An automatic extraction method, system, equipment, and medium for black soil layer interface based on improved U-Net ground penetrating radar

Abstract

This invention discloses a method, system, device, and medium for automatic extraction of black soil layer interfaces using ground-penetrating radar (GPR) based on an improved U-Net, belonging to the field of GPR data processing and intelligent interpretation technology. The method includes acquiring and preprocessing GPR B-scan image data, constructing a training dataset, building a layer extraction model based on an improved U-Net neural network model, constructing a hybrid loss function including at least two combinations of weighted binary cross-entropy loss, Dice loss, structural similarity loss, and smoothness loss, training the layer extraction model based on the training dataset and the hybrid loss function, and inputting the GPR B-scan image data to be detected into the trained layer extraction model to achieve automatic extraction of black soil layer interfaces using GPR. This invention effectively overcomes signal attenuation and noise interference in GPR images, achieving high-precision and continuous soil layer interface extraction.

Description

Technical Field

[0001] This invention relates to the field of ground-penetrating radar (GPR) data processing and intelligent interpretation technology, specifically to a method, system, device and medium for automatic extraction of black soil layer interface based on improved U-Net. Background Technology

[0002] Black soil resources are extremely valuable and crucial for ensuring food security. Ground penetrating radar (GPR), a non-destructive testing technology, effectively detects shallow subsurface structures by emitting high-frequency electromagnetic waves and receiving reflected signals from the interface of underground media. In black soil regions, using GPR to detect the thickness of the black soil layer and identify the interface between black soil and underlying loess is an important means of monitoring black soil loss and protecting the soil.

[0003] However, the automatic interpretation of GPR images (B-scan) faces many challenges. First, the difference in dielectric constant between black soil and loess may be small under certain humidity conditions, resulting in weak reflected signals. Second, GPR images are often subject to a large amount of clutter interference due to underground heterogeneous media (such as plant roots and rocks) and environmental noise. Third, due to signal attenuation or interference, the target stratum interface often appears as discontinuous and blurry lines in the image, rather than continuous and clear edges.

[0004] Traditional layer extraction methods mainly include manual extraction and semi-automatic methods based on edge detection (such as Canny and Sobel operators) or threshold segmentation. Manual extraction is inefficient and highly susceptible to subjective influence; traditional image processing methods have poor noise resistance and struggle to extract continuous layers in complex backgrounds. In recent years, deep learning methods, represented by U-Net, have achieved success in fields such as medical image segmentation. However, directly applying standard U-Net to GPR soil layer extraction often presents the following problems: 1) Due to the limited receptive field of the convolution kernel, it is difficult to connect broken layer signals; 2) The proportion of pixels at soil layer interfaces in GPR images is extremely low (extreme imbalance between positive and negative samples), causing the model to tend to predict the background; 3) The lack of constraints on the continuity of the stratigraphic structure makes the extracted interfaces prone to abrupt changes.

[0005] To address the aforementioned issues, there is an urgent need for an automatic extraction method, system, equipment, and medium for black soil layer interface based on improved U-Net ground penetrating radar, which can solve the problems existing in traditional methods. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic extraction method, system, device and medium for black soil layer interface based on improved U-Net ground penetrating radar, which can effectively overcome signal attenuation and noise interference in ground penetrating radar images and achieve high-precision and continuous soil layer interface extraction.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0008] An automatic extraction method for black soil layer interface based on ground-penetrating radar using an improved U-Net, comprising:

[0009] Step 1: Acquire ground-penetrating radar B-scan image data, preprocess it, and construct a training dataset;

[0010] Step 2: Construct a bit-layer extraction model based on the improved U-Net neural network model. The improved U-Net neural network model includes an encoder, a bottleneck layer, and a decoder. In each layer of the encoder and decoder, the traditional double convolutional layer is replaced by a residual dilated convolutional module, and attention gating is introduced at the skip connections of the decoder.

[0011] Step 3: Construct a hybrid loss function, which includes at least two combinations of weighted binary cross-entropy loss, Dice loss, structural similarity loss, and smoothness loss;

[0012] Step 4: Train the bit-layer extraction model based on the training dataset and the hybrid loss function;

[0013] Step 5: Input the ground-penetrating radar B-scan image data to be detected into the trained stratigraphic extraction model to achieve automatic extraction of the black soil layer interface by ground-penetrating radar.

[0014] Furthermore, in step 1, ground-penetrating radar B-scan image data is acquired and preprocessed to construct a training dataset, specifically as follows:

[0015] Acquire ground-penetrating radar B-scan image data;

[0016] The ground-penetrating radar B-scan image data is processed and augmented to obtain the input dataset;

[0017] Based on the layer annotations of the input dataset, Gaussian soft labels are generated using a Gaussian distribution function to enhance the labels and construct the training dataset.

[0018] Further processing of the ground-penetrating radar B-scan image data is as follows:

[0019] The ground-penetrating radar B-scan image data were sequentially subjected to zero-time correction, direct wave removal, time-varying gain processing, and amplitude normalization processing.

[0020] Furthermore, data augmentation is performed on the ground-penetrating radar B-scan image data, specifically as follows:

[0021] Data augmentation was performed on the processed ground-penetrating radar B-scan image data, including random left and right flipping, random signal strength scaling, and the addition of random Gaussian noise.

[0022] Furthermore, based on the layer annotations of the input dataset, Gaussian soft labels are generated using a Gaussian distribution function to enhance the labels and construct the training dataset, specifically as follows:

[0023] For each column of A-scan data in each image of the input dataset, obtain the manually annotated layer depth location;

[0024] Using the layer depth as the center, a probability distribution band is generated using a one-dimensional Gaussian function. The probability distribution band is then used as the label value of the A-scan data in that column to generate Gaussian soft labels.

[0025] Iterate through all images in the input dataset to generate the training dataset.

[0026] Furthermore, the residual dilated convolution module includes a dilated convolutional layer, a normalization layer, an activation layer, a regular convolutional layer, a normalization layer, and a residual connection path connected in sequence. The dilated convolutional layer is used to expand the receptive field without reducing the feature map resolution in order to capture the contextual information of the fracture interface. The residual connection path is used to add the input to the convolutional features to alleviate the gradient vanishing problem in deep network training.

[0027] Furthermore, the hybrid loss function is:

[0028] (1)

[0029] In the formula, For weighted binary cross-entropy loss, Used to measure the overlap between the predicted region and the actual region. For structural similarity loss, The smoothness loss is calculated using the Laplacian operator to constrain the continuity of the prediction results.

[0030] This invention also provides an automatic extraction system for the interface of black soil layer using ground penetrating radar based on an improved U-Net, applied to the aforementioned automatic extraction method for the interface of black soil layer using ground penetrating radar based on an improved U-Net, comprising:

[0031] The data acquisition module is used to acquire ground-penetrating radar B-scan image data;

[0032] The preprocessing module is used to preprocess the ground-penetrating radar B-scan image data and construct the training dataset;

[0033] The model building module is used to build a bit-layer extraction model based on the improved U-Net neural network model and to construct a hybrid loss function;

[0034] The model training module is used to train the bit-layer extraction model based on the hybrid loss function and the training dataset;

[0035] The interface extraction module is used to automatically extract the interface of the black soil layer from ground-penetrating radar based on the trained sub-layer extraction model.

[0036] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for automatic extraction of black soil layer interface based on improved U-Net ground penetrating radar.

[0037] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for automatic extraction of the black soil layer interface based on improved U-Net ground-penetrating radar.

[0038] In summary, the present invention has at least one of the following beneficial technical effects:

[0039] 1. This invention enhances the ability to extract discontinuous interfaces. By introducing a residual dilated convolution module, the receptive field of the neural network is significantly expanded, enabling the model to automatically connect soil interfaces that are broken due to signal attenuation or interference by utilizing the surrounding context information.

[0040] 2. This invention solves the problem of extreme sample imbalance by using a Gaussian soft label strategy combined with a positive sample weighted loss function, which effectively prevents the model from ignoring the extremely small proportion of soil interface pixels during training, thus greatly improving the recall rate.

[0041] 3. This invention improves positioning accuracy and structural preservation by introducing SSIM and attention mechanisms, making the model pay more attention to the structural and textural features of the strata and suppressing background noise. Experimental results show that the mean absolute error (MAE) of the layer depth extracted by this method is significantly reduced, which can meet the needs of refined soil exploration. Attached Figure Description

[0042] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0043] Figure 2 This is a schematic diagram of a binary hard label;

[0044] Figure 3 This is a schematic diagram of a Gaussian soft label.

[0045] Figure 4This is a schematic diagram comparing the probability distribution of a single channel;

[0046] Figure 5 A schematic diagram of the improved U-Net neural network model structure;

[0047] Figure 6 This is a schematic diagram illustrating the effect of black soil strata identification.

[0048] Figure 7 A schematic diagram of the consistency regression analysis for soil depth prediction;

[0049] Figure 8 This is a schematic diagram of a computer device structure. Detailed Implementation

[0050] 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. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0051] like Figure 1 As shown, this invention provides an automatic extraction method for the interface of black soil layer using ground-penetrating radar based on an improved U-Net, comprising:

[0052] Step 1: Acquire ground-penetrating radar B-scan image data, preprocess it, and construct a training dataset;

[0053] Step 2: Construct a bit-layer extraction model based on the improved U-Net neural network model. The improved U-Net neural network model includes an encoder, a bottleneck layer, and a decoder. In each layer of the encoder and decoder, the traditional double convolutional layer is replaced by a Res-Dilated Block, and an attention gate is introduced at the skip connection of the decoder.

[0054] Step 3: Construct a hybrid loss function, which includes at least two combinations of weighted binary cross-entropy loss, Dice loss, structural similarity loss, and smoothness loss;

[0055] Step 4: Train the bit-layer extraction model based on the training dataset and the hybrid loss function;

[0056] Step 5: Input the ground-penetrating radar B-scan image data to be detected into the trained stratigraphic extraction model to achieve automatic extraction of the black soil layer interface by ground-penetrating radar.

[0057] In step 1, ground-penetrating radar B-scan image data is acquired and preprocessed to construct a training dataset, specifically as follows:

[0058] Step 101: Acquire ground-penetrating radar B-scan image data;

[0059] Step 102: Process and augment the ground-penetrating radar B-scan image data to obtain the input dataset;

[0060] Step 103: Based on the layer annotations of the input dataset, Gaussian soft labels are generated using a Gaussian distribution function to enhance the labels and construct the training dataset.

[0061] In step 101, the ground-penetrating radar B-scan image data is acquired, specifically as follows:

[0062] Ground penetrating radar (GPR) simulation data can be obtained using gprMax simulation software, or measured data can be obtained using professional GPR. At least one type of data is required, and this invention does not impose any restrictions on it. The final result is GPR B-scan image data.

[0063] In step 102, the ground-penetrating radar B-scan image data is processed and augmented to obtain the input dataset, specifically as follows:

[0064] The ground-penetrating radar B-scan image data were sequentially subjected to zero-time correction, direct wave removal, time-varying gain processing, and amplitude normalization.

[0065] After processing, due to the non-stationarity of the GPR signal, data augmentation is performed on the processed ground-penetrating radar B-scan image data, including random left and right flipping, random signal strength scaling, and adding random Gaussian noise.

[0066] In step 103, based on the layer annotations of the input dataset, Gaussian soft labels are generated using a Gaussian distribution function to enhance the labels and construct the training dataset, specifically as follows:

[0067] Traditional binary labels (interface is 1, background is 0) are too sparse in GPR images and are greatly affected by human annotation errors. Therefore, this invention adopts a Gaussian soft labeling strategy.

[0068] For each column of A-scan data in each image of the input dataset, obtain the manually annotated layer depth location. ;

[0069] Based on stratigraphic depth Centered on the data, a probability distribution band is generated using a one-dimensional Gaussian function, and this probability distribution band is used as the label value for the A-scan data in that column. Generate Gaussian soft tags, where:

[0070] (1)

[0071] In the formula, For depth indexing, Using a standard deviation (e.g., 2.0 or 3.0) transforms a "single line" into a "probability band," making the model easier to converge and better reflecting the physical width of the radar wavelet. Figure 2 , Figure 3 and Figure 4 This invention demonstrates the generation principle and effect of Gaussian soft tags, wherein... Figure 2 For the original binary label: Traditional methods only mark the manually annotated layer positions as 1, and the rest of the background as 0, which is represented as a discrete thin line. This is not only too sparse, but also ignores the slight uncertainty of the layer position. Figure 3 As a Gaussian soft label, this invention uses a one-dimensional Gaussian function to smoothly expand the binary label, generating a probability distribution band (heatmap) with a certain width. Figure 4 Using a single-channel probability distribution, the changes in label values ​​of a certain column of data were compared. It can be seen that the soft label has the highest probability (1.0) at the center position and decreases Gaussian towards both sides. This design allows the model to learn the physical law that "the closer to the center, the higher the probability" during training, which effectively alleviates the problem of imbalance between positive and negative samples and improves the convergence speed of the model.

[0072] Subsequently, the profiles of the large-scale ground-penetrating radar B-scan image data were cropped to a fixed size (e.g., or Image patches.

[0073] Iterate through all images in the input dataset to generate the training dataset.

[0074] like Figure 5 As shown, in step 2, a bit-layer extraction model is constructed based on the improved U-Net neural network model. The improved U-Net neural network model includes an encoder, a bottleneck layer, and a decoder. In each layer of the encoder and decoder, the traditional double convolutional layer is replaced by a Res-Dilated Block. Attention gates are introduced at the skip connections of the decoder, specifically:

[0075] First, the residual dilated convolution module is introduced. The residual dilated convolution module includes a dilated convolutional layer, a normalization layer, an activation layer, a regular convolutional layer, a normalization layer, and a residual connection path connected in sequence. The dilated convolutional layer is used to expand the receptive field without reducing the feature map resolution, so as to capture the contextual information of the fracture interface. The residual connection path is used to add the input with the convolutional features to alleviate the gradient vanishing problem in deep network training.

[0076] The encoder, bottleneck layer, and decoder will be introduced next, including:

[0077] 1. Encoder

[0078] The encoder (downsampling path) consists of four encoding stages:

[0079] (1) First stage: Input a single-channel image, pass through a residual dilated convolution module, and output a 64-channel feature map;

[0080] (2) Second to fourth stages: Each stage first performs downsampling through a 2×2 max pooling layer to halve the feature map size; then it passes through a residual dilated convolution module, doubling the number of channels to 128, 256, and 512 respectively;

[0081] 2. Bottleneck layer

[0082] The bottleneck layer is located at the bottom of the network and receives the output of the fourth stage. After one max pooling, it passes through a residual dilated convolution module and outputs 1024 channels (if bilinear interpolation upsampling is used, the number of channels here may be adjusted to adapt to the decoder input. In this invention, the output of the bottleneck layer is upsampled and then concatenated with the encoder features).

[0083] 3. Decoder

[0084] The decoder (upsampling path) consists of four decoding stages. Each stage first enlarges the feature map size by a factor of 2 through bilinear upsampling. Then, the upsampled features are connected to the features of the corresponding encoder layer via a skip connection. Before concatenation, the encoder features are weighted by an attention gate module.

[0085] The concatenated features are input into a residual dilated convolution module, with the number of channels halved sequentially (512, 256, 128, 64).

[0086] 4. Output layer

[0087] The output of the final decoding stage passes through a 1×1 convolutional layer, mapping 64 channels to 1 channel, and outputs a layer probability map with the same size as the original image through a Sigmoid activation function.

[0088] This invention introduces a Res-Dilated Block, which replaces the traditional double convolutional layer in each layer of the encoder and decoder. This module contains dilated convolutions with a dilation rate of 2, which expands the receptive field without increasing the number of parameters and captures laterally continuous stratigraphic features. At the same time, it introduces a Residual Connection, which directly adds the input to the output to solve the degradation problem of deep networks.

[0089] This invention also introduces an attention gate, which adds an attention gate at the skip connection. It uses the deep features of the decoder as the gating signal to weight the shallow features from the encoder, automatically suppressing the response of the background region and focusing on the soil interface region.

[0090] In step 3, a hybrid loss function is constructed, which includes at least two combinations of weighted binary cross-entropy loss, Dice loss, structural similarity loss, and smoothness loss, specifically:

[0091] To address the issues of extreme imbalance between positive and negative samples (where background pixels far outnumber target pixels) and structural breaks, a hybrid loss function is designed. A specific implementation of this invention is provided as follows:

[0092] (2)

[0093] In the formula, To calculate the weighted binary cross-entropy loss, a positive sample weight (pos) is set. weight (For example, set to 20.0-50.0), increase the penalty for missed soil layers. Used to measure the overlap between the predicted region and the actual region. The structural similarity loss is used to constrain the texture structure of the prediction results, ensuring that the layers are not distorted. The smoothness loss is calculated using the Laplacian operator to constrain the continuity of the prediction results.

[0094] In step 4, the bit-layer extraction model is trained based on the training dataset and the hybrid loss function, specifically as follows:

[0095] The model is iteratively trained using the constructed dataset and loss function. During the inference phase, the maximum index (Argmax) of the predicted probability map along the depth direction is taken and converted into physical depth. The accuracy is evaluated by calculating the mean absolute error (MAE). Figure 6 As shown in the figure, experiments demonstrate that the MAE of this method on the test set can reach the 4-5 pixel level, which is significantly better than the traditional envelope detection method. Figure 7 The results of consistency analysis between this method and traditional methods are shown. The horizontal axis of the figure represents the average pixel index of the true depth of the soil layer (Ground Truth), and the vertical axis represents the average pixel index of the depth predicted by the algorithm. The gray dashed line (y=x) at the diagonal of the figure represents the ideal fitting line that is completely consistent with the true value of the predicted value.

[0096] from Figure 7 As can be seen from the data points (cyan circular data points), the data points are closely distributed around the ideal fitted line, almost coinciding with the dashed line. This indicates that the U-Net model combined with the attention mechanism used in this invention can capture the soil layer interface with extremely high accuracy, and the predicted depth has a very high consistency with the actual manually labeled depth (R2 value close to 1), showing stable recognition ability in samples at different depths.

[0097] Traditional envelope detection methods (green "X" shaped data points) exhibit a highly dispersed data point distribution, with most data points significantly deviating from the ideal fitted line and concentrated primarily below it. This indicates that traditional methods are severely affected by noise in complex farmland environments, prone to large prediction biases, and exhibit significant underestimation in identifying deeper soil layers (i.e., failing to accurately identify deep interfaces, resulting in predicted depths lower than the actual depths), demonstrating poor robustness. In summary, the embodiments provided by this invention strongly demonstrate that the method of this invention has significantly higher accuracy and stronger consistency compared to traditional signal processing methods in black soil layer identification tasks, meeting the requirements of precision agriculture for soil layer thickness detection accuracy.

[0098] This invention also provides an automatic extraction system for the interface of black soil layer from ground-penetrating radar based on an improved U-Net, applied to the aforementioned automatic extraction method for the interface of black soil layer from ground-penetrating radar based on an improved U-Net, comprising:

[0099] The data acquisition module is used to acquire ground-penetrating radar B-scan image data;

[0100] The preprocessing module is used to preprocess the ground-penetrating radar B-scan image data and construct the training dataset;

[0101] The model building module is used to build a bit-layer extraction model based on the improved U-Net neural network model and to construct a hybrid loss function;

[0102] The model training module is used to train the bit-layer extraction model based on the hybrid loss function and the training dataset;

[0103] The interface extraction module is used to automatically extract the interface of the black soil layer from ground-penetrating radar based on the trained sub-layer extraction model.

[0104] like Figure 8 As shown, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. When the processor executes the program, it implements the above-described method steps. The computer device includes a processor, a memory, a non-volatile storage medium, a communication interface, and an input / output interface; these components are interconnected via a system bus. The non-volatile storage medium stores an operating system and a computer program. When the computer program is executed by the processor, it implements the automatic extraction method for the black soil layer interface based on the improved U-Net provided in the embodiments of this application using ground-penetrating radar. The memory provides a caching environment for program execution. The processor provides computational and control capabilities, particularly for performing numerous matrix operations on deep learning models. The communication interface receives B-scan data acquired by ground-penetrating radar.

[0105] Embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0109] Contents not described in detail in this specification are prior art known to those skilled in the art. It is hereby indicated that the above description is intended to help those skilled in the art understand this invention, but does not limit the scope of protection of this invention. Any equivalent substitutions, modifications, improvements, or simplifications of the above descriptions that do not depart from the essential content of this invention fall within the scope of protection of this invention.

Claims

1. An automatic extraction method for black soil layer interface based on ground-penetrating radar using an improved U-Net, characterized in that, include: Step 1: Acquire ground-penetrating radar B-scan image data, preprocess it, and construct a training dataset; Step 2: Construct a bit-layer extraction model based on the improved U-Net neural network model. The improved U-Net neural network model includes an encoder, a bottleneck layer, and a decoder. In each layer of the encoder and decoder, the traditional double convolutional layer is replaced by a residual dilated convolutional module, and attention gating is introduced at the skip connections of the decoder. The residual dilated convolution module includes a dilated convolution layer, a normalization layer, an activation layer, a normal convolution layer, a normalization layer, and a residual connection path connected in sequence. The dilated convolution layer is used to expand the receptive field without reducing the feature map resolution in order to capture the contextual information of the fracture interface. The residual connection path is used to add the input to the convolutional features to alleviate the gradient vanishing problem in deep network training. Step 3: Construct a hybrid loss function, which includes at least two combinations of weighted binary cross-entropy loss, Dice loss, structural similarity loss, and smoothness loss; the hybrid loss function is: （1） In the formula, For weighted binary cross-entropy loss, Used to measure the overlap between the predicted region and the actual region. For structural similarity loss, The smoothness loss is calculated using the Laplacian operator to constrain the continuity of the prediction results; Step 4: Train the bit-layer extraction model based on the training dataset and the hybrid loss function; Step 5: Input the ground-penetrating radar B-scan image data to be detected into the trained stratigraphic extraction model to achieve automatic extraction of the black soil layer interface by ground-penetrating radar.

2. The automatic extraction method for black soil layer interface based on improved U-Net using ground-penetrating radar according to claim 1, characterized in that, In step 1, ground-penetrating radar B-scan image data is acquired and preprocessed to construct a training dataset, specifically as follows: Acquire ground-penetrating radar B-scan image data; The ground-penetrating radar B-scan image data is processed and augmented to obtain the input dataset; Based on the layer annotations of the input dataset, Gaussian soft labels are generated using a Gaussian distribution function to enhance the labels and construct the training dataset.

3. The automatic extraction method for black soil layer interface based on improved U-Net using ground-penetrating radar according to claim 2, characterized in that, The ground-penetrating radar B-scan image data is processed as follows: The ground-penetrating radar B-scan image data were sequentially subjected to zero-time correction, direct wave removal, time-varying gain processing, and amplitude normalization processing.

4. The automatic extraction method for black soil layer interface based on improved U-Net using ground-penetrating radar according to claim 3, characterized in that, Data augmentation is performed on the ground-penetrating radar B-scan image data, specifically as follows: Data augmentation was performed on the processed ground-penetrating radar B-scan image data, including random left and right flipping, random signal strength scaling, and the addition of random Gaussian noise.

5. The automatic extraction method for black soil layer interface based on improved U-Net using ground-penetrating radar according to claim 4, characterized in that, Based on the layer annotations of the input dataset, Gaussian soft labels are generated using a Gaussian distribution function for label augmentation, and a training dataset is constructed as follows: For each column of A-scan data in each image of the input dataset, obtain the manually annotated layer depth location; Using the layer depth as the center, a probability distribution band is generated using a one-dimensional Gaussian function. The probability distribution band is then used as the label value of the A-scan data in that column to generate Gaussian soft labels. Iterate through all images in the input dataset to generate the training dataset.

6. An automatic interface extraction system for black soil layer based on improved U-Net ground penetrating radar, applied to the automatic interface extraction method for black soil layer based on improved U-Net as described in any one of claims 1-5, characterized in that, include: The data acquisition module is used to acquire ground-penetrating radar B-scan image data; The preprocessing module is used to preprocess the ground-penetrating radar B-scan image data and construct the training dataset; The model building module is used to build a bit-layer extraction model based on the improved U-Net neural network model and to construct a hybrid loss function; The model training module is used to train the bit-layer extraction model based on the hybrid loss function and the training dataset; The interface extraction module is used to automatically extract the interface of the black soil layer from ground-penetrating radar based on the trained sub-layer extraction model.

7. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the automatic extraction method for black soil layer interface based on improved U-Net as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the automatic extraction method for black soil layer interface based on improved U-Net as described in any one of claims 1 to 5.

Images