Ground penetrating radar underground cavity intelligent detection model, method and system based on dynamic characteristic calibration

Abstract

The invention discloses a ground penetrating radar underground cavity intelligent detection model, method and system based on dynamic characteristic calibration, and the model comprises an encoder, a decoder and a dynamic characteristic calibration module. The dynamic feature calibration module is arranged in a jump connection feature transmission path between the encoder and the decoder, and is used for modeling feature dependency relationships between channels and space through a channel attention branch and a space attention branch respectively, and carrying out dynamic calibration on features extracted by the encoder; and the calibrated features are transmitted to a decoder for feature fusion. The detection model disclosed by the invention has the following advantages: (1) high detection precision: the dynamic feature calibration module can adaptively focus on the features of the cavity target and effectively suppress background clutter and noise, so that the model is significantly superior to the existing mainstream model in key indexes such as average Dice coefficient, recall rate and the like; (2) the small target detection capability is strong: the attention and segmentation capability on small-scale holes and weak echo signals is enhanced, and the small target IoU index is ahead of the comparison model;

Description

Technical Field

[0001] This invention belongs to the field of image recognition in key technologies of artificial intelligence, and specifically relates to an intelligent detection model, method and system for underground cavities based on dynamic feature calibration of ground penetrating radar, which can perform image semantic segmentation on feature maps obtained by ground penetrating radar. Background Technology

[0002] With the continuous expansion of urban underground space development, collapse accidents caused by road defects such as underground cavities and loose soil are occurring frequently, posing a serious threat to infrastructure and public safety. Ground penetrating radar (GPR), as an efficient and non-destructive detection tool, can reconstruct the three-dimensional distribution of underground media by emitting high-frequency electromagnetic waves and receiving reflected signals, and has become the mainstream method for detecting underground cavities.

[0003] Traditional 3D GPR data processing methods rely on manual interpretation or conventional image processing algorithms. However, in complex noisy environments, with weak echo signals, and multiple scale targets (from centimeter-level cracks to meter-level cavities), it is difficult to simultaneously guarantee detection accuracy and robustness. In recent years, deep learning techniques, especially U-Net and its 3D variant (3D U-Net), have been widely used in medical image segmentation and have been introduced into GPR data interpretation. However, directly applying existing models to 3D GPR underground cavity detection faces the following challenges:

[0004] (1) Insufficient feature utilization: The standard 3D U-Net has a fixed skip connection method, which cannot adaptively fuse shallow detail features of the encoder with deep semantic features of the decoder, resulting in insufficient perception of weak echoes and small-scale holes.

[0005] (2) Severe noise interference: GPR data has a low signal-to-noise ratio, and background clutter and noise can overwhelm weak signal targets. Existing methods lack effective mechanisms to suppress noise and enhance target response.

[0006] (3) Extreme class imbalance: Hollow voxels account for a very small proportion in 3D data (usually less than 5%), and model training is very likely to be biased towards the background, resulting in a high false negative rate.

[0007] (4) Scarcity of labeled data: It is costly to obtain large-scale, high-quality 3D GPR labeled data, the model is prone to overfitting and has poor generalization ability.

[0008] Existing technologies such as DCSAU-Net, UNETR++, and UXNet3D either have static attention mechanisms, excessively high computational complexity, or fail to jointly optimize channel and spatial dimension features, and thus cannot effectively solve the unique problems of GPR data within an end-to-end framework. Summary of the Invention

[0009] This invention addresses the problems of poor adaptability of existing 3D image segmentation models to multi-scale cavities in ground-penetrating radar (GPR) data, weak echo signal acquisition capability, high false detection and false negative rates in noisy backgrounds, and insufficient segmentation accuracy of small targets. It provides a ground-penetrating radar underground cavity intelligent detection model, method, and system based on dynamic feature calibration, which can be used for voxel-level accurate segmentation and identification of 3D ground-penetrating radar cavity targets in urban underground space safety assessment.

[0010] The present invention adopts the following technical solution:

[0011] An intelligent detection model for underground cavities based on ground-penetrating radar (GPR) with dynamic feature calibration is improved by including an encoder, a decoder, and a dynamic feature calibration module. The encoder receives data collected by GPR. The dynamic feature calibration module is set in the feature transmission path of the jump connection between the encoder and the decoder. It models the feature dependencies between channels and in space through channel attention branches and spatial attention branches, respectively, to dynamically calibrate the features extracted by the encoder. The calibrated features are then passed to the decoder for feature fusion. The dynamic calibration generates channel attention weights based on the global channel statistics of the features and spatial attention weights based on the spatial context information of the features. The channel attention weights and spatial attention weights are fused to form dynamic calibration weights, which are then used to weight the features. The decoder outputs the underground cavity detection results.

[0012] Furthermore, the encoder contains four downsampling stages, each using residual blocks UnetResBlock for downsampling, and each stage uses 3x3x3 convolution, batch normalization BatchNorm, and ReLU activation function;

[0013] The specific structure of the residual block UnetResBlock is as follows: it contains two consecutive 3×3×3 convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The stride of the first convolutional layer is set to 2, and the stride of the second convolutional layer remains at 1. To match the size and number of channels of the downsampled feature map, the residual block UnetResBlock includes a shortcut connection path consisting of a 1×1×1 convolutional layer and a batch normalization layer. This shortcut connection path is only activated when the number of channels or spatial size of the input and output are inconsistent, and its convolutional stride is synchronized with the first convolutional layer of the main path.

[0014] Furthermore, the channel attention branch uses adaptive 3D average pooling to capture the global context; the spatial attention branch uses asymmetric 3D convolution kernels to focus on local spatial relationships.

[0015] Furthermore, channel attention weights The calculation method is as follows:

[0016]

[0017] In the above formula, For input features, Convolution kernel, The function is Sigmoid, and AdaptiveAvgPool3D is an adaptive 3D average pooling function.

[0018] Spatial attention weights The calculation method is as follows:

[0019]

[0020] In the above formula, Conv3D is three-dimensional convolution, concat is information concatenation, AvgPool is average pooling, and MaxPool is max pooling.

[0021] Dynamic calibration weights The calculation method is as follows:

[0022]

[0023] In the above formula, This is element-wise multiplication.

[0024] Furthermore, in the decoder, the upsampling block UnetUpBlock is used to achieve upsampling through transposed convolution. The upsampled feature map is concatenated with the weighted features from the dynamic feature calibration module along the channel dimension and then fused in subsequent convolutional layers.

[0025] The specific structure of the upsampling block UnetUpBlock is as follows: First, a 3D transposed convolutional layer with a kernel size of 2×2×2 and a stride of 2 is used to upsample the depth, height, and width dimensions of the input feature map to twice their original size. Then, the upsampled feature map is concatenated with the corresponding layer-weighted feature map from the dynamic feature calibration module along the channel dimension. The concatenated feature map is then fed into a feature fusion unit, which consists of two consecutive 3×3×3 convolutional layers. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function.

[0026] Furthermore, the model is trained using the following loss function:

[0027]

[0028] In the above formula, and These are weighting coefficients. + =1, 0.5≤ ≤0.8;

[0029]

[0030] In the above formula, It is the predicted probability of the i-th pixel. It is the true label of the i-th pixel. It is the number of all pixels in the image. It is a constant;

[0031]

[0032] In the above formula, It is a balancing factor. It represents the probability of predicting the correct category when the true label is 1. ,otherwise ; It is the focus parameter.

[0033] A ground-penetrating radar intelligent detection method for underground cavities based on dynamic feature calibration, using the above-mentioned model, is improved by including the following steps:

[0034] Step 1, Data Preparation:

[0035] The collected 3D ground-penetrating radar volume data is normalized.

[0036] Step 2, Model Training:

[0037] The model is trained using the AdamW optimizer, employing the ReduceLROnPlateau learning rate scheduling strategy and early stopping mechanism. The loss function is... and The weighted sum, Weight 0.7, Weight 0.3;

[0038] Step 3, Result Output:

[0039] When receiving new 3D ground-penetrating radar volume data, the system automatically calls the trained model for inference, outputs binary segmentation results, and further calculates the volume and location of cavities. Finally, it displays the detection report and 3D rendering to the user through a visualization interface.

[0040] Furthermore, in step 1, the data is also enhanced, and the enhancement includes at least one of the following:

[0041] Spatial geometry enhancement: random flipping, elastic deformation, affine transformation;

[0042] Intensity Enhancement: Add Gaussian noise, adjust contrast, and scale intensity.

[0043] An intelligent detection system for underground cavities based on ground-penetrating radar (GPR) with dynamic feature calibration is improved by including: a data acquisition module for acquiring three-dimensional GPR volume data; a dynamic feature calibration model, including an encoder, a decoder, and a dynamic feature calibration module, wherein the dynamic feature calibration module is configured to dynamically calibrate the features extracted by the encoder and pass the calibrated features to the decoder for feature fusion; and a segmentation output module for outputting the three-dimensional cavity segmentation results corresponding to the three-dimensional GPR volume data.

[0044] Furthermore, it also includes: a data preprocessing module for normalizing the three-dimensional ground-penetrating radar volume data; and a result postprocessing module for quantitative analysis and visualization of the three-dimensional cavity segmentation results.

[0045] The beneficial effects of this invention are:

[0046] The detection model disclosed in this invention has the following advantages:

[0047] (1) High detection accuracy: The dynamic feature calibration module can adaptively focus on the features of the hole target and effectively suppress background clutter and noise, making the model significantly better than the existing mainstream models in key indicators such as average Dice coefficient and recall rate, with the highest Dice coefficient reaching 78.01%.

[0048] (2) Strong small target detection capability: The introduction of dynamic feature calibration mechanism and Dice-Focal joint loss function enhances the ability to focus on and segment small-scale holes and weak echo signals, and the small target IoU index leads the comparison model.

[0049] The detection method disclosed in this invention has good generalization performance. By combining multiple data augmentation strategies, it reduces the dependence on large-scale labeled data and improves the stability and generalization ability of the detection model under complex geological conditions and noisy environments.

[0050] The detection system disclosed in this invention is highly practical and provides a complete solution from data processing and model training to detection visualization, providing an efficient and reliable technical tool for safety assessment of urban underground space. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the dynamic feature calibration principle of the detection model disclosed in this embodiment;

[0052] Figure 2 This is a schematic diagram of the overall network architecture of the detection model disclosed in this embodiment;

[0053] Figure 3 This is a schematic diagram of the dual-branch structure of the dynamic feature calibration module disclosed in this embodiment;

[0054] Figure 4 This is a comparison diagram of the segmentation effect of the detection model disclosed in this embodiment and other three-dimensional image segmentation models on a typical cross section;

[0055] Figure 5 Figure I shows a comparison of the 3D rendering effects between the detection model disclosed in this embodiment and other 3D image segmentation models.

[0056] Figure 6 Figure II shows a comparison of the 3D rendering effects between the detection model disclosed in this embodiment and other 3D image segmentation models.

[0057] Figure 7 This is a flowchart illustrating the detection method disclosed in this embodiment;

[0058] Figure 8 This is the result after data augmentation;

[0059] Figure 9 This is a comparison chart showing the effects of combining parameters of the joint loss function. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0061] Example 1 discloses an intelligent detection model for underground cavities using ground-penetrating radar based on dynamic feature calibration, such as... Figure 1 As shown, the system includes an encoder, a decoder, and a dynamic feature calibration module. The encoder receives data acquired by ground-penetrating radar. The dynamic feature calibration module is located in the feature transfer path of the jump connection between the encoder and the decoder. It models the feature dependencies between channels and in space through channel attention branches and spatial attention branches, respectively, to dynamically calibrate the features extracted by the encoder. The calibrated features are then passed to the decoder for feature fusion. The dynamic calibration generates channel attention weights based on the global channel statistics of the features and spatial attention weights based on the spatial context information of the features. The channel attention weights and spatial attention weights are fused to form dynamic calibration weights, which are then used to weight the features. The decoder outputs the underground cavity detection results.

[0062] This embodiment uses an improved 3D U-Net network model as an example, referred to as the DCAM U-Net network model. The improved 3D U-Net network model is based on an improved three-dimensional U-Net architecture and incorporates a three-dimensional dynamic contextual attention mechanism and a multi-scale feature fusion mechanism. It should be noted that the dynamic feature calibration mechanism proposed in this invention is also applicable to other types of segmentation network architectures.

[0063] Construct a symmetric encoder-decoder structure DCAM U-Net network model. The encoder contains four downsampling stages. Each stage uses residual blocks UnetResBlock for downsampling. Each stage uses 3x3x3 convolution, batch normalization BatchNorm, and ReLU activation function.

[0064] The specific structure of the residual block UnetResBlock is as follows: it contains two consecutive 3×3×3 convolutional layers, each followed by a batch normalization layer and a ReLU activation function. To achieve downsampling at a specific stage, the stride of the first convolutional layer is set to 2, while the stride of the second convolutional layer remains at 1. Simultaneously, to match the size and number of channels of the downsampled feature map, the residual block UnetResBlock includes a shortcut connection path consisting of a 1×1×1 convolutional layer and a batch normalization layer. This shortcut connection path is only activated when the number of channels or spatial size of the input and output are inconsistent. Its convolutional stride is synchronized with the first convolutional layer of the main path to ensure that the residual addition operation is executed correctly.

[0065] In the decoder, the upsampling block UnetUpBlock is used to achieve upsampling through transposed convolution. The upsampled feature map is concatenated with the weighted features from the dynamic feature calibration module along the channel dimension and then fused in subsequent convolutional layers to ensure the organic combination of multi-scale features and the preservation of hole details.

[0066] The specific structure of the upsampling block UnetUpBlock is as follows: First, a 3D transposed convolutional layer with a kernel size of 2×2×2 and a stride of 2 is used to upsample the depth, height, and width dimensions of the input feature map to twice their original size. Then, the upsampled feature map is concatenated with the corresponding layer-weighted feature map from the dynamic feature calibration module along the channel dimension. The concatenated feature map is then fed into a feature fusion unit, which consists of two consecutive 3×3×3 convolutional layers. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function to complete the deep fusion of cross-scale features.

[0067] Dynamic skip connection layers: such as Figure 2As shown, in order to improve the effectiveness of feature transfer between encoder and decoder, a dynamic feature calibration module (3D-DCAM) is embedded at the jump connection to dynamically calibrate the feature fusion process, suppress background noise interference, and enhance hole-related information.

[0068] The feature maps generated by the encoder are first weighted by 3D-DCAM, and then passed to the corresponding decoder for fusion via skip connections. This weighting process can dynamically adjust channel and spatial attention, effectively suppress irrelevant background features, dynamically balance shallow details (such as crack edges) and deep semantics (such as the overall outline of holes), and optimize the feature fusion effect.

[0069] like Figure 3 As shown, the dynamic feature calibration module consists of two branches: channel attention and spatial attention. The channel attention branch uses adaptive 3D average pooling to dynamically capture global context statistics; the spatial attention branch uses asymmetric bi-branch 3D convolutional kernels to focus on local spatial correlations. To address the rapid signal attenuation in the depth direction of ground-penetrating radar data, an asymmetric 3D convolutional kernel is designed to reduce the computational load in the depth dimension, improving overall efficiency while maintaining feature extraction capabilities.

[0070] Collaborative weighting: By coupling channel and spatial attention weights to input features through element-wise multiplication, the collaborative enhancement of the two is achieved, which suppresses background noise and significantly improves the detection response to multi-scale targets.

[0071] Channel attention weight The calculation method is as follows:

[0072]

[0073] In the above formula, For input features, , Convolution kernel, The function is Sigmoid, and AdaptiveAvgPool3D is an adaptive 3D average pooling function.

[0074] Spatial attention weights The calculation method is as follows:

[0075]

[0076] In the above formula, Conv3D is three-dimensional convolution, concat is information concatenation, AvgPool is average pooling, and MaxPool is max pooling.

[0077] Dynamic calibration weights The calculation method is as follows:

[0078]

[0079] In the above formula, This is element-wise multiplication.

[0080] The model is trained using the Dice-Focal joint loss function:

[0081] In 3D ground-penetrating radar data, void voxels often account for only a small proportion, and the class imbalance causes the model to prioritize background optimization during training, thus ignoring the detection of small targets. To address this, this invention innovatively introduces the Dice–Focal joint loss function, which combines the advantages of Dice Loss and Focal Loss. During the optimization process, it not only improves the overall segmentation performance but also strengthens the focus on sparse, weak-echo voids.

[0082] (1) Dice Loss

[0083] Dice Loss is a metric used to evaluate the overlap between two binary sets (predicted results and true labels). In measured ground-penetrating radar data, void regions are relatively small, while background regions constitute the majority. For void targets with sparse volume and rich boundary details, Dice Loss can effectively suppress background interference, allowing the model to focus more on covering small voxel blocks and encouraging the model to detect as many small void targets as possible. It is defined as:

[0084]

[0085] In the above formula, A is the predicted foreground region, and B is the ground truth foreground region. Dice Loss measures the model's performance by calculating the overlap between the predicted and ground truth label regions. The formula for Dice Loss is as follows:

[0086]

[0087] In 3D ground-penetrating radar data, holes are relatively small and typically appear as localized, low-intensity areas. The core objective of DiceLoss is to maximize the overlap between predicted and true labels, thereby encouraging the model to detect holes more accurately during training. When the model can identify more hole regions, the Dice coefficient will be higher, thus optimizing the model's performance in small target detection. To avoid division by zero, a small constant is added to the denominator. :

[0088]

[0089] (2) Focal Loss

[0090] Focal Loss is an extension of standard cross-entropy loss that increases attention to difficult-to-classify samples by adjusting the weights of easily classified samples. In 3D ground-penetrating radar hole detection, most data belongs to the background class, while the hole target only has a few pixels or voxels. During training, predicting background regions is usually very easy, while predicting hole regions is more difficult, and the model tends to ignore these targets. The formula for Focal Loss is as follows:

[0091]

[0092] In the task of hole detection using 3D ground-penetrating radar data, Focal Loss helps the model focus on harder-to-detect hole targets by increasing the weight of difficult-to-classify hole regions and decreasing the weight of background regions. This helps improve the model's ability to detect small target regions, especially when hole targets are very scarce in the dataset.

[0093] (3) Dice-Focal joint loss function

[0094] The Dice-Focal joint loss function combines the advantages of Dice Loss and Focal Loss, weighting them together to allow the model to focus on edge details and small targets while maintaining overall coherence of the holes. Specifically:

[0095] Dice Loss ensures that the model can accurately capture small, hollow target regions by maximizing the overlap between the predicted results and the true labels.

[0096] Focal Loss helps the model better detect those hard-to-identify hollow targets by reducing attention to background classes and increasing attention to hard-to-classify hollow regions.

[0097] The formula for the Dice-Focal joint loss function is:

[0098]

[0099] In the above formula, It's Dice Loss, which helps the model optimize for hollow target regions. It is FocalLoss, which increases attention to empty targets by reducing the weight given to the background; and It is a weighting coefficient used to adjust the contributions of the two. + =1, 0.5≤ ≤0.8;

[0100]

[0101] In the above formula, It is the predicted probability of the i-th pixel. It is the true label of the i-th pixel. It is the number of all pixels in the image. It is a constant;

[0102]

[0103] In the above formula, It is a balancing factor used to adjust the importance of different categories; It represents the probability of predicting the correct category when the true label is 1. ,otherwise ; It is the focus parameter, usually set to 2, and is used to adjust the influence between easily classified and difficult-to-classify samples.

[0104] By using the Dice-Focal joint loss function, the model can not only maximize the overlap between the hole region and the real label, but also continuously pay attention to those hole targets that are difficult to detect due to weak echoes or sparse voxels during training, which significantly improves the detection accuracy of small targets and the overall segmentation performance.

[0105] This embodiment also discloses a ground-penetrating radar intelligent detection method for underground cavities based on dynamic feature calibration, such as... Figure 7 As shown, using the above model, the following steps are included:

[0106] Step 1, Data Preparation:

[0107] The collected 3D ground-penetrating radar volume data is normalized.

[0108] Data augmentation: During training, spatial geometry augmentation (random flipping, elastic deformation, affine transformation) and / or intensity augmentation (adding Gaussian noise, contrast adjustment, intensity scaling) strategies are used in real time to expand the dataset to improve the model's robustness to complex environments and noise.

[0109] like Figure 8 As shown, the enhancement operation can significantly improve the diversity of the training set while preserving the true features of the holes, providing a solid data foundation for the model to cope with weak echoes, low contrast and complex geological environments.

[0110] Step 2, Model Training:

[0111] The network model was trained using the dataset processed in step 1. The model was trained using the AdamW optimizer (initial learning rate 5e-4, weight decay 1e-4), employing the ReduceLROnPlateau learning rate scheduling strategy and early stopping mechanism. The loss function was... and The weighted sum, Weight 0.7, With a weight of 0.3 and a batch size of 2, the training run on the GPU lasts approximately 150 epochs. Figure 9 This is a comparison chart showing the effects of combining parameters of the joint loss function.

[0112] The model is evaluated using a five-fold cross-validation strategy, and the model parameters are optimized using a backpropagation algorithm until the model converges, resulting in a well-trained underground cavity detection model.

[0113] The trained model is evaluated on the test set. (See attached image) Figure 4 , 5 As shown in Tables 6 and 7 below, the DCAMU-Net model in this embodiment achieves the highest average Dice coefficient, maximum Dice coefficient, and recall rate (73.90%, 78.01%, and 79.73%, respectively), while having the lowest false positive rate (0.66%) and the highest IoU for small targets (20.32%). Its overall performance is significantly better than the other five 3D image segmentation models compared.

[0114]

[0115] Step 3, Result Output:

[0116] When receiving new 3D ground-penetrating radar volume data, the system automatically calls the trained DCAM U-Net model for inference, outputs a probability map of each voxel belonging to a cavity, and after thresholding, obtains the 3D cavity binary segmentation result. It then further calculates indicators such as cavity volume and location, and finally displays the detection report and 3D rendering map to the user through a visualization interface, providing a basis for road safety assessment.

[0117] This embodiment also discloses an intelligent detection system for underground cavities based on dynamic feature calibration using ground-penetrating radar (GPR), comprising: a data acquisition module for acquiring three-dimensional GPR volume data; a dynamic feature calibration model, including an encoder, a decoder, and a dynamic feature calibration module, wherein the dynamic feature calibration module is configured to dynamically calibrate the features extracted by the encoder and pass the calibrated features to the decoder for feature fusion; and a segmentation output module for outputting the three-dimensional cavity segmentation results corresponding to the three-dimensional GPR volume data. It also includes: a data preprocessing module for normalizing the three-dimensional GPR volume data; and a result post-processing module for quantitative analysis and visualization of the three-dimensional cavity segmentation results.

Claims

1. A ground-penetrating radar intelligent detection model for underground cavities based on dynamic feature calibration, characterized in that: The system includes an encoder, a decoder, and a dynamic feature calibration module. The encoder receives data collected by ground-penetrating radar. The dynamic feature calibration module is located in the feature transfer path of the jump connection between the encoder and the decoder. It models the feature dependencies between channels and in space through channel attention branches and spatial attention branches, respectively, to dynamically calibrate the features extracted by the encoder. The calibrated features are then passed to the decoder for feature fusion. The dynamic calibration generates channel attention weights based on the global channel statistics of the features and spatial attention weights based on the spatial context information of the features. The channel attention weights and spatial attention weights are fused to form dynamic calibration weights, which are then used to weight the features. The decoder outputs the underground cavity detection results.

2. The intelligent detection model for underground cavities based on dynamic feature calibration of ground penetrating radar according to claim 1, characterized in that: The encoder consists of four downsampling stages, each using residual blocks UnetResBlock for downsampling, and each stage uses 3x3x3 convolution, batch normalization BatchNorm, and ReLU activation function. The specific structure of the residual block UnetResBlock is as follows: it contains two consecutive 3×3×3 convolutional layers, each followed by a batch normalization layer and a ReLU activation function. The stride of the first convolutional layer is set to 2, and the stride of the second convolutional layer remains at 1. To match the size and number of channels of the downsampled feature map, the residual block UnetResBlock includes a shortcut connection path consisting of a 1×1×1 convolutional layer and a batch normalization layer. This shortcut connection path is only activated when the number of channels or spatial size of the input and output are inconsistent, and its convolutional stride is synchronized with the first convolutional layer of the main path.

3. The intelligent detection model for underground cavities based on dynamic feature calibration of ground penetrating radar according to claim 1, characterized in that: The channel attention branch uses adaptive 3D average pooling to capture the global context; the spatial attention branch uses asymmetric 3D convolution kernels to focus on local spatial relationships.

4. The intelligent detection model for underground cavities based on dynamic feature calibration of ground penetrating radar according to claim 1, characterized in that: Channel attention weight The calculation method is as follows: In the above formula, As input features, Convolution kernel, The function is Sigmoid, and AdaptiveAvgPool3D is an adaptive 3D average pooling function. Spatial attention weights The calculation method is as follows: In the above formula, Conv3D is three-dimensional convolution, concat is information concatenation, AvgPool is average pooling, and MaxPool is max pooling. Dynamic calibration weights The calculation method is as follows: In the above formula, This is element-wise multiplication.

5. The intelligent detection model for underground cavities based on dynamic feature calibration of ground penetrating radar according to claim 1, characterized in that: In the decoder, upsampling is achieved by transposed convolution using the upsampling block UnetUpBlock. The upsampled feature map is then concatenated with the weighted features from the dynamic feature calibration module along the channel dimension and fused in subsequent convolutional layers. The specific structure of the upsampling block UnetUpBlock is as follows: First, a 3D transposed convolutional layer with a kernel size of 2×2×2 and a stride of 2 is used to upsample the depth, height, and width dimensions of the input feature map to twice their original size. Then, the upsampled feature map is concatenated with the corresponding layer-weighted feature map from the dynamic feature calibration module along the channel dimension. The concatenated feature map is then fed into a feature fusion unit, which consists of two consecutive 3×3×3 convolutional layers. Each convolutional layer is followed by a batch normalization layer and a ReLU activation function.

6. The intelligent detection model for underground cavities based on dynamic feature calibration of ground penetrating radar according to claim 1, characterized in that: The model is trained using the following loss function: In the above formula, and These are weighting coefficients. + =1, 0.5≤ ≤0.8; In the above formula, It is the predicted probability of the i-th pixel. It is the true label of the i-th pixel. It is the number of all pixels in the image. It is a constant; In the above formula, It is a balancing factor. It represents the probability of predicting the correct category when the true label is 1. ,otherwise ; It is the focus parameter.

7. A ground-penetrating radar intelligent detection method for underground cavities based on dynamic feature calibration, using the model described in claim 1, characterized in that, Includes the following steps: Step 1, Data Preparation: The collected 3D ground-penetrating radar volume data is normalized. Step 2, Model Training: The model is trained using the AdamW optimizer, employing the ReduceLROnPlateau learning rate scheduling strategy and early stopping mechanism. The loss function is... and The weighted sum, Weight 0.7, Weight 0.3; Step 3, Output the results: When receiving new 3D ground-penetrating radar volume data, the system automatically calls the trained model for inference, outputs binary segmentation results, and further calculates the volume and location of cavities. Finally, it displays the detection report and 3D rendering to the user through a visualization interface.

8. The intelligent detection method for underground cavities based on dynamic feature calibration using ground-penetrating radar according to claim 7, characterized in that, In step 1, the data is further enhanced, including at least one of the following: Spatial geometry enhancement: random flipping, elastic deformation, affine transformation; Intensity Enhancement: Add Gaussian noise, adjust contrast, and scale intensity.

9. A ground-penetrating radar intelligent detection system for underground cavities based on dynamic feature calibration, characterized in that, include: The data acquisition module is used to acquire three-dimensional ground-penetrating radar volume data; The dynamic feature calibration model includes an encoder, a decoder, and a dynamic feature calibration module. The dynamic feature calibration module is configured to dynamically calibrate the features extracted by the encoder and pass the calibrated features to the decoder for feature fusion. The segmentation output module is used to output the three-dimensional cavity segmentation results corresponding to the three-dimensional ground-penetrating radar volume data.

10. The intelligent detection system for underground cavities based on dynamic feature calibration of ground-penetrating radar according to claim 9, characterized in that, Also includes: The data preprocessing module is used to normalize the three-dimensional ground-penetrating radar volume data; The result post-processing module is used to perform quantitative analysis and visualization of the three-dimensional hole segmentation results.

Images