Multi-scale ground penetrating radar full waveform inversion method based on MSTA-UNet

Abstract

The invention discloses a multi-scale ground penetrating radar full-waveform inversion method based on MSTA-UNet. The method comprises the following steps: S1, decomposing original GPR data into low-frequency, medium-frequency and high-frequency three-channel data through wavelet transform; s2, constructing an MSTA-UNet network, and extracting deep features of each frequency band through a multi-branch module; s3, the cross-band features are fused through a Fusion Attention module, and a decoder recovers the resolution in a deconvolution mode; and S4, adopting a mixed loss function training model of a data domain and a physical domain FWI, and outputting an inversion result containing a dielectric constant. According to the method, cyclic wave jump is relieved through multi-scale frequency division, the black box defect is overcome by fusing data and physical driving, the detection depth and resolution are balanced, the inversion precision, generalization ability and calculation efficiency are improved, and the method is suitable for shallow geophysical detection scenes such as urban underground pipeline detection and road disease detection.

Description

Technical Field

[0001] This invention relates to the field of ground penetrating radar detection and geophysical exploration technology, specifically to a multi-scale ground penetrating radar full waveform inversion method based on MSTA-UNet. Background Technology

[0002] Ground-penetrating radar (GPR), as an efficient and non-destructive shallow geophysical exploration technology, has wide applications in urban underground space development and infrastructure inspection. However, accurately reconstructing the physical parameters of the subsurface medium from complex radar echo data remains a core technological challenge in this field.

[0003] While traditional full waveform inversion (FWI) methods can theoretically provide high-resolution subsurface structure information, they are inherently highly nonlinear and ill-posed problems. They are extremely dependent on the initial model, prone to getting trapped in local minima, and computationally expensive, making them unsuitable for real-time applications. In recent years, deep learning-driven inversion methods have offered new insights into this problem; however, these methods generally suffer from a "black box" limitation, insufficient physical interpretability, and their generalization ability is limited by the completeness of the training dataset, making them ill-suited for complex and ever-changing geological exploration environments. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-scale ground-penetrating radar full waveform inversion method based on MSTA-UNet, which solves the problems of strong nonlinearity, easy occurrence of cycle jumps, and significant dependence on the initial model in existing full waveform inversion methods.

[0005] To solve the above problems, the technical solution of the present invention is as follows: A multi-scale ground-penetrating radar full waveform inversion method based on MSTA-UNet includes the following steps: S1: Acquire raw ground-penetrating radar (GPR) data, perform data preprocessing using the multi-resolution analysis characteristics of wavelet transform, and construct a multi-band input dataset: Select wavelet basis functions to perform discrete wavelet transform (DWT) on the raw GPR data, decompose to obtain approximation coefficients and detail coefficients, and separate low-frequency data, mid-frequency data and high-frequency data through coefficient reconstruction, which serve as three independent input channels; S2: Construct the MSTA-UNet multi-scale feature extraction network to perform deep feature encoding on data of each frequency band: After inputting the frequency-divided data into the initial convolutional layer, the features are captured by the multi-branch feature extraction module containing four different sizes of convolutional kernels. Then, through alternating max pooling and downsampling and feature deepening by the "convolution + multi-scale" module, the deep encoded features of each frequency band are obtained. S3: Introducing a skip connection fusion mechanism and bottleneck layer processing to achieve cross-band feature adaptive enhancement and decoding: The Fusion Attention module weighted fusion of coding features from each frequency band, the bottleneck layer generates global bottleneck features, and the decoder restores spatial resolution through cascaded decoding modules and deconvolution; S4: Construct a hybrid loss function that includes data domain loss and physical domain full waveform inversion (FWI) loss. Use backpropagation and optimizer to iteratively update network parameters. After training, input test data and output the inversion results of underground medium physical parameters.

[0006] Furthermore, the wavelet basis function in S1 is the "db4" wavelet, and the number of decomposition levels of the discrete wavelet transform is... The data consists of 3 to 5 layers, and low, medium and high frequency data are obtained through inverse wavelet transform (IDWT) and length correction.

[0007] Furthermore, the four convolutional kernel sizes of the multi-branch feature extraction module described in S2 are 1×1, 3×3, 5×5, and 7×7, respectively, and the feature extraction process follows the formula: To achieve feature splicing and fusion, in the formula, Input features; This represents a convolution operation with a kernel size of k; Concat indicates feature concatenation. These are the features after fusion.

[0008] Furthermore, the data domain loss described in S4 is a weighted combination of mean absolute error (MAE) and structural similarity (SSIM), and the total data loss is calculated using the formula: Calculate; where, For the total data loss, , To balance the coefficients of MAE and SSIM weights, Mean Absolute Error Loss (MSI) measures the pixel-level difference between the predictive and true models, while SSIM measures the structural similarity (SSI) measures the spatial structural similarity between the predictive and true models.

[0009] Furthermore, the physical domain FWI loss described in S4 is used to construct a forward modeling operator using the finite-difference time-domain (FDTD) method. Substituting the network-predicted physical parameters of the subsurface medium into Maxwell's equations generates a simulated radar profile. Calculate the simulated profile With observation radar data The L2 norm residuals; where, Forward simulation operators, The simulated radar profile generated by forward modeling. For actual radar data, the L2 norm is used to quantify the magnitude of the residual between the two.

[0010] Furthermore, the weighted fusion of the Fusion Attention module described in S3 satisfies in, The characteristics after cross-band fusion , , These are deep coding features for low, medium, and high frequency bands, respectively. The weight parameters are automatically learned by the attention mechanism and satisfy the following conditions: (Used to dynamically adjust the contribution of each frequency band feature to the inversion results).

[0011] Furthermore, the activation function of the decoder described in S3 is the ReLU function, and the decoding process follows the formula... To achieve; among which, Let i be the output feature map of the i-th layer decoder. This is a linear rectification activation function (used to introduce nonlinear characteristics). These are the convolution weights of the decoder. This is the convolution operator. This is an upsampling operation (used to restore the spatial resolution of the feature map). This is the output feature map of the (i-1)th layer decoder. For the fusion of skip connection features at the corresponding level, This is the convolution bias for the decoder.

[0012] Furthermore, the optimizer described in S4 is the Adam optimizer, which calculates gradients through the backpropagation algorithm and iteratively updates network parameters until the model converges.

[0013] Furthermore, the deep feature encoding described in S2 is a four-level encoding process, and the downsampling process is implemented through max pooling, which is performed according to the formula... Calculate; where, The output feature map after pooling is located at... The value at that location, For the window region of max pooling, This is the coordinate offset within the window. The feature map is input to the pooling operation. This is the coordinate position of the output feature map.

[0014] Furthermore, the dielectric constant in the physical parameters of the underground medium is used for the identification of underground medium layers and the discrimination of anomalies (such as buried objects and defect areas), and in shallow detection scenarios with a detection depth of 50cm to 10m, the inversion error of the dielectric constant does not exceed 5%.

[0015] The beneficial effects of this invention are as follows: 1. By adopting a wavelet transform multi-scale frequency division strategy, the problem of "cycle jumping" in traditional full waveform inversion is effectively alleviated. The low-frequency component constrains the deep macro background, the high-frequency component depicts the shallow micro-texture, and the mid-frequency component bridges the scale difference, thus robustly balancing the contradiction between detection depth and spatial resolution.

[0016] 2. The constructed MTA-UNet network has the ability of multi-stream parallel feature extraction and attention fusion, which can realize adaptive extraction and complementary fusion of multi-scale features, significantly improving the inversion ability of weak signals and complex geological structures.

[0017] 3. A dual constraint mechanism of physics-driven and data-driven approaches is proposed. By using a hybrid loss function, the accuracy of the image structure and the consistency of the geophysical field in the inversion results are taken into account. This overcomes the "black box" defect of the pure data-driven method and enhances the physical interpretability and generalization ability of the inversion results.

[0018] 4. It enables rapid end-to-end imaging, improving the accuracy of dielectric constant inversion while reducing computational costs. It also solves the problem of difficulty in extracting weak signals in complex environments, providing strong technical support for urban underground space exploration, road defect detection, and quality assessment of hidden works. Attached Figure Description

[0019] The invention will be further described below with reference to the accompanying drawings: Figure 1 This is a schematic diagram of the workflow of the present invention; Figure 2 This is a schematic diagram of the overall wavelet transform process of the present invention; Figure 3 This is a schematic diagram of the feature extraction part of the MSTA-UNet network of the present invention; Figure 4 This is a schematic diagram of the overall network flow of the MSTA-UNet of the present invention; Figure 5 This is a schematic diagram of the single-channel inversion results of the present invention; Figure 6 This is a schematic diagram of the wavelet frequency division result of the present invention; Figure 7 This is a schematic diagram of the multi-channel inversion results after frequency division according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] A multi-scale ground-penetrating radar full-waveform inversion method based on MSTA-UNet, characterized by the following steps: S1: Acquire raw ground-penetrating radar (GPR) data for urban underground detection, road inspection, or geological survey scenarios. This data includes multi-band echo signals reflected from the underground medium, such as... Figure 1 As shown, the original data is preprocessed using the multi-resolution analysis characteristics of wavelet transform. First, a discrete wavelet transform is performed on the original GPR data to decompose it into approximate coefficients representing the low-frequency trend of the signal and detail coefficients representing the mid-to-high frequency details, such as... Figure 2 The wavelet transform process shown is followed by the extraction of low-frequency, mid-frequency, and high-frequency data through a coefficient reconstruction and separation strategy. These three sets of data are then used as three independent feature input channels, achieving decoupling and hierarchical representation of the radar signal's time-frequency characteristics. This effectively alleviates the cycle skipping problem that easily occurs in traditional single-scale inversion, compared to... Figure 5 The single-channel inversion results shown significantly improve the ability to extract deep features; S2: Construct the MSTA-UNet multi-scale feature extraction network specifically for ground-penetrating radar signal inversion, such as... Figure 3 As shown, the three frequency groups of data obtained from S1 are input into the initial convolutional layer of the network for feature mapping dimension unification. Then, they are fed into the feature extraction module with a multi-branch structure. Combined with downsampling operation and "convolution + multi-scale" deepening module, the features are processed hierarchically to gradually improve the abstraction level of the features and the geological information representation ability. Finally, the deep coding features corresponding to each frequency band are obtained, and the accurate capture of underground geological information at different scales is completed, providing high-quality feature input for subsequent fusion and decoding. S3: Introducing a Fusion Attention skip connection fusion mechanism, such as Figure 4 The overall process of the MSTA-UNet network shown is as follows: through this mechanism, the importance weights of deep coding features in each frequency band are automatically learned, and low, medium and high frequency band features are adaptively weighted and fused. Then, the fused features are sent to the bottleneck layer for global information integration and redundant feature filtering. Then, the four-level cascaded decoder gradually restores the spatial resolution of the feature map through deconvolution operation, realizing the complementary enhancement of cross-frequency band features and accurate decoding of underground structures, and strengthening the ability to identify small geological anomalies corresponding to weak signals. S4: Construct a hybrid loss function that includes data domain loss and physical domain full waveform inversion (FWI) loss. The data domain loss focuses on the image accuracy of the inversion results, while the physical domain loss constrains the inversion results to conform to geophysical propagation laws. Based on this hybrid loss function, the gradient is calculated using the backpropagation algorithm, and all trainable parameters of the network are iteratively updated. After the model loss converges to a stable interval, the preprocessed test data is input into the trained network, and the output is the inversion results of underground medium physical parameters, including the dielectric constant, as shown below. Figure 7 The multi-channel inversion results shown combine the advantages of data-driven and physics-driven approaches, overcoming the shortcomings of insufficient physical interpretability in pure data-driven methods, achieving "end-to-end" rapid imaging, and balancing inversion accuracy and computational efficiency.

[0022] Furthermore, the wavelet transform described in S1 uses the "db4" wavelet as the basis function. This wavelet basis possesses excellent time-frequency localization characteristics and smoothness, enabling it to accurately adapt to the decomposition requirements of non-stationary ground-penetrating radar signals. Its time-frequency analysis advantages allow for the precise separation of effective information and interference signals in GPR signals; for example... Figure 2 The wavelet transform process shown uses a discrete wavelet transform with a decomposition level of 3-5 levels. This level range ensures that low-frequency components fully preserve deep geological macroscopic background information, mid-frequency components effectively bridge the differences between deep and shallow scales, and high-frequency components accurately characterize shallow micro-textures, thus robustly balancing the contradiction between detection depth and spatial resolution. Figure 6 The wavelet frequency division results shown are used to reconstruct the signals of the separated coefficients through inverse wavelet transform (IDWT), and a length correction operation is performed on the reconstructed signals to ensure that the lengths of low-frequency, mid-frequency, and high-frequency data are consistent with the original ground-penetrating radar data. This provides standardized multi-channel data for subsequent network input and improves the stability and consistency of model training.

[0023] Furthermore, the multi-branch feature extraction module described in S2 includes convolutional kernels of four sizes: 1×1, 3×3, 5×5, and 7×7, such as... Figure 3 The MSTA-UNet network feature extraction section shown uses a 1×1 convolutional kernel for dimensionality compression and channel fusion of input features, significantly reducing network computation while maintaining feature expressiveness; a 3×3 convolutional kernel for capturing local geological details (such as edges of minor defects); a 5×5 convolutional kernel for extracting mid-range spatial structure features (such as interfaces between different media layers); and a 7×7 convolutional kernel for perceiving global geological background features (such as large-scale homogeneous media regions). These four kernel sizes work synergistically to overcome the technical limitation of traditional single convolutional kernels in simultaneously capturing detailed textures and global structure. The feature extraction process follows the formula: The process involves first concatenating the output features of the four convolutional kernels along the channel dimension using the concat operator, then unifying the feature dimensions and fusing information using a 3×3 convolutional kernel, ultimately yielding a result that incorporates information from multiple receptive fields. By integrating features and enhancing their complementarity, we can improve our ability to characterize complex geological structures such as irregular cavities and pipeline branches.

[0024] Furthermore, the data domain loss described in S4 is a weighted combination of mean absolute error (MAE) and structural similarity (SSIM), and the total data loss is calculated using the formula: Calculation; where For the total data loss, , These are weighting coefficients, with values ​​ranging from 0.3 to 0.7. They can be dynamically adjusted according to the detection scenario (e.g., increasing the weighting coefficient in a homogeneous medium scenario). The complexity of the structural scenarios increases ), adaptable to different geological exploration needs; By calculating the mean absolute error between corresponding pixels in the predicted model and the real model, the numerical accuracy of physical parameters such as dielectric constant is ensured. SSIM optimizes the edge texture and spatial structure integrity of the underground medium by measuring the mean similarity, variance similarity, and covariance similarity between the two models. This weighted combination of both compensates for the shortcomings of a single loss function, which either emphasizes numerical accuracy or ignores structural details, significantly improving the image quality of the inversion results. Figure 7 The inversion results shown are in high agreement with the ground truth model, providing a clear and reliable basis for subsequent geological interpretation.

[0025] Furthermore, the physical domain FWI loss described in S4 is used to construct a forward modeling operator using the finite-difference time-domain (FDTD) method. This operator possesses high-precision time-domain discretization characteristics, enabling it to accurately simulate the propagation process of electromagnetic waves in underground media and ensure the effectiveness of physical constraints. Substituting the network-predicted physical parameters of the underground media (primarily the dielectric constant) into Maxwell's equations, forward modeling is used to generate simulated radar profiles consistent with the observed data format. ; Calculate the simulated profile Compared with actual observation radar data The L2 norm residual is used to quantify the fit between the inversion results and the real physical propagation laws, thereby constraining the network learning direction and ensuring that the inversion results are not only similar to images, but also conform to the propagation characteristics of geophysical fields. This effectively overcomes the "black box" defect of pure data-driven methods, improves the physical interpretability of inversion results, and reduces the dependence on training datasets. Even in complex geological scenarios with incomplete training data, it can still maintain good inversion robustness.

[0026] Furthermore, the weighted fusion of the Fusion Attention module described in S3 satisfies: like Figure 4 The overall flow of the MSTA-UNet network is shown below, where These are global features obtained after cross-band fusion, used for subsequent bottleneck layer processing and decoding; , , These are the encoded deep features of low, medium and high frequency bands, respectively, corresponding to deep macro background, medium scale transition and shallow detail texture information; Let be the weight parameters automatically learned by the attention mechanism through backpropagation, and satisfy . Its value is dynamically adjusted according to the detection scenario (such as in shallow fine detection). As the value increases, during deep macroscopic detection By increasing the weighted values, the adaptive contribution of each frequency band feature to the inversion results is adjusted, solving the problem of frequency band information imbalance caused by traditional fixed-weight fusion. This adaptive fusion mechanism enhances the extraction capability of weak signals (such as high-frequency echoes from minute defects and low-frequency responses from deep media), significantly improving the model's adaptability to complex geological environments. Whether for shallow fine-scale exploration or deep macroscopic investigation, it can output results such as… Figure 7 The high-quality inversion results shown are presented.

[0027] Furthermore, the activation function of the decoder described in S3 is the ReLU function. This function introduces non-linear feature representation capabilities into the network by setting negative input values ​​to zero, effectively alleviating the gradient vanishing problem in deep network training and improving network training stability and feature representation capabilities; for example... Figure 4 The overall MSTA-UNet network flow shown below, the decoding process is based on the formula: Implementation; among which Let i be the output feature map of the i-th layer decoder. These are the convolution weight parameters for the decoder. This is a two-dimensional convolution operator. The upsampling operation (implemented using bilinear interpolation or transposed convolution) is used to upsample the output feature map of the (i-1)th layer decoder. The spatial resolution has been increased to meet the requirements of the current level. The fusion skip connection features corresponding to the encoding stage can transfer the high-resolution detail information preserved in the encoding stage to the decoding stage, avoiding structural blurring caused by resolution loss. The convolutional bias parameters of the decoder are used to adjust the overall offset of the feature map, improving decoding flexibility; the hierarchical decoding logic ensures the accuracy of structural reconstruction from global to local, making the spatial resolution of the inversion result match the original radar data, meeting the requirements of high-precision detection, and finally obtaining... Figure 7 The clear inversion image shown.

[0028] Furthermore, the iterative training described in S4 employs the Adam optimizer, which combines momentum gradient descent with an adaptive learning rate strategy. In the early stages of training, a relatively large learning rate (initial value 1e-4~1e-3) achieves rapid convergence, while in the later stages, the learning rate is automatically reduced for precise parameter fine-tuning, addressing the slow convergence and susceptibility to local minima issues of traditional SGD optimizers. Backpropagation is used to traverse all trainable parameters of the network (including convolution weights, biases, attention weights, etc.), calculating the gradient of the total loss function with respect to each parameter to ensure that the parameter update direction is consistent with the loss descent direction, thus improving model training accuracy. During iterative updates, the criterion for model convergence is "the total loss fluctuation over 10~20 consecutive training epochs does not exceed 1e-6," avoiding overfitting or underfitting. This ensures that the trained model maintains good generalization ability across different test datasets, guaranteeing stable output in practical applications. Figure 7 The high-precision inversion results are shown.

[0029] Furthermore, the deep feature encoding described in S2 is a four-level encoding process, with each level corresponding to one feature extraction and downsampling operation; such as Figure 3 The feature extraction part of the MSTA-UNet network shown below uses a downsampling process implemented through max pooling, which is performed according to the formula: Calculation; where The output feature map after pooling is in coordinates The value at that location, The window region for max pooling (size 2×2 or 3×3). This is the coordinate offset within the window (the value range matches the window size). The feature map input to the pooling operation contains geological feature information of the current level. This operation, by retaining the maximum feature value within the window, reduces the feature map size, the number of network parameters, and computational load (providing technical support for real-time inversion), while strengthening the representation of key geological features (such as medium boundaries and anomaly contours), effectively suppressing noise interference, and improving feature robustness. The hierarchical downsampling of the four-level pooling gradually increases the level of feature abstraction, transforming the original signal features into geological structure features, laying a high-quality foundation for subsequent fusion and decoding, and ultimately contributing to the generation of features such as... Figure 7 The results show the accurate inversion.

[0030] Furthermore, the dielectric constant, a core parameter reflecting the properties of underground media, is a key physical parameter. Different types of media exhibit significant differences in dielectric constant (e.g., soil has a dielectric constant of 4-15, concrete 6-10, and metals are close to infinity). Based on this difference, it is possible to identify underground media layers and determine anomalies (such as buried objects, cavities, pipelines, and defective areas), directly addressing engineering needs such as urban underground pipeline location, roadbed defect detection, and archaeological site investigation, thus enhancing the practicality of the technical solution. The method is applicable to various detection methods. With a depth of 50cm to 10m, this range comprehensively covers core application scenarios such as urban underground pipelines (0.5 to 3m burial depth), roadbed defects (0.5 to 5m burial depth), and shallow exploration of archaeological sites (1 to 10m burial depth), broadening the applicability of the technical solution. Through the synergistic effect of multi-scale frequency division strategy, dual-domain loss constraint, and MTA-UNet network, the dielectric constant inversion error is controlled within 5%, significantly better than traditional full-waveform inversion (error typically above 10%) and pure data-driven methods (error above 8%). Figure 7 The inversion results shown are in high agreement with the real model, fully meeting the actual needs of high-precision engineering detection and geological exploration.

[0031] The embodiments described in this specification are merely examples of implementations of the inventive concept. The scope of protection of this invention should not be considered as limited to the specific forms stated in the embodiments. The scope of protection of this invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A multi-scale ground-penetrating radar full-waveform inversion method based on MSTA-UNet, characterized in that, Includes the following steps: S1: Acquire raw ground-penetrating radar (GPR) data, perform multi-resolution preprocessing through wavelet transform, decompose and reconstruct to obtain low-frequency, medium-frequency and high-frequency data, which serve as three independent input channels; S2: Construct the MTA-UNet multi-scale feature extraction network, and process the frequency-divided data through a multi-branch feature extraction module, downsampling and feature deepening to obtain deep coding features for each frequency band; S3: The Fusion Attention module weighted and fused the coding features of each frequency band. After processing by the bottleneck layer, the decoder restored the spatial resolution through deconvolution, realizing cross-frequency band feature adaptive enhancement and decoding. S4: Construct a hybrid loss function containing data domain loss and physical domain full waveform inversion (FWI) loss. After iteratively training the network parameters, input test data and output the inversion results of underground medium physical parameters, including dielectric constant.

2. The multi-scale ground-penetrating radar full waveform inversion method based on MSTA-UNet according to claim 1, characterized in that, The wavelet basis function in S1 is the "db4" wavelet, and the number of decomposition levels of the discrete wavelet transform is... The data consists of 3 to 5 layers, and low, medium and high frequency data are obtained through inverse wavelet transform (IDWT) and length correction.

3. The multi-scale ground-penetrating radar full waveform inversion method based on MSTA-UNet according to claim 1, characterized in that, The four convolutional kernel sizes of the multi-branch feature extraction module described in S2 are 1×1, 3×3, 5×5, and 7×7, respectively. The feature extraction process follows the formula: To achieve feature splicing and fusion, in the formula, Input features; This represents a convolution operation with a kernel size of k; Concat indicates feature concatenation. These are the features after fusion.

4. The multi-scale ground-penetrating radar full waveform inversion method based on MSTA-UNet according to claim 1, characterized in that, The data domain loss described in S4 is a weighted combination of mean absolute error (MAE) and structural similarity (SSIM), and the total data loss is calculated using the formula: Calculate; where, For the total data loss, , To balance the coefficients of MAE and SSIM weights, SSIM represents the average absolute error loss and structural similarity.

5. The multi-scale ground-penetrating radar full waveform inversion method based on MSTA-UNet according to claim 1, characterized in that, The physical domain FWI loss described in S4 is constructed using the finite-difference time-domain (FDTD) method to create a forward modeling operator. Substituting the network-predicted physical parameters of the subsurface medium into Maxwell's equations generates a simulated radar profile. Calculate the simulated profile With observation radar data The L2 norm residuals.

6. The multi-scale ground-penetrating radar full waveform inversion method based on MSTA-UNet according to claim 1, characterized in that, The weighted fusion of the Fusion Attention module described in S3 satisfies: in, The characteristics after cross-band fusion , , These are deep coding features for low, medium, and high frequency bands, respectively. The weight parameters are automatically learned by the attention mechanism and satisfy the following conditions: .

7. The multi-scale ground-penetrating radar full waveform inversion method based on MSTA-UNet according to claim 1, characterized in that, The activation function of the decoder described in S3 is the ReLU function, and the decoding process follows the formula: To achieve; among which, Let be the output feature map of the i-th layer decoder, and ReLU be the linear rectified activation function. These are the convolution weights of the decoder. Here, "convolution" is the operator, and "upsample" is the upsampling operation. This is the output feature map of the (i-1)th layer decoder. For the fusion of skip connection features at the corresponding level, This is the convolution bias for the decoder.

8. The multi-scale ground-penetrating radar full waveform inversion method based on MSTA-UNet according to claim 1, characterized in that, The optimizer described in S4 is the Adam optimizer, which calculates gradients through backpropagation and iteratively updates network parameters until the model converges.

9. The multi-scale ground-penetrating radar full waveform inversion method based on MSTA-UNet according to claim 1, characterized in that, The deep feature encoding described in S2 is a four-level encoding process. The downsampling process is implemented through max pooling, which is performed according to the formula: Calculate; where, The output feature map after pooling is located at... The value at that location, For the window region of max pooling, This is the coordinate offset within the window. The feature map is input to the pooling operation. This is the coordinate position of the output feature map.

10. The multi-scale ground-penetrating radar full-waveform inversion method based on MSTA-UNet according to claim 1, characterized in that, The dielectric constant in the physical parameters of the underground medium is used for the identification of underground medium layers and the discrimination of anomalies. In shallow detection scenarios with a detection depth of 50cm to 10m, the inversion error of the dielectric constant does not exceed 5%.

Images