal reflection mask prediction unit, and to infer and complete the weak reflection structure of the metal-shaded area through residual learning and spatial correlation modeling to generate the pure radar image.

[0038] Please continue reading. Figure 3 As shown, in an exemplary embodiment of this application, the metal reflection decoupling module further includes an uncertainty map generation unit, which is used to perform uncertainty evaluation on the metal reflection probability map output by the metal reflection mask prediction unit to generate the metal region uncertainty map.

[0039] Specifically, in this embodiment, the physical prior-guided attention unit, guided by the physical feature library of metal reflection, determines attention weights using prior knowledge of the electromagnetic response of the metal target. Then, through the synergistic effect of channel attention and spatial attention, it enhances the feature response of the metal reflection region in the standardized C-scan image while suppressing signal interference from non-metallic reflection regions, thereby achieving preliminary localization of metal features. Specifically, firstly, a 3×3 convolutional layer is used to perform feature mapping on the input standardized C-scan image, adjusting its channel number to a preset dimension to obtain an intermediate feature map. Subsequently, the intermediate feature map is fed into the channel attention branch and the spatial attention branch respectively. In the channel attention branch, global average pooling is used to compress the intermediate feature map in the spatial dimension, obtaining a feature vector with the same dimension and channel number. This vector is then processed and activated by two fully connected layers to generate channel weights, which are used for adaptive weighting of different feature channels. In the spatial attention branch, a 3×3 convolutional layer is used to process the intermediate feature map, and it is activated by the Sigmoid function to generate spatial weights, which are used for adaptive weighting of different spatial locations. The channel weights and spatial weights are fused to obtain joint attention weights, and the joint attention weights are multiplied element-wise with the intermediate feature map to generate an enhanced feature map, which serves as the output of the physical prior guided attention unit.

[0040] In this embodiment, the metal reflection mask prediction unit employs a multi-scale three-dimensional convolutional inversion network to perform pixel-level thresholding of the enhanced feature map output by the physical prior-guided attention unit. This network uses an encoder-decoder structure combined with skip connections to learn the spatial distribution patterns of metal reflection regions and output a metal reflection probability map. This probability map, after thresholding, generates a binary metal reflection mask, thus clearly marking the spatial location of metal reflections. Specifically, the encoder contains four convolutional blocks, each consisting of two 3×3 convolutional layers, a batch normalization layer, and a ReLU activation function. The number of output channels for each convolutional block is set to 32, 64, 128, and 256 respectively. After each convolutional block, downsampling is performed using max pooling with a stride of 2 to gradually reduce the spatial resolution and expand the feature dimension. The decoding end contains four deconvolutional blocks. Each deconvolutional block performs upsampling through transposed convolution to restore spatial resolution and fuses features with the corresponding layer output from the encoding end through skip connections. The number of output channels for each deconvolutional block is set to 128, 64, 32, and 16, respectively. The output layer maps the feature map output from the decoding end to a metal reflection probability map through a 1×1 convolutional layer and a sigmoid activation function. The size of the metal reflection probability map is 256×256×1, where each pixel value represents the predicted probability that the location belongs to a metal reflection region. In this embodiment, a probability threshold of 0.5 is set. When the pixel prediction probability is greater than 0.5, the pixel is determined to belong to a metal reflection region, thereby generating the metal reflection mask to clearly mark the metal reflection regions in the standardized C-scan image. The metal reflection probability map and the metal reflection mask serve as the outputs of the metal reflection mask prediction unit.

[0041] In this embodiment, the pure reflection recovery unit employs a network structure based on residual learning and skip connections to recover weak non-metallic reflection signals that are interfered with or blocked by metallic reflections from the standardized C-scan image. The pure reflection recovery unit uses an encoder-decoder structure, concatenating the standardized C-scan image and the metallic reflection mask along the channel dimension as network input, and outputting a pure radar image of size 256×256×1. By learning the characteristic patterns of metallic reflections, in the metallic regions marked by the metallic reflection mask, the weak reflection structure information of the regions blocked by metallic reflections is reconstructed using the spatial correlation of surrounding effective signals and the laws of electromagnetic wave propagation. In the non-metallic reflection regions, the original effective signals are preserved, and finally, a pure radar image with metallic reflections removed is output. Specifically, the encoding end structure of the pure reflection recovery unit is consistent with that of the metallic reflection mask prediction unit, used to extract multi-scale features. The decoding end, based on the metal reflection mask prediction unit, adds a residual unit after each deconvolution block. This residual unit consists of two 3×3 convolutional layers and a skip connection, used to learn the residual mapping to enhance feature reconstruction capabilities. The clean reflection recovery unit weakens signal interference caused by metal reflection and utilizes the spatial correlation and electromagnetic propagation laws of surrounding non-metallic regions to infer and complete the weak reflection structure of the metal-obscured area, thereby preserving the reflection characteristics of non-metallic targets and improving imaging continuity. The clean radar image after removing metal reflection is the output of the clean reflection recovery unit.

[0042] In this embodiment, the uncertainty generation unit is used to quantify the confidence level of the prediction results of the metal reflection mask prediction unit. The uncertainty generation unit uses the metal region probability map output by the metal reflection mask prediction unit as input to perform uncertainty assessment, calculates the local variance using a 3×3 sliding window, and obtains a local variance map. After normalizing the local variance map, a metal region uncertainty map is generated. The metal region uncertainty map is used to characterize the reliability of the model's decoupling results from metal reflection at different spatial locations. High uncertainty regions typically correspond to metal target boundaries or complex signal regions. The metal region uncertainty map provides a basis for weight adjustment and error control in the subsequent three-dimensional inversion stage. It should be noted that in this embodiment, the uncertainty assessment includes local statistical feature analysis, uncertainty estimation based on Monte Carlo random inactivation inference, or uncertainty assessment based on multi-model ensemble learning.

[0043] It is worth noting that in this embodiment, the metal reflection decoupling module adopts a separate pre-training method, setting hyperparameters such as batch size, initial learning rate, learning rate decay strategy, and training epochs. Its loss function is composed of a metal reflection mask prediction loss and a denoising recovery loss, used to simultaneously constrain the accuracy of metal reflection region recognition and the recovery quality of non-metallic weak reflection signals. Specifically, the metal reflection mask prediction loss measures the difference between the metal reflection probability map output by the metal reflection mask prediction unit and the actual metal region, with a weight of 0.4; the denoising recovery loss constrains the difference between the clean radar map output by the clean reflection recovery unit and the actual non-metallic reflection signal; the denoising recovery loss is composed of a weighted average of the mean square error loss and the structural similarity loss, with weights set to 0.3 and 0.3 respectively. Through the above-mentioned joint optimization of multiple loss functions, accurate localization of the metal reflection region and effective recovery of non-metallic reflection features are achieved.

[0044] Finally, step S300 is executed, in which the metal reflection mask, the pure radar image, and the metal region uncertainty map are input together into the pre-constructed three-dimensional inversion network, and the three-dimensional distribution of the underground dielectric constant is output through multi-scale feature aggregation and voxel-level prediction.

[0045] In an exemplary embodiment of this application, step S300 further includes the following steps: S310: The metal reflective mask, the pure radar image, and the metal region uncertainty map are stitched together along the channel dimension to form a multi-channel three-dimensional feature tensor; S320: Input the multi-channel three-dimensional feature tensor into the pre-constructed three-dimensional inversion network, and output the three-dimensional distribution of the underground dielectric constant through multi-scale feature aggregation and voxel-level prediction.

[0046] It should be noted that, in this embodiment, the three-dimensional inversion network is a neural network model pre-built and trained based on a deep learning framework, used to achieve high-precision three-dimensional inversion of underground structures.

[0047] In an exemplary embodiment of this application, the three-dimensional inversion network includes a multi-scale feature aggregation module, an inversion weight adjustment module, and a voxel-level prediction head; The multi-scale feature aggregation module is used to extract multi-scale features from the multi-channel three-dimensional feature tensor using multiple three-dimensional convolutional kernels of different sizes, and to aggregate the extracted multi-scale features using a feature fusion gating unit to generate aggregated features. It should be noted that the feature fusion gating unit is a mechanism for adaptively controlling the flow of multi-scale feature information, used to dynamically filter and weight features of different scales, rather than simply adding or splicing them together.

[0048] The inversion weight adjustment module is used to dynamically adjust the loss weight of the metal reflection region according to the metal reflection mask, and to assign feature enhancement weights to non-metal reflection regions with uncertainty values ​​below a preset threshold according to the metal region uncertainty map; The voxel-level prediction head is used to output the predicted value of the underground dielectric constant on a voxel-by-voxel basis in order to construct a three-dimensional distribution of the underground dielectric constant.

[0049] Specifically, the input to the 3D inversion network is a multi-channel 3D feature tensor obtained by stitching the metal reflection mask, the clean radar image, and the metal region uncertainty map along the channel dimension, and the output is the 3D dielectric constant distribution of the underground medium. In this embodiment, the 3D inversion network adopts a block voxel prediction strategy, dividing the clean radar image after metal reflection removal into multiple overlapping or non-overlapping sub-voxel blocks according to a preset voxel size, and performing parameter inversion prediction on each sub-voxel block respectively; subsequently, through stitching, weighted fusion, or smoothing, the 3D inversion result of the overall underground structure is obtained, resulting in the 3D distribution of the underground dielectric constant. The 3D distribution of the underground dielectric constant can characterize the changes in the electromagnetic properties of the underground medium and can be further used for underground structure identification and imaging analysis. It is understood that in another embodiment, the 3D inversion network can also adopt a sliding window prediction strategy, using a 3D window of a preset size to slide along the spatial or depth direction in the clean radar image, performing local 3D inversion prediction on the data within the window, and fusing the prediction results at different window positions to form a continuous 3D inversion result. By employing a block voxel prediction strategy or a sliding window prediction strategy, the GPU memory resources required for a single computation can be significantly reduced while ensuring the accuracy of 3D inversion.

[0050] Please see Figure 4 As shown, in an exemplary embodiment of this application, the metal reflection decoupling module and the three-dimensional inversion network are optimized by joint training. The network parameters are iteratively updated through the backpropagation algorithm to minimize the multi-task loss function, which includes metal reflection mask prediction loss, denoising recovery loss, three-dimensional inversion reconstruction loss, and physical consistency loss.

[0051] Specifically, the metal feature decoupling module and the 3D inversion network are optimized through joint training. In this embodiment, a gradient descent-based optimization algorithm is used to iteratively update the network parameters. During training, hyperparameters such as batch size, learning rate, and training epochs are set to ensure the stability of network convergence. A multi-task loss function is introduced during network training to jointly constrain metal reflection separation, non-metal reflection signal recovery, and underground structure inversion reconstruction. The multi-task loss function includes metal reflection mask prediction loss, denoising recovery loss, and 3D inversion reconstruction loss. The weights of each loss term are set according to the importance of the task and can be dynamically adjusted during training. During training, the metal feature decoupling module can be pre-trained using the metal reflection mask prediction loss and denoising recovery loss. In the end-to-end joint training stage, the 3D inversion reconstruction loss is introduced, which, together with the above losses, constitutes a multi-task joint optimization objective, thereby achieving synergistic optimization of metal reflection decoupling and 3D inversion.

[0052] In an exemplary embodiment of this application, a physical consistency loss is introduced to constrain the inversion results to conform to the physical laws of electromagnetic wave propagation in underground media while meeting data fitting accuracy requirements. Specifically, in this embodiment, the three-dimensional spatial distribution of the underground dielectric constant output by the three-dimensional inversion network is used as input. A forward model based on the electromagnetic wave propagation mechanism is used to calculate the simulated radar echo signal under the corresponding conditions. By comparing the difference between the simulated radar echo signal and the target radar signal, a physical consistency loss is constructed, thereby guiding the inversion results during network training and avoiding non-physical calculations that do not conform to the laws of electromagnetic propagation. It is understood that in other embodiments of this application, the physical consistency loss can be constructed offline through high-precision electromagnetic forward simulation or implemented through a pre-trained forward model to avoid high-computational-cost online electromagnetic numerical simulations during training. The forward model can be implemented using an electromagnetic numerical simulation model based on the finite-difference time-domain method, a pre-trained radar signal generation proxy model, or other computational models that satisfy the physical constraints of electromagnetic propagation.

[0053] In summary, the present invention provides a method for decoupling and enhancing metal reflection based on ground-penetrating radar (GPR). First, raw radar echo data collected in-situ by GPR is acquired and preprocessed to obtain a standardized C-scan image. Then, the standardized C-scan image is input into a pre-built metal reflection decoupling module. This module, based on the standardized C-scan image and a pre-built metal reflection physical feature library, utilizes the difference in electromagnetic response characteristics between metallic and non-metallic targets to decouple metal and non-metallic reflections in the standardized C-scan image, outputting a metal reflection mask, a clean radar image, and a metal region uncertainty map. Finally, the metal reflection mask, the clean radar image, and the metal region uncertainty map are input into a pre-built three-dimensional inversion network. Through multi-scale feature aggregation and voxel-level prediction, the three-dimensional distribution of the underground dielectric constant is output. This invention, based on the physical characteristics of metal reflection, precisely separates the metal reflection component from the original radar echo signal through a metal reflection decoupling module. This fundamentally solves the problem of confusion between strong metal reflection and effective signal, achieving selective attenuation of metal reflection rather than simple filtering. This preserves the signal characteristics of weakly reflective targets such as underground cavities, water-rich bodies, and soil layer changes, significantly improving the detection rate of small targets. Simultaneously, through residual learning and spatial correlation modeling of the pure reflection recovery unit, the weak reflection structure in metal-obscured areas is inferred and completed, filling detection blind spots that traditional methods cannot cover, thus improving the integrity of the inversion imaging. Furthermore, diverse metal reflection datasets are generated through electromagnetic forward modeling, enabling the model to adapt to complex metal environments. Reflection decoupling and 3D inversion are integrated into the same network framework for multi-task joint optimization, achieving synergistic effects and ultimately reducing the inversion error of underground dielectric constant, making target boundaries clearer and significantly improving overall imaging quality.

[0054] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for decoupling and three-dimensional inversion enhancement of metal reflections based on ground-penetrating radar, characterized in that, include: The raw radar echo data collected by ground penetrating radar in the field is acquired, and the raw radar echo data is preprocessed to obtain a standardized C-scan image. The standardized C-scan image is input into a pre-built metal reflection decoupling module. The metal reflection decoupling module decouples the metal reflection and non-metal reflection in the standardized C-scan image based on the standardized C-scan image and the pre-built metal reflection physical feature library, and utilizes the difference in electromagnetic response characteristics between metal targets and non-metal targets to output a metal reflection mask, a clean radar image, and a metal region uncertainty map. The metal reflection mask, the pure radar image, and the metal region uncertainty map are input into a pre-constructed three-dimensional inversion network. Through multi-scale feature aggregation and voxel-level prediction, the three-dimensional distribution of the underground dielectric constant is output. The metal reflection decoupling module includes a physical prior-guided attention unit, a metal reflection mask prediction unit, a pure reflection recovery unit, and an uncertainty map generation unit. The physical prior-guided attention unit is used to determine attention weights based on the physical feature library of metal reflection features, and to perform weighted enhancement on the input standardized C-scan image through the synergistic effect of channel attention and spatial attention to generate an enhanced feature map. The metal reflection mask prediction unit is used to perform pixel-level metal reflection probability prediction on the enhanced feature map output by the physical prior guided attention unit to obtain a metal reflection probability map, and to perform threshold segmentation on the metal reflection probability map to generate the metal reflection mask. The pure reflection recovery unit is used to perform channel stitching of the standardized C-scan image and the metal reflection mask output by the metal reflection mask prediction unit, and to infer and complete the weak reflection structure of the metal-shaded area through residual learning and spatial correlation modeling to generate the pure radar image. The uncertainty map generation unit is used to evaluate the uncertainty of the metal reflection probability map output by the metal reflection mask prediction unit in order to generate the metal region uncertainty map.

2. The method for decoupling and three-dimensional inversion enhancement of metal reflection based on ground-penetrating radar according to claim 1, characterized in that, The process of acquiring raw radar echo data collected on-site by ground-penetrating radar and preprocessing the raw radar echo data to obtain a standardized C-scan image includes: Acquire A-scan data generated by ground-penetrating radar scanning point by point along the survey line; Multiple A-scan data are spliced ​​together along the survey line direction to obtain B-scan data; Multiple B-scan data are spatially combined and interpolated to obtain C-scan data, which is a three-dimensional data volume. The C-scan data is subjected to depth slicing, amplitude normalization, and spatial size adjustment to obtain the standardized C-scan image.

3. The method for decoupling and three-dimensional inversion enhancement of metal reflection based on ground-penetrating radar according to claim 1, characterized in that, The metal reflection physical feature library is constructed in the following way: Multiple radar echo samples are acquired, and each radar echo sample corresponds to preset metal parameters and environmental parameters; Each radar echo sample is subjected to feature analysis to extract the corresponding set of physical features of metal reflection. The set of physical features of metal reflection includes at least the peak amplitude, the proportion of high-frequency energy, the number of multiple peak values, the trail length, and the amplitude of local phase perturbation. The physical feature set of metal reflection for each radar echo sample is associated with and stored with the corresponding metal parameters and environmental parameters to form the physical feature library of metal reflection.

4. The method for decoupling and three-dimensional inversion enhancement of metal reflection based on ground-penetrating radar according to claim 3, characterized in that, The radar echo samples were generated using electromagnetic forward modeling tools or obtained through on-site detection and collection by ground-penetrating radar.

5. The method for decoupling and three-dimensional inversion enhancement of metal reflection based on ground-penetrating radar according to claim 1, characterized in that, The uncertainty assessment includes local statistical feature analysis, uncertainty estimation based on Monte Carlo random inactivation inference, or uncertainty assessment based on multi-model ensemble learning.

6. The method for decoupling and three-dimensional inversion enhancement of metal reflection based on ground-penetrating radar according to claim 1, characterized in that, The process involves inputting the metal reflective mask, the clean radar image, and the metal region uncertainty map into a pre-constructed three-dimensional inversion network. Through multi-scale feature aggregation and voxel-level prediction, the network outputs a three-dimensional distribution of the underground dielectric constant, including: The metal reflective mask, the pure radar image, and the metal region uncertainty map are stitched together along the channel dimension to form a multi-channel three-dimensional feature tensor. The multi-channel three-dimensional feature tensor is input into a pre-constructed three-dimensional inversion network, and through multi-scale feature aggregation and voxel-level prediction, the three-dimensional distribution of the underground dielectric constant is output.

7. The method for decoupling and three-dimensional inversion enhancement of metal reflection based on ground-penetrating radar according to claim 6, characterized in that, The three-dimensional inversion network includes a multi-scale feature aggregation module, an inversion weight adjustment module, and a voxel-level prediction head; The multi-scale feature aggregation module is used to extract multi-scale features from the multi-channel three-dimensional feature tensor using multiple three-dimensional convolutional kernels of different sizes, and to aggregate the extracted multi-scale features using a feature fusion gating unit to generate aggregated features. The inversion weight adjustment module is used to dynamically adjust the loss weight of the metal reflection region according to the metal reflection mask, and to assign feature enhancement weights to non-metal reflection regions with uncertainty values ​​below a preset threshold according to the metal region uncertainty map; The voxel-level prediction head is used to output the predicted value of the underground dielectric constant on a voxel-by-voxel basis in order to construct a three-dimensional distribution of the underground dielectric constant.

8. The method for decoupling and three-dimensional inversion enhancement of metal reflection based on ground-penetrating radar according to claim 1, characterized in that, The metal reflection decoupling module and the three-dimensional inversion network are optimized through joint training. The network parameters are iteratively updated through the backpropagation algorithm to minimize the multi-task loss function, which includes metal reflection mask prediction loss, denoising and recovery loss, three-dimensional inversion reconstruction loss, and physical consistency loss.

Images