A geological structure recognition method based on deep learning

Abstract

The present application relates to the field of image recognition, and more particularly to a geological structure recognition method based on deep learning. First, the original geological image collected is preprocessed, and features are extracted based on the preprocessed geological image to generate a feature map; based on the feature map, the gradient vector and gradient norm of the pixel points are calculated; based on the gradient norm of the pixel points, the local connectivity factor is calculated; then, based on the gradient norm of the pixel points and the local connectivity factor, the reconstructed feature map is obtained by processing using a dynamic boundary feature reconstruction algorithm; finally, based on the reconstructed feature map, the geological structure is classified by a geological structure probability classification algorithm to generate a classification probability vector. The technical problems of insufficient precision of automatic geological structure recognition, limitations of boundary feature extraction, imbalance between feature enhancement and noise suppression, and insufficient generalization ability of geological structure classification are solved.

Description

Technical Field

[0001] This invention relates to the field of image recognition, and more particularly to a method for identifying geological structures based on deep learning. Background Technology

[0002] Geological structure identification is a crucial research area in geology and geophysics. Its purpose is to identify structural features within subsurface rock strata, such as faults, folds, and fracture systems, by analyzing geological data, including seismic data, well logging data, and core images. Geological structures are not only essential for understanding geological evolution but also directly impact practical applications such as oil and gas exploration, mineral resource development, and earthquake hazard assessment. Traditional methods for geological structure identification primarily rely on manual interpretation or physical model-based processing techniques, which suffer from low efficiency, high subjectivity, and limited ability to handle complex data. In recent years, the rapid development of deep learning technology has provided a novel solution for geological structure identification. Deep learning's powerful feature extraction and data processing capabilities have significantly improved the accuracy and efficiency of identification. Summary of the Invention

[0003] This invention provides a deep learning-based method for identifying geological structures, which solves the technical problems of insufficient accuracy in automatic identification of geological structures, limitations in boundary feature extraction, imbalance between feature enhancement and noise suppression, and insufficient generalization ability of geological structure classification.

[0004] The present invention provides a geological structure identification method based on deep learning, which specifically includes the following technical solutions: A deep learning-based method for identifying geological structures includes the following steps: S1. Perform image preprocessing on the acquired raw geological images to obtain preprocessed geological images; extract features from the preprocessed geological images to generate feature maps; calculate the gradient vector and gradient norm of each pixel based on the feature maps; calculate the local connectivity factor based on the gradient norm of each pixel. S2. Based on the gradient norm and local connectivity factor of pixels, the reconstructed feature map is obtained by using a dynamic boundary feature reconstruction algorithm. S3. Based on the reconstructed feature map, the geological structures are classified using a geological structure probability classification algorithm to generate a classification probability vector.

[0005] Preferably, S1 specifically includes: The size of the feature map is the height multiplied by the width multiplied by the number of feature channels.

[0006] Preferably, S1 specifically includes: For each channel of the feature map, a gradient vector is calculated at each pixel. The gradient vector consists of the partial derivatives of the eigenvalues ​​in the horizontal and vertical directions.

[0007] Preferably, S1 specifically includes: Based on the gradient vector, the gradient norm of each pixel is calculated by taking the square root of the sum of the squares of the partial derivatives of the eigenvalues ​​in the horizontal and vertical directions.

[0008] Preferably, S1 specifically includes: For each pixel in each channel of the feature map, define the neighborhood of the pixel, which includes the center pixel and the neighboring pixels around the center pixel; calculate the gradient similarity between the gradient vector of each neighboring pixel in the neighborhood and the gradient vector of the center pixel; take the average of the gradient similarity between the gradient vectors of all neighboring pixels in the neighborhood and the gradient vector of the center pixel to obtain the local connectivity factor of the pixel.

[0009] Preferably, S2 specifically includes: In the implementation of the dynamic boundary feature reconstruction algorithm, for each pixel of each feature channel, the square of the gradient norm is calculated and normalized by an exponential function to obtain the gradient Gaussian weight; based on the gradient Gaussian weight and the local connectivity factor, the weighted sum is normalized to generate the dynamic reconstruction factor.

[0010] Preferably, S2 specifically includes: Based on the dynamic reconstruction factor, the weights of the feature map are adjusted to generate the reconstructed feature map.

[0011] Preferably, S2 specifically includes: In the process of adjusting the weights of the feature map, a learnable channel weight vector is introduced. For each channel, an enhancement factor is generated based on the gradient norm and the learnable channel weight vector. The enhancement factor is then multiplied by the feature map to generate the reconstructed feature map.

[0012] Preferably, S3 specifically includes: In the implementation of the geological structure probabilistic classification algorithm, the classification probability vector is obtained through three operations: global average pooling, fully connected layer mapping, and probability normalization.

[0013] The beneficial effects of the technical solution of the present invention are: 1. High-resolution cameras are used to capture geological profiles, revealing fine structural details of outcrops or rock surfaces, generating high-quality raw geological images. Image preprocessing, including normalization and denoising, effectively removes noise interference and enhances image clarity, laying the foundation for subsequent feature extraction. High-quality image input significantly improves the accuracy of feature extraction, providing reliable data support for subsequent analysis.

[0014] 2. A pre-trained convolutional neural network is used to extract features from the preprocessed geological image, generating feature maps containing information such as edges and textures. By calculating the gradient norm, the boundary strength of the feature map is quantified, highlighting key features of geological structures, such as the boundary regions of faults or folds. At the same time, a local connectivity factor is introduced, and by analyzing the gradient similarity of neighboring pixels, the continuity of boundary features is quantified, effectively capturing the local and global features of geological structures, enhancing the feature map's ability to represent complex geological structures, and providing a high-precision feature foundation for subsequent boundary enhancement and classification.

[0015] 3. Based on gradient norm and local connectivity factor, a dynamic boundary feature reconstruction algorithm is used to calculate dynamic reconstruction factor. The feature map weight is adjusted by the dynamic reconstruction factor, which significantly enhances the feature expression of the boundary region and suppresses the interference of the background region. The reconstructed feature map not only retains the spatial structure of the original features, but also highlights the boundary information of the geological structure, making structural features such as folds and faults clearer and more distinguishable, thus improving the recognition accuracy of geological structures. Especially in complex geological environments, it can effectively distinguish different types of structural features.

[0016] 4. Based on the reconstructed feature map, the geological structure probabilistic classification algorithm transforms the high-dimensional feature map into a low-dimensional classification probability vector through global average pooling, fully connected layer mapping, and probability normalization, generating a probability distribution for each geological structure category. This enables automatic classification of geological structures such as folds and faults, significantly reducing the workload of manual analysis while improving the accuracy and consistency of classification. The output of the probability distribution provides geologists with intuitive structural identification results, supporting subsequent geological analysis and decision-making. Attached Figure Description

[0017] Figure 1 This is a flowchart of a deep learning-based geological structure identification method according to the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0020] The following description, in conjunction with the accompanying drawings, details a specific scheme for a deep learning-based geological structure identification method provided by the present invention.

[0021] See attached document Figure 1 The diagram illustrates a flowchart of a deep learning-based geological structure identification method according to an embodiment of the present invention. The method includes the following steps: S1. Perform image preprocessing on the acquired raw geological images to obtain preprocessed geological images; extract features from the preprocessed geological images to generate feature maps; calculate the gradient vector and gradient norm of each pixel based on the feature maps; calculate the local connectivity factor based on the gradient norm of each pixel. Using high-resolution cameras, such as the Canon EOS 5D Mark IV with a resolution of 6720×4480 pixels, geological profiles were photographed in the field to record local structural details of outcrops or rock surfaces. These were then compiled into a raw geological image dataset. All raw geological images were annotated by professional geologists, with annotations including the type of geological structure, such as faults, folds, and fissures, as well as their boundary locations. Annotation employed both pixel-level segmentation and image-level classification to ensure the accuracy of feature extraction and classification tasks. The raw geological image dataset was divided into a training set (4500 images) and a test set (500 images) at a 9:1 ratio to support subsequent model training and performance evaluation.

[0022] Image preprocessing, including normalization and denoising, is performed on the original geological image to obtain the preprocessed geological image. The normalization process uses min-max normalization, mapping the pixel values ​​of the original geological image to the [0,1] interval. This effectively addresses the extreme differences in pixel values ​​caused by strong light areas, such as reflections from rock surfaces or shadow areas, in the original geological image. Furthermore, min-max normalization is computationally simple, preserves the relative intensity relationships of the original geological image, and helps improve the training stability of the convolutional neural network, avoiding gradient explosion or vanishing problems caused by excessively large pixel value ranges. Denoising is performed using bilateral filtering. Simultaneously suppressing noise while preserving boundary details in the original geological image, since the original geological image may contain both salt-and-pepper noise caused by camera acquisition interference and Gaussian noise caused by ambient lighting, bilateral filtering, by combining spatial and value domain Gaussian kernels, can effectively suppress noise while preserving boundary details of geological structures, such as the edges of faults or folds. Compared to Gaussian filtering, which blurs boundaries, or median filtering, which has limited effect on suppressing Gaussian noise, bilateral filtering performs better when dealing with complex textures and noise types in the original geological image. The image preprocessing operations used are all techniques well known to those skilled in the art and will not be described in detail here.

[0023] Feature maps are generated by extracting features from pre-trained convolutional neural networks on pre-processed geological images in the training set. The feature map has a height of Multiply by width Multiply by the number of feature channels The number of feature channels is either 64 or 128. When dealing with simple geological structures, such as single faults or folds, C=64 is sufficient to capture the main edge and texture features with low computational complexity. When dealing with complex geological structures, such as multiple faults or complex folds, C=128 can provide rich feature representation and capture subtle boundary details, but it will increase the computational cost by about 2 times. Therefore, the trade-off in the choice of the number of channels is based on the texture complexity of the original geological image and the limitations of computational resources. The pre-trained convolutional neural network architecture is ResNet-50. Based on the deep residual structure of ResNet-50, it can effectively extract multi-scale features, making it suitable for complex textures in geological images, ranging from macroscopic bedding to microscopic fault zones. The pre-training dataset is ImageNet, which provides general edge and texture feature extraction capabilities. To adapt to the texture features of geological images, a transfer learning strategy is adopted: first, pre-training is performed using ImageNet, and then fine-tuning is done on geological image samples. The fine-tuning dataset uses a geologically specific dataset, such as GeoImageNet, containing approximately 10,000 labeled rock profiles and outcrop images, covering various structural types such as folds, faults, and bedding. The fine-tuning process uses the Adam optimizer, with a learning rate of [missing information]. The algorithm is trained for 20 epochs to adapt to the special texture features of geological images, such as the periodicity of rock bedding or the discontinuity of fault fracture zones, thereby improving the specificity and robustness of feature extraction.

[0024] For each channel of the feature map, at each pixel location , i.e., pixel The gradient vector is calculated, consisting of the partial derivatives of the eigenvalues ​​in the horizontal and vertical directions. It represents the direction and intensity of change in the feature map at each pixel. The calculation formula is: , in, The feature map is represented by the first... Channels at pixels The gradient vector at that point; The feature map is represented by the first... Channels at pixels Eigenvalues ​​at; The feature map is represented by the first... Channels at pixels The horizontal partial derivative at that point; The feature map is represented by the first... Channels at pixels The partial derivative in the vertical direction at that point.

[0025] Based on the gradient vector, the gradient norm, i.e., the gradient magnitude, is calculated for each pixel. It is obtained by calculating the square root of the sum of the squares of the partial derivatives of the feature values ​​in the horizontal and vertical directions. The gradient norm is a scalar that reflects the boundary strength of the feature map at each pixel. The calculation formula is as follows: , in, The feature map is represented by the first... Channels at pixels The gradient norm at point 1, i.e., the feature map at point 2. Channels at pixels The gradient magnitude at that point is used to quantize the boundary strength.

[0026] To measure the continuity of boundary features in a feature map, a local connectivity factor is calculated. Specifically, for each channel of the feature map at each pixel (a pixel is typically defined as the smallest unit in a digital image, representing an image unit with a fixed location and a specific color or brightness value), a 3×3 pixel neighborhood is defined, including the center pixel and its eight neighboring pixels. The neighborhood size is appropriate, capturing the continuity of local features without introducing excessive noise or irrelevant information. For each pixel within the neighborhood, the gradient similarity between the gradient vectors of the neighboring pixels and the center pixel is calculated. This gradient similarity is calculated using cosine similarity, which is obtained by dividing the dot product by a factor of 1 / 2. The product of the magnitudes of two gradient vectors measures the directional consistency between them. A value close to 1 indicates a high degree of directional consistency, suggesting that the two pixels may belong to the same continuous boundary, such as the linear features of a fault. To avoid division by zero, a very small constant is added to the denominator during calculation. Based on the gradient similarity of all pixels in the neighborhood, a local connectivity factor is calculated. The gradient similarity of each neighboring pixel is mapped to a positive range using an exponential function to enhance the influence of high-similarity regions. The average gradient similarity of all pixels in the neighborhood is taken as the local connectivity factor of the pixel. The level of the local connectivity factor reflects the continuity of the pixel with its neighboring pixels in terms of boundary features. The specific calculation formula is as follows: , in, The feature map is represented by the first... The channel is at the center pixel. The local connectivity factor at point 1 measures the local connectivity factor at point 2. The channel is at the center pixel. The continuity of boundary features at the location; The feature map is represented by the first... The channel is at the center pixel. The local neighborhood of the location; The feature map is represented by the first... The channel is at the center pixel. The number of pixels in the local neighborhood of a given location; This represents the summation of pixels in the neighborhood, used to aggregate the gradient similarity contributions of all pixels in the neighborhood, thereby calculating the local connectivity factor; This means that the gradient similarity is mapped to a positive range through an exponential function, which enhances the weight of high similarity regions and highlights the coherence of boundary features; Representing neighboring pixels The gradient similarity between the gradient vector of the center pixel and the gradient vector of the center pixel is used to measure the gradient similarity of the center pixel. and neighboring pixels The gradient vector has consistent direction, and the calculation formula is: , in, The feature map is represented by the first... Channels at neighboring pixels The gradient vector at that point; The feature map is represented by the first... The channel is at the center pixel. The gradient vector at point and at neighboring pixels The dot product of the gradient vectors at a given point is used to calculate the consistency of the directions of the two gradient vectors; the larger the dot product, the more similar the directions. The feature map is represented by the first... Channels at neighboring pixels The gradient norm at point 1, i.e., the feature map at point 2. Channels at neighboring pixels The gradient magnitude at that point is used to quantize the boundary strength; Represents a very small constant, used to avoid division by zero, such as ; The feature map is represented by the first... The channel is at the center pixel. The gradient norm and in neighboring pixels The product of the gradient norms at each point, used as the denominator, normalizes the dot product of the gradient vectors.

[0027] S2. Based on the gradient norm and local connectivity factor of pixels, the reconstructed feature map is obtained by using the dynamic boundary feature reconstruction algorithm. Based on the gradient norm and local connectivity factor of the feature map, the coherence and discriminativeness of the boundary features are enhanced by a dynamic boundary feature reconstruction algorithm, and the reconstructed feature map is obtained. The dynamic boundary feature reconstruction algorithm emphasizes the weight of high gradient regions, i.e., the weight of boundary regions, through gradient Gaussian weighting, while suppressing the contribution of low gradient regions, i.e., the contribution of background regions. Specifically, for each pixel of each feature channel, the square of the gradient norm is calculated and normalized by an exponential function. The denominator of the exponential function is a normalization parameter, which is set as the standard deviation of the gradient norm to control the weighting intensity. To avoid the mismatch between channel weights and boundary features caused by static calculation of dynamic reconstruction factors and dynamic learning of learnable channel weight vectors (e.g., assigning high weights to channels with weak boundary features, leading to excessive amplification of background noise), a collaborative optimization mechanism is introduced. This mechanism utilizes learnable scalar parameters... Adaptive adjustment is performed, with the learnable scalar parameters optimized jointly through gradient descent and the learnable channel weight vector to ensure that the calculation of the dynamic reconstruction factor dynamically adapts to the importance of the channel weights. The gradient Gaussian weights and local connectivity factors are multiplied to obtain the weighted value for each pixel. This summation and normalization over the spatial dimensions of the entire feature map (height and width) is combined with the learnable scalar parameters to generate the dynamic reconstruction factor. The dynamic reconstruction factor integrates boundary strength and boundary coherence and is used to adaptively adjust the feature weights. The calculation formula is as follows: , in, This represents the dynamic reconstruction factor, which measures the strength and coherence of boundary features in the feature map; The feature map is represented by the first... The learnable scalar parameters of the channel; Represents the first feature map Channels at pixels The weighted values ​​at each point are summed to aggregate the weighted contributions of all pixels. The weighted values ​​consist of two parts: the gradient Gaussian weight and the local connectivity factor. This represents an exponential function, i.e., gradient Gaussian weights, based on the feature map. Channels at pixels Gaussian weighted by the square of the gradient norm at the given point; The normalization parameter is set as the standard deviation of the gradient norm of the feature map. That is, the standard deviation of the gradient norm of all pixels is calculated for each feature channel of each feature map. Since the gradient distribution of geological images varies greatly due to the image content, such as rock type and structural complexity, the local statistics of a single geological image can reflect the gradient characteristics of the current feature map, avoiding the global bias introduced by batch or full training set statistics, thereby ensuring the stability and specificity of the gradient Gaussian weight. This represents the total number of pixels in the spatial dimension. It is used for normalization to ensure that the value of the dynamic reconstruction factor is within a reasonable range and to prevent numerical instability caused by differences in feature map size.

[0028] Furthermore, the weights of the feature maps are adjusted using a dynamic reconstruction factor. Specifically, for each channel, the dot product of the gradient norm and the learnable channel weight vector is calculated and mapped to the hyperbolic tangent function. Range, enhancing the contrast of boundary features; The learnable channel weight vectors are optimized using the Adam optimizer, reflecting the importance of different feature channels for geological structure identification. The Adam optimizer combines the advantages of momentum and adaptive learning rates, making it suitable for handling complex non-convex optimization problems and accelerating gradient descent convergence. The hyperparameters are set as follows: the initial learning rate is... An exponential decay strategy was adopted, with the batch size decreasing to 0.9 times the original size every 10 epochs to ensure the stability of later training. The batch size was set to 32 to balance computational efficiency and gradient estimation stability. The training lasted for a total of 100 epochs, which was sufficient for the model to converge while avoiding overfitting. The dynamic reconstruction factor is multiplied by the mapped value and then incremented by 1 to form an enhancement factor, which is used to multiply the feature map element by element to generate the reconstructed feature map. The reconstructed feature map not only retains the spatial structure of the original features but also enhances the boundary information of the geological structure. The formula for calculating the reconstructed feature map is: , in, This represents the reconstructed feature map; Indicates the enhancing factor; This represents the dynamic reconstruction factor, used to measure the strength and coherence of boundary features in the feature map; Represents the hyperbolic tangent function, mapping the input value to... Intervals are used to enhance the contrast of boundary features while avoiding excessive weight changes and maintaining the stability of feature enhancement. The gradient norm of the feature map is used to quantify the boundary strength of the feature map. This represents a learnable channel weight vector, where each element corresponds to the weight of a feature channel. It is automatically learned through backpropagation during the training process of deep learning models such as convolutional neural networks.

[0029] S3. Based on the reconstructed feature map, the geological structures are classified using a geological structure probability classification algorithm to generate a classification probability vector; Based on the reconstructed feature map, geological structures such as folds and faults are classified using a geological structure probability classification algorithm, generating a classification probability vector that reflects the probability distribution of each geological structure category. The geological structure probabilistic classification algorithm transforms high-dimensional feature maps into low-dimensional classification probability vectors through three sub-steps: global average pooling, fully connected layer mapping, and probability normalization. It is suitable for automatically identifying geological structures from geological images. Global average pooling calculates the spatial dimensional average of each channel in the reconstructed feature map. Specifically, it averages all feature values ​​in the height and width dimensions of the spatial grid for each channel, generating a fixed-length feature vector. The length of the feature vector is equal to the number of channels in the feature map, and each element represents the average feature intensity of the corresponding channel. The purpose of global average pooling is to compress the high-dimensional reconstructed feature map into a low-dimensional feature vector, retaining the main feature information while significantly reducing computational complexity. The fully connected layer mapping operation maps feature vectors to a new vector space with a dimension equal to the number of geological structure categories using a learnable weight matrix. Each element of the feature vector is multiplied by the corresponding row of the weight matrix to generate a score vector reflecting the correlation between the feature vector and each geological structure category. The training process of the learnable weight matrix of the fully connected layer is similar to... and By sharing the same optimizer and loss function, namely using the Adam optimizer and cross-entropy loss, collaborative optimization is ensured. The probability normalization operation normalizes the score vector by applying the softmax function, outputting a classification probability vector that reflects the probability distribution of each geological structure category. The calculation formula is as follows: , in, Represents the classification probability vector; Represents the normalization function; Indicates a fully connected layer; This represents global average pooling, used to compute the reconstructed feature map. Spatial mean of each channel; This represents the weight matrix of the fully connected layer. It is a learnable weight matrix that is automatically learned through backpropagation during the training process of the deep learning model.

[0030] To verify the advantages of the deep learning-based geological structure recognition method, geologists manually analyzed geological images in the test set, recording the structure types and boundary locations as a baseline. Manual interpretation relies on expert experience, is time-consuming, and is easily influenced by subjectivity. Simultaneously, a pre-trained ResNet-50 model was used to directly classify the pre-processed geological images in the test set, without introducing a dynamic boundary feature reconstruction algorithm, to test the ability to recognize geological structures. To quantitatively evaluate the effectiveness of different methods, the F1 score, which is the harmonic mean of precision and recall, is calculated to comprehensively evaluate the classification performance of different methods, demonstrating the superiority of a deep learning-based geological structure identification method compared with traditional manual interpretation methods and existing deep learning methods, such as the pre-trained ResNet-50 model.

[0031] In summary, a deep learning-based method for geological structure identification has been developed.

[0032] The order of the embodiments is for illustrative purposes only and does not represent the superiority or inferiority of the embodiments. The processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0033] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0034] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A geological structure identification method based on deep learning, characterized in that, Includes the following steps: S1. Perform image preprocessing on the acquired raw geological images to obtain preprocessed geological images; extract features from the preprocessed geological images to generate feature maps; calculate the gradient vector and gradient norm of each pixel based on the feature maps; calculate the local connectivity factor based on the gradient norm of each pixel. S2. Based on the gradient norm and local connectivity factor of each pixel, a dynamic boundary feature reconstruction algorithm is used. For each pixel in each feature channel, the square of the gradient norm is calculated and normalized using an exponential function to obtain the gradient Gaussian weight. Based on the gradient Gaussian weight and the local connectivity factor, a weighted sum is calculated and normalized to generate the dynamic reconstruction factor. The calculation formula is as follows: in, Represents the dynamic reconstruction factor; The feature map is represented by the first... The learnable scalar parameters of the channel; Indicates the gradient Gaussian weights; Indicates the normalization parameter; The total number of pixels in the spatial dimension; The feature map is represented by the first... The channel is at the center pixel. The local connectivity factor at the location; The feature map is represented by the first... Channels at pixels Gradient norm at; The feature map is represented by the first... Channels at pixels The gradient vector at that point; The feature map is represented by the first... Channels at pixels Eigenvalues ​​at; Based on the dynamic reconstruction factor, a learnable channel weight vector is introduced to adjust the weights of the feature map. For each channel, an enhancement factor is generated based on the gradient norm and the learnable channel weight vector. The enhancement factor is multiplied by the feature map to obtain the reconstructed feature map. The calculation formula is as follows: in, This represents the reconstructed feature map; Represents the feature map; Indicates the enhancing factor; Represents the hyperbolic tangent function; The gradient norm of the feature map; This represents the learnable channel weight vector, which is optimized using the Adam optimizer, reflecting the importance of different feature channels for geological structure identification. S3. Based on the reconstructed feature map, the geological structures are classified using a geological structure probability classification algorithm to generate a classification probability vector.

2. The geological structure identification method based on deep learning according to claim 1, characterized in that, S1 specifically includes: The size of the feature map is the height multiplied by the width multiplied by the number of feature channels.

3. The geological structure identification method based on deep learning according to claim 2, characterized in that, S1 specifically includes: For each channel of the feature map, a gradient vector is calculated at each pixel. The gradient vector consists of the partial derivatives of the eigenvalues ​​in the horizontal and vertical directions.

4. The geological structure identification method based on deep learning according to claim 3, characterized in that, S1 specifically includes: Based on the gradient vector, the gradient norm of each pixel is calculated by taking the square root of the sum of the squares of the partial derivatives of the eigenvalues ​​in the horizontal and vertical directions.

5. The geological structure identification method based on deep learning according to claim 3, characterized in that, S1 specifically includes: For each pixel in each channel of the feature map, define the neighborhood of the pixel, which includes the center pixel and the neighboring pixels around the center pixel; calculate the gradient similarity between the gradient vector of each neighboring pixel in the neighborhood and the gradient vector of the center pixel; take the average of the gradient similarity between the gradient vectors of all neighboring pixels in the neighborhood and the gradient vector of the center pixel to obtain the local connectivity factor of the pixel.

6. The geological structure identification method based on deep learning according to claim 1, characterized in that, S3 specifically includes: In the implementation of the geological structure probabilistic classification algorithm, the classification probability vector is obtained through three operations: global average pooling, fully connected layer mapping, and probability normalization.

Images