A Karst Anomaly Identification Method Based on Improved Faster R-CNN

Abstract

This invention provides a karst anomaly identification method based on an improved Faster R-CNN, comprising the following steps: Step 1, establishing a karst anomaly image dataset based on shear wave velocity geophysical images; Step 2, constructing an improved Faster R-CNN network; Step 3, training the improved Faster R-CNN network; Step 4, visualizing the test results; Step 5, identifying karst anomaly regions; Step 6, visualizing the karst anomaly region results. This invention constructs an improved Faster R-CNN network to accurately and efficiently identify karst anomaly regions in shear wave velocity geophysical images.

Description

Technical Field

[0001] This invention relates to a karst anomaly identification method based on an improved Faster R-CNN, belonging to the field of geological engineering exploration technology. Background Technology

[0002] In the field of geological engineering exploration, the identification of karst anomaly areas is of great significance for assessing the stability of underground geological structures, preventing geological disasters, and developing mineral resources. Currently, the identification of karst anomaly areas from shear wave velocity maps mainly relies on manual interpretation. However, this method suffers from problems such as strong dependence on expert experience, low interpretation efficiency, and a high risk of misjudgment or omission. In recent years, target detection algorithms have made significant progress in the field of image processing and are gradually being applied to geological exploration. For example, the Faster R-CNN network has demonstrated excellent accuracy and speed in ground-penetrating radar pipeline identification. The Faster R-CNN network is a classic deep learning-based target detection algorithm. This model achieves high-precision target recognition through two-stage detection and has strong versatility, but it still has limitations in small target detection capability and insufficient positioning accuracy. Summary of the Invention

[0003] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a karst anomaly identification method based on an improved Faster R-CNN. This method constructs an improved Faster R-CNN network, identifies karst anomaly regions in shear wave velocity geophysical images through the improved Faster R-CNN network, and visualizes the karst anomaly region identification results based on the improved Faster R-CNN network. This achieves automatic identification, annotation, and visualization of karst anomaly regions based on shear wave velocity geophysical images.

[0004] The objective of this invention is achieved through the following technical solution: a karst anomaly identification method based on an improved Faster R-CNN, comprising the following steps: Step 1, establishing a karst anomaly image dataset based on shear wave velocity geophysical images; Step 2, constructing an improved Faster R-CNN network: replacing the feature extraction backbone network of the Faster R-CNN network with a ResNeXt network, setting a CBAM attention mechanism module between the second convolutional layer and the batch normalization layer of each incomplete block of the ResNeXt network, retaining the Region Proposal Network (RPN) and classification and regression layers of the Faster R-CNN network, adding a Feature Pyramid Network (FPN) between the ResNeXt network and the RPN, and replacing the Region of Interest Pooling (ROIPooling) of the Faster R-CNN network with Region of Interest Alignment (ROI Align) to establish the improved Faster R-CNN network; Step 3, training the improved Faster R-CNN network: training the improved Faster R-CNN network based on the karst anomaly image dataset and outputting test results, wherein the test results are a test set based on the karst anomaly image dataset, after being improved by Faster R-CNN. Step 4: Visualize the test results: Input the test results and the corresponding shear wave velocity geophysical images from the shear wave velocity geophysical image set into OpenCV, and draw bounding boxes and labels on the corresponding shear wave velocity geophysical images from the shear wave velocity geophysical image set to visualize the test results; Step 5: Identify karst anomaly areas: Input the shear wave velocity geophysical images of the karst anomaly areas to be identified into the trained improved Faster R-CNN network to obtain the classification and regression results output by the improved Faster R-CNN network, i.e., the karst anomaly area results; Step 6: Visualize the karst anomaly area results: Input the karst anomaly area results identified in Step 5 and the corresponding shear wave velocity geophysical images of the karst anomaly areas to be identified into OpenCV, and draw bounding boxes and labels on the corresponding shear wave velocity geophysical images of the karst anomaly areas to be identified to obtain the karst anomaly area result map.

[0005] In one specific implementation, multiple shear wave velocity geophysical images are obtained using the multi-source frequency domain seismic wave method. Each shear wave velocity geophysical image is preprocessed and then amplified to form a shear wave velocity geophysical image set. The shear wave velocity geophysical image set is then labeled, and the labeled shear wave velocity geophysical image set is divided into a training set, a validation set, and a test set in a ratio of 8:1:1 to form a karst anomaly image dataset.

[0006] In one specific implementation, the sample labeling of the shear wave velocity geophysical image set is as follows: karst anomalies are divided into two categories: the first category is loose and non-dense soil layers, and the second category is karst or bedrock fractured or weak interlayers. The loose and non-dense soil layers in each shear wave velocity geophysical image set are represented by the label S, and the karst or bedrock fractured or weak interlayers in each shear wave velocity geophysical image set are represented by the label Y.

[0007] In one specific implementation, in step 3, the AdamW optimizer is used to adjust the weights to be trained in the improved Faster R-CNN network, and the weight decay parameter is set to 1×10. -4 The learning rate is adjusted using a cosine annealing learning rate scheduler.

[0008] In one specific implementation, step 3 specifically comprises: step 3.1, initializing the weights to be trained in the improved Faster R-CNN network, and setting the learning rate and batch size; step 3.2, inputting the karst anomaly image dataset into the improved Faster R-CNN network for training, wherein the training set is used to train the improved Faster R-CNN network to adjust the weights of the improved Faster R-CNN network, the validation set is used to adjust the hyperparameters of the improved Faster R-CNN network and to preliminarily evaluate the capabilities of the improved Faster R-CNN network, and the test set is used to finally evaluate the capabilities of the improved Faster R-CNN network.

[0009] In one specific implementation, precision P, recall R, and F1 score are used to ultimately evaluate the improved Faster R-CNN network capabilities. The formula for calculating precision P is: P=TP / (TP+FP), the formula for calculating recall R is: R=TP / (TP+FN), and the formula for calculating F1 score F1-score is: F1-score=2P*R / (P+R), where TP represents the number of correctly identified targets, FP represents the number of incorrectly identified targets, and FN represents the number of unidentified targets.

[0010] Compared with existing technologies, the beneficial effects of this invention are as follows: The improved Faster R-CNN network constructed in this invention extracts features from the input shear wave velocity geophysical image using a ResNeXt network and incorporates the CBAM attention mechanism. It acquires features that fuse low-level detail features and high-level semantic features through a Feature Pyramid Network (FPN). The localization accuracy of the model is enhanced by ROIAlign calibration using the Region of Interest (ROIAlign). Testing of the improved Faster R-CNN network shows that when identifying loose and non-dense soil layers, its precision is 91.4%, 1.7 percentage points higher than the original Faster R-CNN network; its recall is 89.7%, 1.4 percentage points higher than the original Faster R-CNN network; and its F1 score is 90.5%, 1.5 percentage points higher than the original Faster R-CNN network. When identifying karst, fractured bedrock, or weak interlayers, its precision is 93.2%, 1.7 percentage points higher than the original Faster R-CNN network; and its recall is 91.8%, higher than the original Faster R-CNN network. The R-CNN network outperforms the Faster R-CNN network by 1.7 percentage points, achieving an F1 score of 92.5%, which is 1.7 percentage points higher than the Faster R-CNN network. The improved Faster R-CNN network demonstrates superior performance in identifying karst anomaly regions in shear wave velocity geophysical images, while also achieving a 16-fold increase in identification efficiency compared to traditional manual interpretation. Therefore, this invention can accurately and efficiently identify karst anomaly regions in shear wave velocity geophysical images, providing effective technical support for the field of geological exploration. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating the karst anomaly identification method based on the improved Faster R-CNN of this invention.

[0012] Figure 2 The graph shows the loss values ​​during the training process of each model.

[0013] Figure 3 The graph shows the loss values ​​during the validation process of each model.

[0014] Figure 4 This is a geophysical image of shear wave velocity after sample annotation.

[0015] Figure 5 for Figure 4 A visualization of the test results based on the improved Faster R-CNN network.

[0016] Figure 6 for Figure 4 A visualization of the test results based on the Faster R-CNN network.

[0017] Figure 7 for Figure 4 A visualization of the test results based on RetinaNet.

[0018] Figure 8 for Figure 4 A visualization of the test results based on YOLOv8. Detailed Implementation

[0019] The present invention will now be described in detail with reference to the embodiments and accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0020] refer to Figure 1 This invention discloses a karst anomaly identification method based on an improved Faster R-CNN, including the following steps: Step 1, establishing a karst anomaly image dataset based on shear wave velocity geophysical images.

[0021] Specifically, multiple shear wave velocity geophysical images were obtained using the multi-source frequency domain seismic wave method. Each shear wave velocity geophysical image was preprocessed and then amplified to form a shear wave velocity geophysical image set. The shear wave velocity geophysical image set was then labeled, and the labeled shear wave velocity geophysical image set was divided into a training set, a validation set, and a test set in a ratio of 8:1:1 to form a karst anomaly image dataset.

[0022] To identify karst anomaly areas along railway sections, multi-source frequency domain seismic wave methods were used to obtain geophysical images of shear wave velocity due to interference from the track.

[0023] In this embodiment, multi-source frequency domain seismic wave method was used to conduct karst exploration on the K2250+220-K2255+100 section of the Beijing-Guangzhou Railway to obtain shear wave velocity geophysical images, with a total survey line of 22 km. Karst anomaly areas in the rock strata characterized by relatively low shear wave velocity and closed loops or depressions in velocity contour lines are loose and non-dense soil layers; karst anomaly areas in the rock strata characterized by relatively low shear wave velocity and obvious depressions in velocity contour lines are karst or fractured bedrock or weak interlayers. 600 shear wave velocity geophysical images were obtained by sliding a square window.

[0024] Spatial and pixel-level variations were used to expand the shear wave velocity geophysical images from 600 to 1600, making the training data more diverse and beneficial to model training. Spatial variations included horizontal mirroring and local magnification, while pixel-level variations included increasing brightness, increasing contrast, decreasing brightness, and decreasing contrast.

[0025] Based on the opinions of multiple experts, LabelImg was used to mark karst anomaly areas in 1600 shear wave velocity geophysical images. (See reference...) Figure 4 .

[0026] Specifically, the sample labeling of the shear wave velocity geophysical image set is as follows: karst anomalies are divided into two categories: the first category is loose and non-dense soil layers, and the second category is karst or bedrock fractured or weak interlayers. Loose and non-dense soil layers in each shear wave velocity geophysical image set are represented by the label S, and karst or bedrock fractured or weak interlayers in each shear wave velocity geophysical image set are represented by the label Y.

[0027] The karst anomaly image dataset is saved in VOC format, and the shear wave velocity geophysical image dataset is saved in JPG format.

[0028] The 1600 annotated shear wave velocity geophysical images were divided into training, validation, and test sets in a ratio of 8:1:1 to form a karst anomaly image dataset.

[0029] Step 2: Construct an improved Faster R-CNN network: Replace the feature extraction backbone of the Faster R-CNN network with a ResNeXt network. Set up a CBAM attention mechanism module between the second convolutional layer and the batch normalization layer of each incomplete block of the ResNeXt network. Retain the Region Proposal Network (RPN) and classification and regression layers of the Faster R-CNN network. Add a Feature Pyramid Network (FPN) between the ResNeXt network and the RPN. Replace the Region of Interest (ROI) Pooling of the Faster R-CNN network with Region of Interest Alignment (ROI) to build an improved Faster R-CNN network.

[0030] The improved Faster R-CNN network includes a ResNeXt network, a Feature Pyramid Network (FPN), a Region Proposal Network (RPN), a Region of Interest (ROI) Alignment network, and classification and regression layers. The output of the ResNeXt network is connected to the Feature Pyramid Network (FPN), the output of the FPN is connected to the RPN, the output of the RPN is connected to the ROI Alignment network, and the output of the ROI Alignment network is connected to the classification and regression layers.

[0031] The second convolutional layer of each incomplete block in the ResNeXt network uses grouped convolutions to perform preliminary feature extraction and transformation, resulting in feature maps containing relatively rich information. Inserting a CBAM attention mechanism module between the second convolutional layer and the batch normalization layer can better capture key channel and spatial information in the feature map, effectively improving the model's attention to the target region and enhancing feature extraction accuracy. Inserting the CBAM attention mechanism module after the first convolutional layer would increase its computational cost due to the large number of channels in the feature map; inserting it after the third convolutional layer might miss some feature patterns formed during the grouped convolution process. Specifically, the operation is as follows: First, assuming the output feature map of the second convolutional layer is f, f is input into the CBAM attention mechanism module. Channel attention weights are obtained through CAM, and spatial attention weights are obtained through SAM. Then, feature map f is multiplied by the channel attention weights, and the resulting feature map is multiplied by the spatial attention weights to obtain a channel- and spatial attention-weighted feature map F. Finally, batch normalization is performed on F.

[0032] A Feature Pyramid Network (FPN) is added between the ResNeXt network and the Region Proposal Network (RPN). The FPN fuses low-level detail features and high-level semantic features to enhance the model's ability to detect small targets. Specifically, the multi-scale feature maps C2, C3, C4, and C5 extracted by the feature extraction network are input into the FPN. Multi-scale feature maps P2, P3, P4, and P5 are generated through lateral connections (1×1 convolutions to adjust the number of channels) and a top-down path (upsampling + feature map addition). P2 is used to detect small targets, and P5 is used to detect large targets. Subsequently, the multi-scale feature maps P2, P3, P4, and P5 output by the FPN are input into the RPN to generate candidate regions. Non-maximum suppression is then used to filter out the final candidate regions.

[0033] Region of Interest (ROI) Align uses bilinear interpolation to avoid quantization errors, thereby enhancing the model's localization accuracy.

[0034] Step 3: Train the improved Faster R-CNN network: Train the improved Faster R-CNN network based on the karst anomaly image dataset and output the test results. The test results are the classification and regression results output by the improved Faster R-CNN network based on the test set of the karst anomaly image dataset.

[0035] Specifically, in step 3, the AdamW optimizer is used to adjust the weights to be trained in the improved Faster R-CNN network, and the weight decay parameter is set to 1×10. -4This method can improve the convergence speed and generalization ability of the improved Faster R-CNN network. It uses a cosine annealing learning rate scheduler to adjust the learning rate, which helps the improved Faster R-CNN network to learn quickly in the early stage of training and make fine adjustments in the later stage of training, thereby improving the convergence speed and final performance of the improved Faster R-CNN network.

[0036] The formula for adjusting the learning rate using the cosine annealing learning rate scheduler is as follows: In the formula: η t η is the learning rate in the t-th cycle during training. min It is the minimum value of the learning rate, η. max It is the maximum learning rate, T cur It is the index of the current training cycle, T max This indicates the length of a complete cosine annealing cycle.

[0037] Specifically, step 3 is as follows: Step 3.1: Initialize the weights to be trained in the improved Faster R-CNN network, and set the learning rate and batch size.

[0038] Step 3.2: Input the karst anomaly image dataset into the improved Faster R-CNN network for training. The training set is used to train the improved Faster R-CNN network to adjust its weights. The validation set is used to validate the improved Faster R-CNN network to adjust its hyperparameters and preliminarily evaluate its capabilities. The test set is used to test the improved Faster R-CNN network to ultimately evaluate its capabilities, namely its generalization ability, reliability, and practicality.

[0039] In the training of the improved Faster R-CNN network, the loss value during each training process and the loss value during each validation process are recorded in real time. When the loss values ​​during training and validation converge, the weights of the improved Faster R-CNN network are saved and input into the test set to test the improved Faster R-CNN network to finally evaluate the capabilities of the improved Faster R-CNN network.

[0040] Specifically, the precision P, recall R, and F1 score are used to ultimately evaluate the improved capabilities of the Faster R-CNN network. The formula for calculating precision P is: P=TP / (TP+FP), the formula for calculating recall R is: R=TP / (TP+FN), and the formula for calculating F1 score F1 is: F1-score=2P*R / (P+R).

[0041] Where TP represents the number of correctly identified targets, FP represents the number of incorrectly identified targets, and FN represents the number of unidentified targets.

[0042] Precision represents the proportion of samples predicted as positive that are actually positive, recall represents the proportion of samples that are actually positive that are correctly predicted, and the F1 score is the harmonic mean of precision and recall, balancing the performance of the two.

[0043] Step 4: Visualize the test results: Input the test results and the corresponding shear wave velocity geophysical images in the shear wave velocity geophysical image set into OpenCV, and draw annotation boxes and labels on the corresponding shear wave velocity geophysical images in the shear wave velocity geophysical image set to visualize the test results.

[0044] Since the classification and regression results output by the improved Faster R-CNN network are numerical data, including bounding box coordinates and labels, visualization is usually required in practical applications for easy viewing and statistical analysis. OpenCV (Open Source Computer Vision Library) is a cross-platform computer vision library that can implement many common algorithms in image processing and computer vision. Therefore, OpenCV is used to visualize the classification and regression results output by the improved Faster R-CNN network.

[0045] In this embodiment, a karst anomaly image dataset including 1600 shear wave velocity geophysical images is input into the improved Faster R-CNN network, Faster R-CNN network, RetinaNet, and YOLOv8, respectively, to train the improved Faster R-CNN network, Faster R-CNN network, RetinaNet, and YOLOv8 and output test results.

[0046] The initial batch size and learning rate settings for each model were determined using a grid search method. The batch size and learning rate for the improved Faster R-CNN network were 16 and 0.001, respectively; for the Faster R-CNN network, they were 8 and 0.001, respectively; for RetinaNet, they were 16 and 0.001, respectively; and for YOLOv8, they were 16 and 0.01, respectively.

[0047] During the training of each model, the loss value during each training process and the loss value during each validation process are recorded in real time. When the loss values ​​during training and validation converge, the weights of the corresponding models are saved, and the test set is input into each model to finally evaluate the capabilities of each model.

[0048] See Figure 2 and Figure 3 After 100 iterations, the loss values ​​of each model converged during both the training and validation processes. The loss values ​​of the improved Faster R-CNN network, RetinaNet, and YOLOv8 during the training process were 0.063, 0.081, 0.089, and 0.114, respectively, while the loss values ​​during the validation process were 0.056, 0.072, 0.083, and 0.113, respectively. The difference between the loss values ​​during training and validation was very small, indicating that none of the models experienced overfitting. The improved Faster R-CNN had lower loss values ​​in both processes than the other three models, indicating that the improved Faster R-CNN had the highest accuracy in identifying karst anomaly areas in shear wave velocity geophysical maps on both the training and validation sets.

[0049] When identifying loose and non-dense soil layers, the improved Faster R-CNN network achieved a precision of 91.4%, which is 1.7, 4.7, and 7.7 percentage points higher than Faster R-CNN, RetinaNet, and YOLOv8, respectively; a recall of 89.7%, which is 1.4, 4.5, and 7.6 percentage points higher than Faster R-CNN, RetinaNet, and YOLOv8, respectively; and an F1 score of 90.5%, which is 1.5, 4.6, and 7.6 percentage points higher than Faster R-CNN, RetinaNet, and YOLOv8, respectively. When identifying karst, fractured bedrock, or weak interlayers, the improved Faster R-CNN network achieved a precision of 93.2%, which is 1.7, 3.9, and 6.5 percentage points higher than Faster R-CNN, RetinaNet, and YOLOv8, respectively; and a recall of 91.8%, which is higher than Faster R-CNN, RetinaNet, and YOLOv8, respectively. The improved Faster R-CNN network outperformed RetinaNet and YOLOv8 by 1.7, 3.7, and 6.5 percentage points, respectively; its F1 score was 92.5%, which was 1.7, 3.8, and 6.5 percentage points higher than Faster R-CNN, RetinaNet, and YOLOv8, respectively. This analysis indicates that the improved Faster R-CNN network performs best in identifying karst anomaly regions on the test set.

[0050] The test results obtained from each model, along with the corresponding shear wave velocity geophysical images from a set of 1600 shear wave velocity geophysical images, are all input into OpenCV. Labels and bounding boxes are then drawn onto the corresponding shear wave velocity geophysical images in the image set, resulting in a visualization of the test results based on each model. (See also...) Figures 4 to 8 ,from Figure 4 As can be seen from the markings, there is one area with loose and non-dense soil, and two areas with karst or bedrock fractures or weak interlayers. Figure 5 In the middle section, there is one area with loose and non-dense soil, and two areas with karst or broken bedrock or weak interlayers; the identification is completely correct. Figure 6 In the middle section, there are two areas with loose and non-dense soil layers, and two areas with karst or fractured or weak interlayered bedrock. Among them, one of the areas with loose and non-dense soil layers is a mislabeled location. Figure 7 In the middle section, there is one area with loose and non-dense soil layer, and two areas with karst or broken or weak interlayered bedrock. The identification of the loose and non-dense soil layer is incorrect. Figure 8 In the middle, there are two areas with karst or broken or weak interlayered bedrock. In one of them, the identification area is small and the soil layer is loose and not dense, so the solid objects were not identified.

[0051] Compared to other models, the improved Faster R-CNN network shows significantly fewer missed and false positives in karst anomaly areas, and a higher degree of agreement with expert annotations. The improved Faster R-CNN network achieves 16 times the efficiency of traditional manual identification.

[0052] Step 5: Identify karst anomalies: Input the shear wave velocity geophysical image of the karst anomaly to be identified into the trained improved Faster R-CNN network to obtain the classification and regression results output by the improved Faster R-CNN network, i.e., the karst anomaly area results.

[0053] The shear wave velocity geophysical image is input into the ResNeXt network to extract features introduced by the CBAM attention mechanism. The features extracted by the ResNeXt network are input into the Feature Pyramid Network (FPN) to output features that fuse low-level detail features and high-level semantic features. The features obtained by the Feature Pyramid Network (FPN) are input into the Region Proposal Network (RPN) to generate candidate regions. The candidate regions generated by the RPN are input into the Region of Interest (ROI) Alignment. The features obtained by the ROI Alignment are input into the classification and regression layers to determine whether the candidate regions belong to the target category and to fine-tune the position of the candidate regions to obtain the final target detection result, namely the karst anomaly region result.

[0054] Step 6: Visualize the results of karst anomaly areas: Input the results of the karst anomaly areas identified in Step 5 and the corresponding shear wave velocity geophysical images of the karst anomaly areas to be identified into OpenCV, draw annotation boxes and labels on the corresponding shear wave velocity geophysical images of the karst anomaly areas to be identified, and obtain the karst anomaly area result map.

[0055] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions and substitutions can be made without departing from the inventive concept, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for identifying karst anomalies based on an improved Faster R-CNN, characterized in that, Includes the following steps: Step 1: Establish a dataset of karst anomaly images based on shear wave velocity geophysical images; Step 2: Construct an improved Faster R-CNN network: Replace the feature extraction backbone network of the Faster R-CNN network with a ResNeXt network. Set a CBAM attention mechanism module between the second convolutional layer and the batch normalization layer of each incomplete block of the ResNeXt network. Retain the Region Proposal Network (RPN) and classification and regression layers of the Faster R-CNN network. Add a Feature Pyramid Network (FPN) between the ResNeXt network and the RPN. Replace the Region of Interest (ROI) Pooling of the Faster R-CNN network with Region of Interest Alignment (ROI) to construct an improved Faster R-CNN network. Step 3: Train the improved Faster R-CNN network: Train the improved Faster R-CNN network based on the karst anomaly image dataset and output the test results. The test results are the classification and regression results output by the improved Faster R-CNN network on the test set of the karst anomaly image dataset. Step 4: Visualize the test results: Input the test results and the corresponding shear wave velocity geophysical images in the shear wave velocity geophysical image set into OpenCV, and draw annotation boxes and labels on the corresponding shear wave velocity geophysical images in the shear wave velocity geophysical image set to visualize the test results; Step 5: Identify karst anomaly areas: Input the shear wave velocity geophysical image of the karst anomaly area to be identified into the trained improved Faster R-CNN network to obtain the classification and regression results output by the improved Faster R-CNN network, i.e., the karst anomaly area results. Step 6: Visualize the results of karst anomaly areas: Input the results of the karst anomaly areas identified in Step 5 and the corresponding shear wave velocity geophysical images of the karst anomaly areas to be identified into OpenCV, draw annotation boxes and labels on the corresponding shear wave velocity geophysical images of the karst anomaly areas to be identified, and obtain the karst anomaly area result map.

2. The karst anomaly identification method based on improved Faster R-CNN as described in claim 1, characterized in that, Multiple shear wave velocity geophysical images were obtained using the multi-source frequency domain seismic wave method. Each shear wave velocity geophysical image was preprocessed and then amplified to form a shear wave velocity geophysical image set. The shear wave velocity geophysical image set was labeled, and the labeled shear wave velocity geophysical image set was divided into a training set, a validation set, and a test set in a ratio of 8:1:1 to form a karst anomaly image dataset.

3. The karst anomaly identification method based on improved Faster R-CNN as described in claim 2, characterized in that, The sample labeling of the shear wave velocity geophysical image set is as follows: karst anomalies are divided into two categories: the first category is loose and non-dense soil layers, and the second category is karst or bedrock fractured or weak interlayers. Loose and non-dense soil layers in each shear wave velocity geophysical image set are represented by the label S, and karst or bedrock fractured or weak interlayers in each shear wave velocity geophysical image set are represented by the label Y.

4. The karst anomaly identification method based on improved Faster R-CNN as described in claim 3, characterized in that, In step 3, the AdamW optimizer is used to adjust the weights to be trained in the improved Faster R-CNN network, with the weight decay parameter set to 1×10. -4 The learning rate is adjusted using a cosine annealing learning rate scheduler.

5. The karst anomaly identification method based on improved Faster R-CNN as described in claim 4, characterized in that, Step 3 specifically involves: Step 3.1: Initialize the weights to be trained in the improved Faster R-CNN network, and set the learning rate and batch size; Step 3.2: Input the karst anomaly image dataset into the improved Faster R-CNN network for training. The training set is used to train the improved Faster R-CNN network to adjust its weights. The validation set is used to adjust the hyperparameters of the improved Faster R-CNN network and to preliminarily evaluate its capabilities. The test set is used to finally evaluate the capabilities of the improved Faster R-CNN network.

6. The karst anomaly identification method based on improved Faster R-CNN as described in claim 5, characterized in that, The improved Faster R-CNN network's capabilities are ultimately evaluated using precision (P), recall (R), and F1 score (F1-score). Precision (P) is calculated as: P = TP / (TP + FP), recall (R) is calculated as: R = TP / (TP + FN), and F1 score (F1-score) is calculated as: F1-score = 2P * R / (P + R), where TP represents the number of correctly identified targets, FP represents the number of incorrectly identified targets, and FN represents the number of unidentified targets.

Images