Karst cave area detection and correction system based on deep learning

Abstract

The invention discloses a deep learning-based karst cave area detection and correction system, and relates to the field of karst cave area detection, the deep learning-based karst cave area detection and correction system comprises a data acquisition and preprocessing module, a deep learning model construction module and a data correction and result output module, the data acquisition and preprocessing module comprises multi-source data acquisition and data preprocessing, and is used for processing a collected database; and the deep learning model construction module performs automatic feature extraction and pattern recognition on the multi-source geophysical data. The convolutional neural network, the recurrent neural network and the generative adversarial network form a deep learning model architecture, and model training and optimization are matched, so that the whole deep learning model architecture module can learn complex modes and laws in karst cave detection data; therefore, an accurate judgment basis is provided for subsequent data correction and result output, and the accuracy and reliability of a detection result are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cave area detection, and specifically to a cave area detection correction system based on deep learning. Background Art

[0002] Karst cave area exploration refers to the use of a variety of technical means to conduct comprehensive exploration and analysis of the geological conditions of the karst cave area.

[0003] However, in the existing technology, the complex geological structure of the karst cave area poses a great challenge to geological exploration. Traditional detection methods often have problems of inaccurate data and low resolution when faced with the irregular shape of the karst cave, variable fillings and complex surrounding geological environment. In addition, the geological conditions in the karst cave area are extremely complex, and the original detection data is easily affected by various factors and has deviations. Data correction is very necessary. Among the original data, geological radar data may be affected by electromagnetic interference, and seismic exploration data will have abnormal wave velocity due to uneven strata. Although drilling and coring are accurate, the samples are limited. These all affect the accuracy of the data. Summary of the Invention

[0004] The purpose of the present invention is to provide a deep learning-based cave area detection and correction system to solve the problem raised in the above background technology that the cave area poses great challenges to geological exploration due to its complex geological structure. Traditional detection methods often have problems of inaccurate data and low resolution when faced with the irregular shape of the cave, variable fillings and complex surrounding geological environment.

[0005] To achieve the above objectives, the present invention provides the following technical solutions: a deep learning-based cave area detection and correction system, comprising a data acquisition and preprocessing module, a deep learning model construction module, and a data correction and result output module. The data acquisition and preprocessing module includes multi-source data acquisition and data preprocessing, which are used to process the collected database. The deep learning model construction module automatically extracts features and performs pattern recognition on multi-source geophysical data. The data correction and result output module converts the original model output into cave detection results through multi-dimensional verification, physical constraint fusion, and visual expression.

[0006] The deep learning model construction module includes deep learning model architecture and model training and optimization. The deep learning model architecture includes convolutional neural networks, recurrent neural networks and their variants, and adversarial networks. The convolutional neural network adopts the U-Net architecture to capture the spatial features in geological radar image data, and uses a bidirectional LSTM network to process seismic exploration data. The adversarial network uses a conditional generative adversarial network based on the traditional GAN, and additional input conditional information controls the properties of the generated data. Model training and optimization maps the data to a unified feature space through comparative learning to eliminate the dimensional differences of different geophysical methods.

[0007] Preferably, model training and optimization include training data set preprocessing, loss function training, algorithm optimization and model optimization;

[0008] In addition to collecting a large amount of multi-source exploration data from known cave areas and their corresponding real geological conditions and performing conventional divisions, the training data preprocessing method combines data oversampling and undersampling to process unbalanced data sets.

[0009] The loss function training adopts the focus loss function for classification tasks, adjusts the degree of attention to samples of different difficulty levels, and increases the learning weight of difficult-to-classify samples;

[0010] Algorithm optimization adds weight decay to the Adam algorithm to dynamically adjust the learning rate;

[0011] Model optimization uses the validation set to monitor the model's accuracy, recall rate, mean square error and other conventional indicators, and introduces F1 score and average precision to conduct a more comprehensive evaluation of model performance.

[0012] Preferably, the multi-source data acquisition includes geological radar, seismic survey and borehole coring;

[0013] Geological radar emits high-frequency electromagnetic waves and detects the distribution of underground media by receiving reflected waves, thus obtaining the general outline and location information of the cave.

[0014] Seismic exploration uses the differences in propagation characteristics of artificially excited seismic waves in different geological media to analyze the information of reflected and refracted waves to determine the depth and size of caves;

[0015] Drilling and coring directly obtain underground rock samples, and through core analysis, the composition of the filling material and rock properties in the cave can be accurately determined.

[0016] Preferably, data preprocessing includes denoising, data standardization and data fusion;

[0017] Denoising is done by using wavelet transform, adaptive filtering and other methods to remove electromagnetic interference noise in geological radar data and environmental noise in seismic exploration data.

[0018] Data standardization: Standardize data from different sources and dimensions to a unified scale for subsequent data analysis and model training.

[0019] Data fusion integrates pre-processed geological radar, seismic exploration and drill core data.

[0020] Preferably, the deep learning model architecture includes the following steps:

[0021] A1. Extracting multi-scale features of caves through convolutional neural networks;

[0022] A2. Use U-Net++ or V-Net to perform voxel-level segmentation of caves and output cave probability cubes;

[0023] A3. A bidirectional LSTM network analyzes borehole acoustic logging time series data to determine the vertical development of caves.

[0024] Preferably, the data correction and result output module includes the following steps:

[0025] B1. Input new exploration data: Newly collected cave area exploration data, including geological radar data, seismic exploration data, and drill core data, are formatted and input into the trained deep learning model.

[0026] B2. Model processing and correction: The deep learning model processes the input raw data based on the knowledge learned during the training phase, such as cave characteristic patterns, data distribution patterns, and data variation patterns under different geological conditions;

[0027] B3. Multi-source data fusion correction: After the model corrects the single-source data separately, the multi-source data is fused and corrected.

[0028] B4. Generate a geological structure report. The deep learning model combines its own prediction results to generate a detailed geological structure report of the cave area. The report covers key information such as the location, shape, size, depth, and filling type of the cave.

[0029] B5. Data visualization: Display the cross-sectional shape of the cave at different depths and the surrounding geological conditions in the form of a two-dimensional cross-section diagram, and vividly present the three-dimensional form of the cave and its spatial distribution in the entire detection area through a three-dimensional model;

[0030] B6. Output the generated geological structure report file containing visualization content.

[0031] Preferably, the loss function training includes the following steps:

[0032] C1. Clarify the task type and loss function selection: In cave detection, if you need to determine whether a cave exists or not, choose the focal loss function. If you need to predict continuous values ​​such as the size and depth of a cave, use the Huber loss function.

[0033] C2. Initialize model parameters: After selecting the loss function, initialize the parameters of the deep learning model;

[0034] C3, Forward Propagation: The training data (including multi-source data such as geological radar and seismic exploration) is input into the initialized model. The model processes the input data layer by layer according to its own architecture and parameter settings. After a series of operations, the model outputs the prediction results.

[0035] C4. Calculate the loss value: Compare the model's prediction results with the corresponding true labels, and calculate the difference between the prediction results and the true values ​​according to the selected loss function to obtain the loss value;

[0036] C5. Backpropagation and parameter update: Based on the calculated loss value, the gradient of the loss value with respect to the model parameters is calculated through the backpropagation algorithm. The AdamW optimization algorithm is used to update the model parameters according to the calculated gradient. The updated model parameters will be used for the next round of forward propagation and backpropagation.

[0037] Preferably, the algorithm optimization comprises the following steps:

[0038] D1. Clarify optimization goals;

[0039] D2. Data cleaning and preprocessing enhancement: Perform more detailed cleaning of raw data;

[0040] D3, data expansion and balancing;

[0041] D4. Loss function optimization: Try different loss function combinations or improve existing loss functions. In addition to the focal loss function, combine the triplet loss function;

[0042] D5. Hyperparameter tuning: Determine the hyperparameters that need to be tuned and set a reasonable value range for each hyperparameter.

[0043] Compared with the prior art, the present invention has the following beneficial effects:

[0044] 1. In the present invention, convolutional neural networks, recurrent neural networks and generative adversarial networks constitute a deep learning model architecture. Since cave detection data has spatial characteristics, its convolutional layer can automatically extract local features, and the pooling layer can reduce the feature dimension, which is effective in processing geological radar image data and identifying cave boundaries and internal structural characteristics. Recurrent neural networks and their variants LSTM can effectively process such sequence data and capture long-term dependencies because seismic exploration data has correlation in time or space sequences. They are used to analyze time series data of seismic wave propagation to predict cave characteristics. Generative adversarial networks can generate simulated cave data through the generator, and the discriminator judges the similarity between the generated data and the real data, thereby expanding the training data set and improving the model generalization ability. Combined with model training and optimization, the entire deep learning model construction module can learn the complex patterns and laws in the cave detection data, thereby providing accurate judgment basis for subsequent data correction and result output, greatly improving the accuracy and reliability of the detection results.

[0045] 2. In the present invention, model training and optimization include training data set preprocessing, loss function training, algorithm optimization and model optimization. In addition to collecting a large amount of multi-source detection data of known cave areas and their corresponding real geological conditions and performing conventional divisions for training data preprocessing, a method combining data oversampling and undersampling is used to process unbalanced data sets. Loss function training adopts a focus loss function for classification tasks, adjusts the degree of attention to samples of different difficulty levels, and increases the learning weight of difficult-to-classify samples. Algorithm optimization adds weight attenuation on the basis of the Adam algorithm to dynamically adjust the learning rate. Model optimization uses a validation set to monitor the model's accuracy, recall rate, mean square error and other conventional indicators, and introduces F1 score and average precision mean to conduct a more comprehensive evaluation of model performance.

[0046] 3. In the present invention, the newly collected karst cave area detection data, including geological radar data, seismic exploration data and drilling core sampling related data, are unified and formatted and input into the trained deep learning model. The model uses the learned relationship between strata and wave velocity to correct abnormal wave velocity values. In this process, the model will analyze and judge each data point, and adjust and correct the data according to its internal parameters and algorithm logic. After the data is corrected, the deep learning model combines its own prediction results to generate a detailed geological structure report of the karst cave area. The report content covers key information such as the location, shape, size, depth, and filling type of the karst cave. The model calculates and obtains these specific geological information through a comprehensive analysis of the corrected data and the mapping relationship between the features and geological structure parameters established during the training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1The present invention proposes a block diagram of a karst cave area detection and correction system based on deep learning. DETAILED DESCRIPTION

[0048] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the implementation regulations described are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0049] Example 1: A deep learning-based cave area detection and correction system includes a data acquisition and preprocessing module, a deep learning model construction module, and a data correction and result output module. The data acquisition and preprocessing includes multi-source data acquisition and data preprocessing, which are used to process the collected database. The deep learning model construction module automatically extracts features and recognizes patterns from multi-source geophysical data. The data correction and result output module converts the original model output into cave detection results through multi-dimensional verification, physical constraint fusion, and visual expression.

[0050] The deep learning model construction module includes deep learning model architecture and model training and optimization. The deep learning model architecture includes convolutional neural networks, recurrent neural networks and their variants, and adversarial networks. The convolutional neural network adopts the U-Net architecture to capture the spatial features in geological radar image data, and uses a bidirectional LSTM network to process seismic exploration data. The adversarial network uses a conditional generative adversarial network based on the traditional GAN, and additional input conditional information controls the properties of the generated data. Model training and optimization uses contrastive learning to map data to a unified feature space, eliminating the dimensional differences between different geophysical exploration methods.

[0051] The deep learning model architecture consists of the following steps:

[0052] A1. Extracting multi-scale features of caves through convolutional neural networks;

[0053] A2. Use U-Net++ or V-Net to perform voxel-level segmentation of caves and output cave probability cubes;

[0054] A3. A bidirectional LSTM network analyzes borehole acoustic logging time series data to determine the vertical development of caves.

[0055] In this embodiment, the convolutional neural network (CNN) is effective in processing geological radar image data and identifying cave boundaries and internal structural features, as cave detection data has spatial characteristics. Its convolution layer can automatically extract local features, and the pooling layer can reduce feature dimensions. The RNN and its variant LSTM are effective in processing seismic exploration data and identifying cave boundaries and internal structural features, as seismic exploration data has correlations in time or space sequences. They can effectively process such sequence data and capture long-term dependencies, and are used to analyze time series data of seismic wave propagation to predict cave characteristics. The generative adversarial network (GAN) can generate simulated cave data through a generator, and the discriminator can determine the similarity between the generated data and the real data, thereby expanding the training data set and improving the generalization ability of the model.

[0056] Prepare a large amount of multi-source detection data of known cave areas and their corresponding real geological conditions and divide them into training sets, validation sets and test sets to provide rich materials for model learning. Select appropriate loss functions according to different tasks and data types, such as cross entropy loss function for classification tasks and mean square error loss function for regression tasks. Use optimization algorithms such as stochastic gradient descent to adjust model parameters to minimize the loss function. Dynamically adjust the learning rate during training, use the validation set to monitor model performance to prevent overfitting, and finally use the test set to verify the model generalization ability. Through this series of operations, the deep learning model construction module can learn the complex patterns and laws in the cave detection data, thereby providing accurate judgment basis for subsequent data correction and result output, greatly improving the accuracy and reliability of the detection results.

[0057] Example 2: Model training and optimization include training data set preprocessing, loss function training, algorithm optimization and model optimization. In addition to collecting a large amount of multi-source detection data of known cave areas and their corresponding real geological conditions and performing conventional divisions, the training data preprocessing adopts a method combining data oversampling and undersampling to process unbalanced data sets. Loss function training adopts a focus loss function for classification tasks, adjusts the degree of attention to samples of different difficulty levels, and increases the learning weight of difficult-to-classify samples. Algorithm optimization adds weight decay to the Adam algorithm for dynamic adjustment of the learning rate. Model optimization uses a validation set to monitor the model's accuracy, recall rate, mean square error and other conventional indicators. In addition, the F1 score and average precision mean are introduced to conduct a more comprehensive evaluation of model performance.

[0058] C1. Clarify the task type and loss function selection: In the detection of cave areas, if you need to determine whether a cave exists or other classification problems, select the focus loss function. When you need to predict continuous values ​​such as the size and depth of the cave, use the Huber loss function. C2. Initialize model parameters: After selecting the loss function, initialize the parameters of the deep learning model. C3. Forward propagation: Input the training data (including multi-source data such as geological radar and seismic exploration) into the initialized model. The model processes the input data layer by layer according to its own architecture and parameter settings. After a series of operations, the model outputs the prediction result. C4. Calculate the loss value: Compare the model's prediction result with the corresponding true label. According to the selected loss function, calculate the difference between the prediction result and the true value to obtain the loss value. C5. Backpropagation and parameter update: Based on the calculated loss value, the gradient of the loss value to the model parameters is calculated through the backpropagation algorithm. The AdamW optimization algorithm is used to update the model parameters according to the calculated gradient. The updated model parameters will be used for the next round of forward propagation and backpropagation processes.

[0059] D1. Clarify the optimization goal. D2. Data cleaning and preprocessing enhancement: Perform more detailed cleaning of the raw data. D3. Data expansion and balancing. D4. Loss function optimization: Try different loss function combinations or improve existing loss functions. In addition to the focal loss function, combine the triplet loss function. D5. Hyperparameter tuning: Determine the hyperparameters that need to be tuned, such as the learning rate, batch size, number of iterations, number of model layers, number of neurons, etc., and set a reasonable value range for each hyperparameter.

[0060] In this embodiment, after completing the data set preprocessing, the data after cleaning, enhancement, normalization, etc. is divided into training set, validation set and test set according to a certain ratio. This step is to provide data sets with different functions for subsequent loss function training. The training set is used for model training and learning parameters, the validation set is used to evaluate the model performance during the training process to adjust the hyperparameters, and the test set is used to finally evaluate the generalization ability of the trained model. In the loss function training stage, the divided training set data needs to be loaded into the model in batches, which is usually achieved using a data loader (DataLoader). It can automatically load data in batches during the training process, and can perform operations such as data shuffling and parallel loading to improve training efficiency. Convert to the format and tensor type required for model input to ensure that the data can be correctly input into the model for forward propagation calculation. During the loss function training process, continuously monitor the changes in the model's training loss, validation loss, and other related indicators (such as accuracy, recall rate, etc.). By analyzing the changing trends of these indicators, determine whether the model has problems such as overfitting, underfitting, or slow convergence. For example, if the training loss continues to decrease but the validation loss starts to increase at a certain moment, it means that the model may be overfitting; if both the training loss and the validation loss decrease slowly, it may be that the learning rate is improperly set or the model structure is unreasonable, resulting in slow convergence. Based on the monitoring and analysis results of the loss function training process, determine the direction of algorithm optimization, such as If the model is found to be overfitting, you can consider adding regularization terms, reducing model complexity, increasing the Dropout probability, etc.; if the convergence speed is slow, you can try adjusting the learning rate, changing the optimizer or optimizing its parameters, etc. According to the determined algorithm optimization direction, make specific adjustments and improvements to the model training algorithm. For example, after adjusting the learning rate, restart the model training and observe the performance of the model during the training process to see if it can improve the previous problems. If the optimizer is changed, the optimizer parameters need to be reinitialized and the model training should be performed under the new optimizer. During the model training process after the algorithm is optimized, continue to use the validation set to evaluate the model to observe whether the performance indicators of the model have improved. When the model performs well on the validation set, When the performance reaches a certain level of satisfaction or no longer shows significant improvement, the algorithm optimization is considered to have achieved certain results. At this point, the model can be finally evaluated using the test set to determine the model's generalization ability on unseen data. Based on the performance evaluation results of the model after algorithm optimization, further consideration can be given to optimizing the model structure. If it is found that the model still performs poorly in some aspects, such as low accuracy in detecting certain types of caves, the model architecture may need to be adjusted, such as increasing or decreasing the number of network layers, changing the size of the convolution kernel, introducing new modules, etc., and then the model is trained again, repeating the process of loss function training, algorithm optimization, and model performance evaluation, and continuously iterating and optimizing the model until the model performance reaches the expected goal.

[0061] Example 3: The data correction and result output module includes the following steps:

[0062] B1. Inputting New Exploration Data: Newly collected cave exploration data, including geological radar data, seismic exploration data, and drill core data, is formatted and input into the trained deep learning model. B2. Model Processing and Correction: The deep learning model processes the input raw data based on knowledge learned during the training phase, including cave characteristic patterns, data distribution patterns, and data variation patterns under different geological conditions. B3. Multi-Source Data Fusion and Correction: After the model corrects single-source data, it fuses and corrects multi-source data to generate a geological structure report. The deep learning model combines its own prediction results to generate a detailed geological structure report for the cave area. The report covers key information such as the cave's location, shape, size, depth, and infill type. Data Visualization: The cross-sectional shape of the cave at different depths and the surrounding geological conditions are displayed in the form of a two-dimensional profile. The three-dimensional model vividly presents the three-dimensional form of the cave and its spatial distribution throughout the exploration area. The generated geological structure report containing visualization content is output in a common file format for easy viewing and printing.

[0063] In this embodiment, the newly collected cave area detection data, including geological radar data, seismic exploration data, and borehole coring related data, are unified and formatted and input into the trained deep learning model. These data are uncorrected raw data and may have noise interference, measurement errors, and abnormal data points due to complex geological conditions. The deep learning model processes the input raw data based on the knowledge of cave characteristic patterns, data distribution laws, and data change laws under different geological conditions learned during the training phase. For example, for abnormal reflection signals caused by electromagnetic interference in geological radar data, the model can correct the abnormal signals to signals that conform to the actual geological conditions based on the normal reflection signal characteristics mastered during the training process. For abnormal wave velocity values ​​caused by uneven strata in seismic exploration data, the model uses the learned relationship between strata and wave velocity to correct the abnormal wave velocity values. In this process, the model will analyze and judge each data point, and adjust the data based on its internal parameters and algorithm logic. and correction. Since the data obtained by different detection methods have their own advantages and disadvantages, after the model corrects the single-source data respectively, it will also perform multi-source data fusion correction. For example, the corrected geological radar image features are integrated with the wave velocity characteristics of seismic exploration and the rock composition characteristics obtained by drilling core sampling. By fusing the corrected data from different data sources, the accuracy and completeness of the data are further improved, and the limitations that may exist in a single data source are eliminated. After data correction, the deep learning model combines its own prediction results to generate a detailed geological structure report for the cave area. The report content covers key information such as the location, shape, size, depth, and filling type of the cave. The model calculates and obtains these specific geological information through a comprehensive analysis of the corrected data and the mapping relationship between the features and geological structure parameters established during the training process. In order to make the results more intuitive and easy to understand, and convenient for engineers and decision makers to use, the generated geological structure report data will be visualized to display the cross-sectional shape of the cave at different depths and the surrounding geological conditions in the form of a two-dimensional profile.

[0064] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A deep learning-based cave area detection and correction system, characterized by: It includes a data acquisition and preprocessing module, a deep learning model building module, and a data correction and result output module. Data acquisition and preprocessing includes multi-source data acquisition and data preprocessing, which are used to process the collected database. The deep learning model building module automatically extracts features and performs pattern recognition on multi-source geophysical data. The data correction and result output module converts the original model output into cave detection results through multi-dimensional verification, physical constraint fusion, and visual expression. The deep learning model construction module includes deep learning model architecture and model training and optimization. The deep learning model architecture includes convolutional neural networks, recurrent neural networks and their variants, and adversarial networks. The convolutional neural network adopts the U-Net architecture to capture the spatial features in geological radar image data, and uses a bidirectional LSTM network to process seismic exploration data. The adversarial network uses a conditional generative adversarial network based on the traditional GAN, and additional input conditional information controls the properties of the generated data. Model training and optimization maps the data to a unified feature space through comparative learning to eliminate the dimensional differences of different geophysical methods.

2. The deep learning-based cave area detection and correction system according to claim 1, characterized in that: Model training and optimization include training data set preprocessing, loss function training, algorithm optimization, and model optimization; In addition to collecting a large amount of multi-source exploration data from known cave areas and their corresponding real geological conditions and performing conventional divisions, the training data preprocessing method combines data oversampling and undersampling to process unbalanced data sets. The loss function training adopts the focus loss function for classification tasks, adjusts the degree of attention to samples of different difficulty levels, and increases the learning weight of difficult-to-classify samples; Algorithm optimization adds weight decay to the Adam algorithm to dynamically adjust the learning rate; Model optimization uses the validation set to monitor the model's accuracy, recall rate, mean square error and other conventional indicators, and introduces F1 score and average precision to conduct a more comprehensive evaluation of model performance.

3. The deep learning-based cave area detection and correction system according to claim 2, characterized in that: Multi-source data acquisition includes geological radar, seismic surveys, and drill core sampling; Geological radar emits high-frequency electromagnetic waves and detects the distribution of underground media by receiving reflected waves, thus obtaining the general outline and location information of the cave. Seismic exploration uses the differences in propagation characteristics of artificially excited seismic waves in different geological media to analyze the information of reflected and refracted waves to determine the depth and size of caves; Drilling and coring directly obtain underground rock samples, and through core analysis, the composition of the filling material and rock properties in the cave can be accurately determined.

4. The deep learning-based cave area detection and correction system according to claim 3, characterized in that: Data preprocessing includes denoising, data standardization and data fusion; Denoising is done by using wavelet transform, adaptive filtering and other methods to remove electromagnetic interference noise in geological radar data and environmental noise in seismic exploration data. Data standardization: Standardize data from different sources and dimensions to a unified scale for subsequent data analysis and model training. Data fusion integrates pre-processed geological radar, seismic exploration and drill core data.

5. The deep learning-based karst cave detection and correction system according to claim 4, characterized in that: The deep learning model architecture consists of the following steps: A1. Extracting multi-scale features of caves through convolutional neural networks; A2. Use U-Net++ or V-Net to perform voxel-level segmentation of caves and output cave probability cubes; A3. A bidirectional LSTM network analyzes borehole acoustic logging time series data to determine the vertical development of caves.

6. The deep learning-based cave area detection and correction system according to claim 5, characterized in that: The data correction and result output module includes the following steps: B1. Input new exploration data: Newly collected cave area exploration data, including geological radar data, seismic exploration data, and drill core data, are formatted and input into the trained deep learning model. B2. Model processing and correction: The deep learning model processes the input raw data based on the knowledge learned during the training phase, such as cave characteristic patterns, data distribution patterns, and data variation patterns under different geological conditions; B3. Multi-source data fusion correction: After the model corrects the single-source data separately, the multi-source data is fused and corrected. B4. Generate a geological structure report. The deep learning model combines its own prediction results to generate a detailed geological structure report of the cave area. The report covers key information such as the location, shape, size, depth, and filling type of the cave. B5. Data visualization: Display the cross-sectional shape of the cave at different depths and the surrounding geological conditions in the form of a two-dimensional cross-section diagram, and vividly present the three-dimensional form of the cave and its spatial distribution in the entire detection area through a three-dimensional model; B6. Output the generated geological structure report file containing visualization content.

7. The deep learning-based cave area detection and correction system according to claim 2, characterized in that: Loss function training includes the following steps: C1. Clarify the task type and loss function selection: In cave detection, if you need to determine whether a cave exists or not, choose the focal loss function. If you need to predict continuous values ​​such as the size and depth of a cave, use the Huber loss function. C2. Initialize model parameters: After selecting the loss function, initialize the parameters of the deep learning model; C3, Forward Propagation: The training data (including multi-source data such as geological radar and seismic exploration) is input into the initialized model. The model processes the input data layer by layer according to its own architecture and parameter settings. After a series of operations, the model outputs the prediction results. C4. Calculate the loss value: Compare the model's prediction results with the corresponding true labels, and calculate the difference between the prediction results and the true values ​​according to the selected loss function to obtain the loss value; C5. Backpropagation and parameter update: Based on the calculated loss value, the gradient of the loss value with respect to the model parameters is calculated through the backpropagation algorithm. The AdamW optimization algorithm is used to update the model parameters according to the calculated gradient. The updated model parameters will be used for the next round of forward propagation and backpropagation.

8. The deep learning-based karst cave detection and correction system according to claim 3, characterized in that: Algorithm optimization includes the following steps: D1. Clarify optimization goals; D2. Data cleaning and preprocessing enhancement: Perform more detailed cleaning of raw data; D3, data expansion and balancing; D4. Loss function optimization: Try different loss function combinations or improve existing loss functions. In addition to the focal loss function, combine the triplet loss function; D5. Hyperparameter tuning: Determine the hyperparameters that need to be tuned and set a reasonable value range for each hyperparameter.

Images