Road cavity detection method based on ground penetrating radar three-dimensional forward modeling and SAM2-Unet fusion

Abstract

The invention belongs to the technical field of road detection, and particularly discloses a road cavity detection method based on ground penetrating radar three-dimensional forward modeling and SAM2-Unet fusion, and the method comprises the following steps: S1, transmitting electromagnetic waves to a to-be-detected road through a ground penetrating radar, receiving a reflection signal, generating a ground penetrating radar image, and analyzing the ground penetrating radar image; s2, data preprocessing; s3, constructing an underground medium three-dimensional velocity model; s4, performing geometric ray tracing in the three-dimensional velocity model, and determining a propagation path and a time parameter from an electromagnetic wave reflection point to a ground surface receiving point; s5, superposing diffracted wave contributions by adopting a three-dimensional Kirchhoff migration algorithm; s6, inputting the offset result image into a pre-trained SAM2-Unet network for high-precision segmentation; and S7, analyzing three-dimensional space coordinates, extension directions and burial depth parameters of the road cavity. According to the method, three-dimensional migration imaging and a deep learning method are combined, the positioning precision and the morphological representation capability of road cavity detection are improved, technical support can be provided for intelligent diagnosis of underground diseases, and the method has remarkable engineering application value.

Description

Technical Field

[0001] The present invention belongs to the technical field of non-destructive detection of road underground diseases, and particularly relates to a road cavity detection method based on the fusion of three-dimensional forward modeling of ground penetrating radar and SAM2-Unet. Background Technique

[0002] With the rapid development of transportation infrastructure, the demand for road underground disease detection has become increasingly urgent. However, the traditional two-dimensional ground penetrating radar technology is limited by the information dimension and it is difficult to construct an underground disease model with spatial continuity. Especially for the complex cavity diseases caused by subgrade void and pipe gallery leakage, the two-dimensional profile imaging has inherent defects such as fuzzy spatial feature interpretation and inaccurate geometric parameter quantification, which seriously restricts the accuracy and reliability of disease detection. In recent years, to meet the growing transportation demand, China has accelerated the promotion of large-scale road engineering construction. However, with the deepening of the urbanization process, the underground space structure of urban roads has become increasingly complex, and the operating load of underground pipe networks has continued to increase, which has led to the increasingly prominent problem of road diseases. According to statistics, the direct economic loss caused by road collapses in Beijing in 2022 exceeded 5 million yuan, and the relevant maintenance costs were as high as 200 million yuan. In addition, a main road in Guangzhou caused a major traffic jam due to a subgrade cavity caused by underground pipeline leakage, and its indirect economic loss was immeasurable. In cities built earlier, urban roads often suffer from frequent occurrence of adverse diseases such as subgrade cavities and voids due to factors such as overloading, excessive groundwater extraction, underground construction damage, and pipeline rupture and water seepage, which seriously threaten road safety and significantly increase the urban operation and maintenance costs.

[0003] In the field of road disease detection, the accurate identification of subgrade cavities and voids plays a crucial role in ensuring the safe and stable operation of roads and has become a key link in the detection work. The ground penetrating radar technology occupies an important position in the detection of underground structures with its non-destructive detection characteristics and is one of the main widely used means at present. However, the traditional research mode mainly relies on the analysis of two-dimensional radar profile images. Due to the limited information dimension of two-dimensional images, only limited cross-sectional information can be presented, and it is difficult to completely and accurately depict the complex three-dimensional structure of the subgrade. There are obvious deficiencies in the reflection of key information such as the specific location, shape, and distribution range of voids, which greatly increases the risk of misjudgment and reduces the reliability of the detection results in actual detection.

[0004] Compared with two-dimensional imaging, three-dimensional forward simulation has significant advantages. It can more accurately and comprehensively depict key features such as the detailed shape, spatial position, and influence range of underground cavities, providing a more solid and reliable theoretical support for the accurate detection of road diseases and becoming one of the important research directions at present. In recent years, the organic combination of three-dimensional visualization technology and ground penetrating radar forward simulation has created new methods and approaches for underground cavity detection. However, due to the high complexity of the geological environment and the diversity of underground structure characteristics, in practical engineering applications, accurately obtaining cavity data and achieving high-precision simulation still face many challenges, such as signal interference and insufficient model construction accuracy under complex geological conditions, which need to be solved urgently. Therefore, how to achieve accurate modeling and efficient identification of underground cavities in complex environments has become a key technical problem that needs to be broken through in the current field.

[0005] Therefore, further optimizing the accuracy of forward simulation, improving the effect of three-dimensional visualization, and at the same time enhancing the efficiency of data collection and processing not only have important significance for enriching and improving the theoretical system of road disease detection, but also can provide targeted and guiding solutions for practical engineering practice, with extremely important practical value. At present, significant progress and achievements have been made in the three-dimensional simulation research based on ground penetrating radar through the efforts of many scholars. For example, Tan Cai et al. (2024, CN202411517494.0) innovatively proposed a three-dimensional imaging method and system for termite nests in dikes based on ground penetrating radar. Through data preprocessing, feature extraction and encryption, a deep learning model is constructed based on the U-shaped network for imaging, which can accurately capture the complex features of termite nests and dike structures, improve the imaging accuracy, and be used in the field of termite nest detection. Zhou Feng et al. (2024, CN202411149800.X) proposed a three-dimensional ground penetrating radar target recognition training method based on the cyclic generative adversarial network. Through data slicing, augmentation, and point cloud training, the efficiency and accuracy of underground target recognition are improved, and false detections are reduced. Huang Zhiyong (2022, CN202210436380.8) proposed a three-dimensional ground penetrating radar crack disease recognition method and system for road disease detection. First, a training sample set is composed of road sample images containing pixel label information. A crack Unet neural network is constructed based on the spatial attention mechanism, VGG16, and Unet neural network and trained using the training sample set to obtain a crack recognition model. Then, the road to be recognized is scanned with a three-dimensional ground penetrating radar to obtain an image, and the road cracks are accurately determined by classifying the image pixels according to the model.

[0006] Despite many research results, the optimization of three-dimensional modeling for cavities of different sizes, the efficient application of simulation signals in practical engineering, and the high-precision reconstruction of cavity structures are still key problems in the research. Summary of the Invention

[0007] In view of this, the object of the present invention is to provide a road cavity detection method based on the fusion of three-dimensional forward modeling of ground penetrating radar and SAM2-Unet. The present invention aims to solve the problems of incomplete detection and low detection accuracy in existing road cavity detection methods.

[0008] The present invention provides a road cavity detection method based on the fusion of three-dimensional forward modeling of ground penetrating radar and SAM2-Unet, comprising the following steps:

[0009] S1. An electromagnetic pulse wave is transmitted to the road to be detected through a ground penetrating radar system, the reflected signal is received and a ground penetrating radar image is generated, and the amplitude, frequency change and abnormal distribution characteristics of the signal are analyzed;

[0010] Among them, the ground penetrating radar system uses a pseudo-three-dimensional detection method for road cavity detection;

[0011] S2. The data collected in step S1 is preprocessed to avoid confusion or weakening of the interpretation of the overall signal characteristics;

[0012] S3. According to the actual situation or prior knowledge of the underground medium, a three-dimensional velocity model of the underground medium is constructed based on the finite-difference time-domain method (FDTD), and YEE grids are used for discretization to ensure numerical stability;

[0013] S4. Based on the three-dimensional velocity model constructed in step S3, geometric ray tracing calculations are performed to determine the propagation paths and times of waves from each point underground to the surface receiving points;

[0014] S5. After the geometric ray tracing calculations are completed, for each data point received on the surface, using the three-dimensional Kirchhoff migration algorithm, according to the path and time information of the geometric ray tracing, the diffraction wave contributions from all underground reflection points are calculated and superimposed;

[0015] S6. The migrated result image is input into a pre-trained SAM2-Unet network for high-precision segmentation;

[0016] Among them, the encoding-decoding architecture of the SAM2-Unet deep learning model incorporates multi-scale dilated convolution modules, which can automatically extract the geometric features of underground targets;

[0017] S7. Combining the output results of deep learning for verification and further quantification of the three-dimensional cavity model, considering the wave propagation path, velocity change and wavefront curvature, migration calculations are performed for all surface receiving points to generate a three-dimensional migrated imaging result of the underground structure, thereby reflecting the spatial position of the underground cavity and revealing the shape and properties of the cavity.

[0018] Further, in the step S1, the pseudo-3D detection method is to detect through multiple parallel survey lines or regular grid survey lines, and use the coordinate information in the ground penetrating radar file to fuse the radar data on different survey lines into a 3D data volume.

[0019] Further, in the step S2, the preprocessing of the data includes exponential gain and direct wave removal processing;

[0020] Exponential gain: The time exponential gain function is used to perform gain processing on each ground penetrating radar signal; among them, the maximum gain multiple y max is 100;

[0021] Direct wave removal processing: The average tracking method is used to remove the direct wave. Specifically, the average waveform of the direct wave is extracted from the original data, and the direct wave is suppressed by subtracting the average waveform channel by channel.

[0022] Further, in the step S7, the specific steps for reflecting the spatial position of the underground cavity are as follows:

[0023] I. Use the isosurface feature extraction method to complete the accurate reconstruction of the cavity shape, so as to present the shape and position of the cavity in the visualization image;

[0024] II. Quantitatively evaluate the accuracy of the reconstructed position, obtain the vertical parameters of the cavity based on the wavelet time-energy density analysis, and verify the height measurement accuracy in combination with deep learning; calculate the 3D volume through the spatial integration of the top-down projected area and the vertical parameters;

[0025] III. Use the model for verification to further ensure the practicality and reliability of the results;

[0026] The SAM2-Unet network model is used to train the radar image dataset after the offset of the underground cavity of the road. After the training is completed, the offset data of the surface receiving points is imported into the SAM2-Unet network model, so as to obtain the segmentation result of the cavity boundary of the underground structure, and then reflect the spatial position of the underground cavity.

[0027] Further, during the training process of the SAM2-Unet network model:

[0028] The number of iteration rounds is 50, the batch size is 4, and the initial learning rate is 0.001;

[0029] When updating the network weights, the Adam optimizer is used to optimize the loss function, and the training platform uses an Intel Core i7-12700F processor and an NVIDIA GeForce RTX3060 12GB graphics card.

[0030] Beneficial effects:

[0031] The present invention proposes a road cavity detection method based on the fusion of 3D forward modeling of ground penetrating radar and SAM2-Unet. The present invention uses GprMax software to construct a 3D road model with cavities of three different sizes to simulate real road cavity scenarios. Through 3D visualization processing of the simulated radar signals, the signal characteristics and distribution are intuitively observed. On this basis, the 3D Kirchhoff migration imaging technique is used to process the cavity reflection waves, achieving precise focusing of the signals and improving the recognition and positioning accuracy of the reflection signals. In addition, the isosurface extraction method is also used to extract the features of the signals related to the cavities, completing the refined reconstruction of the cavity shape and clearly presenting the 3D geometric features of the cavities. To verify the reliability and practical application value of the proposed method, on-site test simulations are further carried out. A model with a preset specific size cavity is buried in a sandbox to simulate the actual road environment, and a ground penetrating radar device is used to collect signals. The collected signals are processed and analyzed through the method system established in the previous research. Finally, the accurate visualization imaging of the position, trend, and depth of the actual cavity is successfully achieved, providing accurate and reliable technical means and data support for the detection and evaluation of road cavity diseases, and verifying the effectiveness and feasibility of the method in actual engineering applications.

[0032] Other advantages, objectives, and features of the present invention will be described to some extent in the subsequent specification, and to some extent, based on the study of the following text, will be obvious to those skilled in the art or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following specification. Brief Description of the Drawings

[0033] Figure 1 It is a flowchart of the road cavity detection method of the present invention;

[0034] Figure 2 It is a schematic diagram of 3D slice imaging data acquisition;

[0035] Figure 3 It is a structural diagram of the SAM2-Unet network model;

[0036] Figure 4 It is the segmentation result of the underground cavity by the SAM2-Unet network model;

[0037] Figure 5 It is a YEE grid structure;

[0038] Figure 6 It is a flowchart of 3D migration imaging and visualization processing;

[0039] Figure 7 It is the preprocessing result of the measured data;

[0040] Figure 8 It is the extraction result of the three-dimensional cavity feature of the measured data. Specific implementation mode

[0041] To make the technical solutions, advantages and objectives of the present invention clearer, the following will clearly and completely describe the technical solutions of the embodiments of the present invention in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work belong to the protection scope of this application.

[0042] Ground penetrating radar is an advanced geophysical exploration technology that uses high-frequency electromagnetic pulse waves to non-destructively detect underground structures and is widely used in multiple fields such as geological exploration, transportation engineering (especially road disease detection), and archaeology. In road engineering, ground penetrating radar can effectively identify diseases such as cavities and cracks under the road. This technology provides a way for researchers and engineers to deeply understand the underground structure and geological characteristics of the road by accurately detecting different underground medium layers (including soil, subgrade materials, and potential cavity areas), and provides accurate geological information support for road maintenance, repair, and new construction projects. There are significant differences in the reflection, refraction, and scattering characteristics of electromagnetic waves by different media under the road (such as solid subgrade and cavities). When electromagnetic waves encounter these medium layers, especially abnormal structures such as cavities, part of the electromagnetic waves will be reflected back to the ground. The ground penetrating radar system captures these reflected electromagnetic waves and carefully analyzes the amplitude, frequency changes of the received waves, and the distribution characteristics of abnormal signals, so as to accurately infer the position, scale of the cavity under the road, and the properties and structural states of the surrounding media.

[0043] As Figure 1 shown, the present invention provides a road cavity detection method based on the fusion of three-dimensional forward modeling of ground penetrating radar and SAM2-Unet, including the following steps:

[0044] S1. Transmit electromagnetic pulse waves to the road to be detected through the ground penetrating radar system, receive the reflected signals and generate ground penetrating radar images, and analyze the amplitude, frequency changes, and abnormal distribution characteristics of the signals;

[0045] In practical applications, by moving the antenna in the direction of the survey line, the ground penetrating radar system automatically records each reflected signal according to a preset sampling step, thereby obtaining the ground penetrating radar profile of the survey line. To accurately obtain the spatial geometric shape characteristics of underground targets, three-dimensional detection is particularly important. The three-dimensional detection methods are mainly divided into true three-dimensional and pseudo-three-dimensional (2.5D) types. True three-dimensional detection requires that the step increments in different directions are the same and less than the Nyquist sampling step. Pseudo-three-dimensional detection is carried out by multiple parallel survey lines or regular grid survey lines, and using the coordinate information in the ground penetrating radar file, the radar data on different survey lines are fused into a three-dimensional data volume. As Figure 2 shown, the present invention uses a pseudo-three-dimensional detection method for data construction, and realizes a three-dimensional and all-round scan of the internal structure of the road by moving and scanning along multiple parallel or grid-shaped survey lines in the plane.

[0046] S2. Preprocess the data collected in step S1 to avoid confusion or weakening of the interpretation of the overall signal characteristics;

[0047] In step S2, the preprocessing of the data includes exponential gain and direct wave removal processing;

[0048] During the propagation of electromagnetic waves, due to the difference in dielectric constants between air and the road soil layer medium, energy loss, signal interference and fluctuations will occur, which will affect the signal quality in other areas. To ensure the objectivity and accuracy of the research results, the forward modeling data of all survey lines need to be uniformly preprocessed to avoid confusion or weakening of the interpretation of the overall signal characteristics. In this paper, a time exponential gain function is used to perform gain processing on each ground penetrating radar signal. To prevent excessive gain, the maximum gain multiple y max is set to 100 to ensure that the gain operation is carried out within a reasonable range, thereby realizing effective exponential gain of the ground penetrating radar signal. The direct wave is the wave that directly propagates from the transmitting antenna to the receiving antenna, and it often masks the reflection signal of shallow underground targets. To remove the direct wave, an average tracking method is adopted, which realizes by subtracting the direct wave component from the original data. Since the energy of the direct wave is mainly concentrated in the first few sampling points, while the reflection signals of underground targets are relatively scattered in the subsequent sampling points, this method can effectively remove the direct wave and retain the reflection signals of underground targets, especially suitable for the case where the direct wave component in the echo signal changes little.

[0049] S3. According to the actual situation or prior knowledge of the underground medium, construct a three-dimensional velocity model of the underground medium based on the finite-difference time-domain method (FDTD), and ensure numerical stability through YEE grid discretization;

[0050] S4. Based on the three-dimensional velocity model constructed in step S3, perform geometric ray tracing calculations to determine the propagation paths and times of waves from various points underground to surface receiving points;

[0051] S5. After the geometric ray tracing calculations are completed, for each data point received on the surface, use the three-dimensional Kirchhoff migration algorithm to calculate and stack the diffracted wave contributions from all underground reflection points according to the path and time information of the geometric ray tracing;

[0052] Three-dimensional migration imaging can relocate the diffracted waves generated by horizontal reflection interfaces and cuboid cavity anomalies in the model. The energy completely converges to the accurate positions where the horizontal reflection interfaces and cuboid cavities are located, and clearly depicts the true shape and spatial distribution of the cuboid cavities. Thus, by analyzing the migration slice maps in different directions and positions, the three-dimensional spatial position distribution of the cavity anomalies can be accurately located. By comparing the radar profiles before and after migration, it can be seen that this three-dimensional migration algorithm can relocate the reflected waves and converge the diffracted waves in the forward modeling profile, greatly improving the resolution of the radar profile.

[0053] S6. Input the migration result image into the pre-trained SAM2-Unet network for high-precision segmentation. Its encoder-decoder architecture integrates multi-scale dilated convolution modules to automatically extract the geometric features of underground targets;

[0054] S7. Combine the output results of deep learning for verification and further quantification of the three-dimensional cavity model. Considering the wave propagation path, velocity changes, and wavefront curvature, perform migration calculations for all surface receiving points to generate the three-dimensional migration imaging results of the underground structure, thereby reflecting the spatial positions of underground cavities and revealing the shapes and properties of the cavities;

[0055] In step S7, the specific steps to reflect the spatial positions of underground cavities are as follows:

[0056] I. Use the isosurface feature extraction method to accurately reconstruct the shape of the cavity, so as to present the shape and position of the cavity in the visualization image;

[0057] II. Quantitatively evaluate the accuracy of the reconstructed position, obtain the vertical parameters of the cavity based on wavelet time-energy density analysis, and combine deep learning to verify the height measurement accuracy; calculate the three-dimensional volume through the spatial integration of the top-down projected area and the vertical parameters;

[0058] III. Use the model for verification to further ensure the practicality and reliability of the results;

[0059] The SAM2-Unet network model is used to train the road underground cavity dataset. After training, the offset data of the surface receiving points is imported into the SAM2-Unet network model to obtain the three-dimensional offset imaging result of the underground structure, thereby reflecting the spatial position of the underground cavity.

[0060] SAM2-UNet is a segmentation model based on the U-Net structure. Its encoder uses the Hiera backbone network from (SegmentAnything Model2, SAM2), while the decoder adopts the classic U-shaped design. As Figure 3 shown, the overall architecture of this model consists of four main components: Encoder, Decoder, Receptive Field Blocks (RFB), and Adapters. SAM2-UNet constructs a concise, efficient, and powerful image segmentation framework by integrating the Hiera pre-trained backbone network of SAM2 and the U-shaped decoder. In the model, the receptive field blocks are used to reduce the number of channels, and the adapters implement a parameter-efficient fine-tuning mechanism. This design enables SAM2-UNet to exhibit good generality in various image segmentation tasks and can effectively adapt to different types of image segmentation requirements.

[0061] The input of the SAM2-Unet network model is the preprocessed three-dimensional offset data of the ground penetrating radar, and the output is the semantic segmentation result of the underground cavity. During the training process, the following key parameters are set: the number of iterations is 50 rounds, the batch size is 4, and the initial learning rate is 0.001 to optimize the feature extraction ability of the network model and thus improve the recognition accuracy. When updating the network weights, the Adam optimizer is used to optimize the loss function, and the training platform uses an Intel Core i7-12700F processor and an NVIDIA GeForce RTX3060 12GB graphics card. After each round of training, the loss value results of training and validation are retained.

[0062] After 50 rounds of iterative calculations, the loss values of the model on the training and validation sets have converged. Subsequently, the trained SAM2-Unet network model was used to test the test set data. To comprehensively evaluate the model performance, in addition to the Dice coefficient and recall rate, the Mean Intersection over Union (MIoU) and Mean Pixel Accuracy (MPA) are also introduced as evaluation metrics. The Mean Intersection over Union refers to the ratio of the intersection and union of two images of the ground truth and the prediction, and the Mean Pixel Accuracy is the proportion of the accurately recognized pixels in each category on average among all categories. The calculation formulas for the Mean Intersection over Union and the Mean Pixel Accuracy are as follows:

[0063]

[0064] In the formula: k is the number of categories; p ij represents the number of the j-th category predicted as the i-th category. The test results of the SAM2-Unet network model on the test set are shown in Table 1.

[0065] It can be seen from the test results that the Dice coefficient of the SAM2-Unet network model in identifying underground road cavities is 0.9876, the recall rate is 0.9937, the intersection over union and pixel accuracy are 0.9757 and 0.9980 respectively. Overall, the average intersection over union and average pixel accuracy of the model reach 0.9867 and 0.9980 respectively, and the average Dice coefficient and average recall rate are 0.9933 and 0.9960. To further display and evaluate the recognition results of the SAM2-Unet network model, some image samples are randomly selected from the test set of road underground cavities for testing. The SAM2-Unet network model is used to perform semantic segmentation on the location and size of the underground cavities, and the segmentation results are shown in Figure 4 , where Figure 4 (a) is the ground penetrating radar image, Figure 4 (b) is the label image, Figure 4 (c) is the model segmentation result.

[0066] Table 1 Test results of the SAM2-Unet network model

[0067]

[0068] From Figure 4 it can be seen that the SAM2-Unet network model can accurately identify and extract the underground road cavities in the ground penetrating radar image. The average intersection over union and average pixel accuracy of the network model are both close to 0.98, indicating that it can accurately extract the size of the underground cavities in the radar image. Therefore, the SAM2-Unet network model meets the requirements of rapid identification and quantitative analysis of underground cavities in engineering.

[0069] Example 1

[0070] GprMax forward simulation

[0071] FDTD is an important method in the numerical calculation of electromagnetic fields, and its core is the difference solution based on Maxwell's equations. By transforming the continuous electromagnetic field problem into a difference equation system on a discrete grid, the numerical simulation of the spatio-temporal evolution of the electromagnetic field is realized. The electromagnetic phenomena at the macroscopic scale can be described by a set of Maxwell's equations, and the expression of the first-order partial differential is:

[0072]

[0073] Where: E is the electric field strength; H is the magnetic field strength; B is the magnetic induction intensity; D is the displacement vector of the point; t is the electromagnetic induction time; q v is the charge density; J is the current density.

[0074] In the implementation process of the FDTD method, the YEE grid structure plays a key role. The YEE grid structure is as Figure 5 shown. By dividing the space into regular cubic grids, defining the electric and magnetic field components at each grid node respectively, and using the central difference scheme to discretize the time and space derivatives in the Maxwell equations, it not only ensures the staggered distribution of the electromagnetic field components in space, but also can effectively maintain the consistency of the discretized equations with the original Maxwell equations in terms of physical essence.

[0075] To ensure the numerical stability and calculation accuracy of the FDTD algorithm, parameter selection is crucial. The spatial grid size needs to meet the requirement of being less than one-tenth of the minimum wavelength to effectively suppress the dispersion error. The selection of the time step must follow the Courant-Friedrichs-Lewy (CFL) stability condition:

[0076]

[0077] Where: c represents the propagation speed of electromagnetic waves in the medium.

[0078] The Kirchhoff migration imaging principle is based on the high-frequency approximate solution of the wave equation. By using the geometric ray theory, it calculates the superposition of diffracted waves received at each point underground from all possible reflection points to reconstruct the image of the underground structure. During the migration process, it is necessary to construct the velocity model of the underground medium and perform geometric ray tracing based on this model to determine the path and time for waves to propagate from each point underground to the surface receiving point. For each data point received on the surface, according to the geometric ray tracing results and the migration principle, the contributions of reflected waves from different depths and directions underground are superimposed, so as to achieve the accurate relocation of the reflected waves and generate a high-precision image of the underground target.

[0079] Through the 3D forward simulation data of GprMax, the propagation process of electromagnetic waves in the underground 3D space can be simulated, 3D electromagnetic wave field data containing the information of the underground structure of the road can be generated, and slice processing in the X and Y directions is performed on it to obtain 2D B-scan profiles in multiple directions. A 3D velocity model is constructed according to the actual situation or prior knowledge of the underground medium. Based on the 3D velocity model, geometric ray tracing calculations are carried out to determine the exact propagation paths and times of waves from various underground points to the surface receiving points. The results of geometric ray tracing will directly affect the migration accuracy of reflected waves. After completing the geometric ray tracing, for each data point received on the surface, using the 3D Kirchhoff migration algorithm, according to the path and time information of geometric ray tracing, the contributions of diffracted waves from all possible underground reflection points are calculated and superimposed. Considering factors such as the wave propagation path, velocity changes, and wavefront curvature, ensure that the reflected waves can be accurately migrated to their true underground positions. Through the migration calculations of all surface receiving points, a 3D migration imaging result of the underground structure is generated, reflecting the spatial positions of underground targets and revealing their shapes (see Figure 6 ).

[0080] Analysis of test results

[0081] To systematically verify the technical effects of the underground cavity detection method of the present invention, an outdoor physical model test was conducted in this study. During the test, a cubic cavity was pre-buried in the model sand box. The upper surface of the model is a rectangle of 1.25m×1.00m, the size of the cubic cavity is 0.20m in length, width, and height, and the depth of the upper interface of the cavity from the simulated road surface is 0.1m. The detection system is configured with an Italian IDS K2 ground penetrating radar. Through parameter optimization experiments, the combination of a 1600MHz center frequency and a 15ns sampling time window is determined to be the best, and the number of sampling points is 1024. Scanning is carried out along the pre-arranged survey lines to ensure that the center position of the ground penetrating radar antenna is aligned with the survey lines. The survey network layout adopts an orthogonal matrix design, with 18 horizontal survey lines (spacing 0.5m) and 24 vertical survey lines (spacing 0.5m). Through comparative experiments, it is verified that this density can balance the detection efficiency and accuracy requirements.

[0082] In this embodiment, precise detection of underground cavities is achieved through a systematic signal processing process. In the standardized preprocessing stage, the Reflexw 5.6 professional software is used to perform zero-offset correction, direct wave elimination, and background removal, and a 3D data volume is constructed. In the core processing link, the method proposed by the present invention is used to perform processing such as time exponential gain and improved 3D Kirchhoff migration imaging on the radar data. Since the environment where the measured data is located is relatively complex and there is a lot of clutter interference in the original data, a filtering and noise reduction processing link is additionally added in the processing process to ensure that the position of the finally 3D reconstructed cavity is clear and accurate, and to minimize the adverse effects of noise on the results. The preprocessing results of the measured data are shown inFigure 7 , where Figure 7 (a) is the data before filtering, Figure 7 (b) is the data after Bior2.2 wavelet adaptive threshold filtering, Figure 7 (c) is the data after Gaussian smoothing.

[0083] It can be clearly observed that Figure 7 by using the bior2.2 adaptive wavelet threshold filtering method, most of the clutter in the background is effectively filtered out, and the main reflection signal characteristics of the cavity are successfully retained. The data of each survey line are uniformly processed by wavelet adaptive threshold filtering technology to further optimize the data quality. After filtering, the Gaussian smoothing technology is used to optimize the image to make the cavity signal clearer and more prominent. In order to further accurately extract the main shape characteristics from the processed three-dimensional volume data, the method of extracting the isosurface used in the simulation is used to deeply process the three-dimensional volume data. The three-dimensional extraction results and positions of the main characteristics of the measured cubic cavity are shown as Figure 8 shown. It can be intuitively observed from Figure 8 the accurate presentation of the cavity characteristics after processing and their distribution in space, which helps to deeply understand the key information such as the shape and position of the cavity.

[0084] By deeply analyzing Figure 8 the three-dimensional reconstructed underground cavity in, it is found that the method proposed in the present invention can effectively extract the signal characteristics of the measured cubic cavity in three key dimensions: the position, trend and depth of the cavity. In terms of position, the spatial coordinates of the reconstructed cavity are highly consistent with the position of the cavity in the actual model, and the deviation is controlled within a very small range, accurately reflecting the lateral and longitudinal distribution of the cavity in the subgrade. In terms of trend, the reconstructed result clearly shows the extension direction consistent with the actual cavity, achieving accurate reproduction and truly restoring the extension trend of the cavity in three-dimensional space. In the depth direction, the vertical position of the reconstructed cavity is highly consistent with the depth information of the actual model, and the depth error is extremely small, accurately depicting the position characteristics of the cavity at the depth level of the subgrade. The three-dimensional characteristics shown by the reconstructed signal are highly similar to the characteristics of the actual model, strongly proving the efficiency and accuracy of the method of the present invention in extracting and restoring the cavity signal characteristics.

[0085] It is hereby stated that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the present technical solution, and they should all be covered within the scope of the claims of the present invention.

Claims

1. A road cavity detection method based on the fusion of 3D forward simulation of ground-penetrating radar and SAM2-Unet, characterized in that The following steps are involved: S1. Use the ground penetrating radar system to transmit electromagnetic pulse waves to the road to be inspected, receive the reflected signal and generate a ground penetrating radar image, and analyze the amplitude, frequency change and abnormal distribution characteristics of the signal; Among them, the ground penetrating radar system uses a pseudo-three-dimensional detection method to detect road holes; S2. preprocessing the data collected in step S1 to avoid confusion or weakening of the interpretation of the overall signal characteristics; S3. According to the actual situation or prior knowledge of the underground medium, a three-dimensional velocity model of the underground medium is constructed based on the finite-difference time-domain method (FDTD), and the numerical stability is ensured by YEE grid discretization; S4. Based on the three-dimensional velocity model constructed in step S3, a geometric ray tracing calculation is performed to determine the propagation path and time of the wave from each underground point to the surface receiving point; S5. After the geometric ray tracing calculation is completed, for each data point received on the surface, the three-dimensional Kirchhoff migration algorithm is used to calculate and superimpose the diffraction wave contributions from all reflection points underground according to the path and time information of the geometric ray tracing; S6. Input the migration result image into the pre-trained SAM2-Unet network for high-precision segmentation; Among them, the encoding-decoding architecture of the SAM2-Unet deep learning model integrates a multi-scale dilated convolution module, which can automatically extract the geometric features of underground targets; S7. Combined with the output results of deep learning to verify and further quantify the three-dimensional cavity model, the wave propagation path, velocity change and wavefront curvature are comprehensively considered to perform offset calculations on all surface receiving points to generate three-dimensional offset imaging results of the underground structure, thereby reflecting the spatial position of the underground cavity and revealing the morphology and properties of the cavity.

2. The road cavity detection method based on the fusion of 3D forward modeling of ground penetrating radar and SAM2-Unet according to claim 1, characterized in that: In step S1, the pseudo three-dimensional detection method is to perform detection through multiple parallel measurement lines or regular grid measurement lines, and use the coordinate information in the ground penetrating radar file to merge the radar data on different measurement lines into a three-dimensional data volume.

3. A road cavity detection method based on the fusion of 3D forward modeling of ground penetrating radar and SAM2-Unet according to claim 2, characterized in that: In step S2, the data preprocessing includes exponential gain and direct wave removal.

4. A road cavity detection method based on the fusion of 3D forward modeling of ground penetrating radar and SAM2-Unet according to claim 3, characterized in that: In step S7, the specific steps of reflecting the location of the underground cavity space are: I. Use the isosurface feature extraction method to accurately reconstruct the shape of the cavity, so that the shape and position of the cavity can be presented in the visual image; II. Quantitatively evaluate the accuracy of the reconstructed position, obtain the vertical parameters of the cavity based on wavelet time energy density analysis, and verify the height measurement accuracy by combining deep learning; calculate the three-dimensional volume by spatial integration of the top-view projection area and the vertical parameters; III. Use the model for verification to further ensure the practicality and reliability of the results; The SAM2-Unet network model is used to train the radar image dataset after the migration of underground voids in the road. After the training is completed, the migration data of the surface receiving points are imported into the SAM2-Unet network model to obtain the segmentation result of the void boundary of the underground structure, thereby reflecting the spatial position of the underground void.

5. A road cavity detection method based on the fusion of 3D forward modeling of ground penetrating radar and SAM2-Unet according to claim 4, characterized in that: During the training process of the SAM2-Unet network model: The number of iterations is 50, the batch size is 4, and the initial learning rate is 0.001; When updating the network weights, the Adam optimizer is used to optimize the loss function, and the training platform is based on an Intel Core i7-12700F processor and an NVIDIA GeForce RTX3060 12GB graphics card.

Images