A residual structure-based U-Net shale authigenic mineral identification method

Abstract

The application relates to a U-Net shale authigenic mineral identification method based on a residual structure, and relates to the technical field of geophysics. The method comprises the following steps: acquiring sample backscattering, cathodeluminescence and element distribution image data; constructing a data set and making labels; inputting the data set into a U-Net neural network model based on a residual structure to perform training, so as to obtain a network model capable of identifying different types of authigenic minerals; and using the trained U-Net model to perform pixel-level fine segmentation on different types of authigenic minerals in a to-be-detected image, so as to realize high-precision automatic classification of different types of authigenic minerals. The method effectively solves the problems of great difficulty and low efficiency in identifying different types of authigenic minerals in fine-grained shale, and realizes the needs of rapid and accurate segmentation and quantitative analysis of a large number of samples, thereby providing important microscopic image basis for the research contents of shale oil and gas reservoir physical property changes and reservoir reconstruction schemes.

Description

Technical Field

[0001] This invention relates to the field of geophysical technology, specifically to a method for identifying authigenic minerals in U-Net shale based on residual structures. Background Technology

[0002] The formation of authigenic minerals in rocks significantly alters their pore structure and physical properties, and is also key to analyzing diagenesis and its co-evolution with organic matter. When authigenic minerals precipitate from pore fluids as cement, they encroach on pore space, significantly reducing porosity and permeability. When calcareous or siliceous cements form in the early diagenetic stage, these authigenic minerals can also increase the rock's compressive strength. Authigenic clay minerals can inhibit the formation of siliceous cements in the late diagenetic stage, protecting pore space; the development of intercrystalline pores in clay minerals can also increase the pore volume of shale. Different authigenic minerals also indicate differences in the acid-base conditions and composition of pore fluids. Therefore, the identification of authigenic minerals is fundamental to classifying diagenetic stages, a prerequisite for analyzing the evolution of pore-throat structure and permeability changes, and an important basis for clarifying hydrocarbon accumulation mechanisms, predicting high-quality reservoirs, and formulating reservoir stimulation strategies.

[0003] Authigenic minerals can usually be identified based on microscopic crystal morphology and cathodoluminescence characteristics. In sandstone and conglomerate, where grain size and intergranular pores are relatively large, authigenic minerals typically exhibit well-developed crystal morphology. However, shale, with its strong heterogeneity, small grain size, and well-developed micro- and nano-pores, presents greater challenges for authigenic mineral identification. Therefore, a combined approach of scanning electron microscopy (SEM) and cathodoluminescence is commonly used for comprehensive characterization. Traditional authigenic mineral content statistics employ thin-section methods for manual point counting. However, for small-grained shale, fine-grained minerals can be confused with peripheral minerals or pore fractures, leading to biases and reduced accuracy in mineral composition statistics using point counting. Furthermore, this method is inefficient for statistical analysis of large batches of fine-grained samples. Therefore, how to efficiently and accurately identify authigenic minerals in shale using high-resolution images remains a pressing problem. Summary of the Invention

[0004] To address the problems in existing technologies, this application proposes for the first time a U-Net method for identifying authigenic minerals in shale based on residual structures. It utilizes an image set composed of scanning electron microscopy backscattered emission, elemental distribution, and cathodoluminescence (CEL) images. Based on the microscopic morphology and cathodoluminescence characteristics of different authigenic minerals, combined with material composition information, it achieves accurate segmentation of different authigenic minerals. This meets the need for rapid and accurate segmentation and quantitative analysis of large batches of samples, providing important microscopic image evidence for subsequent research on pore-throat structure evolution and permeability changes, clarifying shale oil and gas enrichment mechanisms, predicting high-quality reservoirs, and designing reservoir stimulation schemes.

[0005] The technical problem solved by this invention is achieved by the following technical solution: A U-Net method for identifying authigenic minerals in shale based on residual structure includes the following steps:

[0006] A method for identifying authigenic minerals in shale based on residual structure U-Net, characterized by the following steps;

[0007] Step (1): Acquire sample images:

[0008] Rock samples were prepared into block-shaped samples measuring 1.0cm × 1.0cm × 2.5cm. The sample surfaces were argon-ion polished for observation under a scanning electron microscope. For each analytical field of view, backscattered images, elemental distribution images, and cathodoluminescence images were acquired at the same magnification, and these three types of images were combined to form an image set. Each image set includes information on the microstructure, material composition, and cathodoluminescence characteristics of the rock, providing a basis for the identification of authigenic minerals.

[0009] Step (2): Construct a self-generated mineral identification network model:

[0010] 1) Constructing and partitioning the image dataset;

[0011] 2) Construct a U-Net neural network model based on residual structure;

[0012] 3) Design the loss function and select evaluation indicators;

[0013] 4) Train a U-Net shale authigenic mineral model based on residual structure;

[0014] Based on the model built above, the image dataset formed by combining backscattered-elemental distribution-catholuminescence images and the corresponding label images are input into the U-Net model based on residual structure for training. The training process of the model is monitored by loss function and validation set evaluation index. Finally, the segmentation accuracy of the model is tested using test set data to see if it meets the requirements.

[0015] Step (3): Identify authigenic minerals based on the U-Net model with residual structure.

[0016] Backscattered images, elemental distribution maps, and cathodoluminescence images of the region to be detected are acquired, forming image groups and cropping them into image patches. These patches are then input into a U-Net model based on residual structure to predict the target authigenic minerals in the image patches, resulting in segmented images of the target authigenic minerals. The image patches are then stitched together to obtain the authigenic mineral picking results for each image group.

[0017] The specific steps for constructing and partitioning the image dataset in step (2) above are as follows:

[0018] Based on the microstructure, material composition, and cathodoluminescence image information of each image group, different authigenic minerals are manually labeled with different category names to generate mask images. A mapping dictionary is constructed to realize the mapping relationship between the RGB color values ​​of the mask images and the category index. Each image group and its corresponding mask image are cropped into 512×512 pixel image blocks. The image block dataset is randomly divided into training set, validation set, and test set as the data basis for model training and evaluation.

[0019] The specific steps in step (2) above for building the U-Net neural network model based on the residual structure are as follows:

[0020] The U-Net model based on residual structure uses the traditional U-Net network structure. The encoder initially consists of one 7×7 convolutional layer and one 3×3 max-pooling layer, downsampling the input image size to 1 / 2 and 1 / 4 of its original size, respectively. The convolutional layer adjusts the number of channels to 64. The encoder path then uses a 4-level residual structure instead of the conventional convolutional layers. Each residual structure consists of two residual modules (BasicBlock) to enhance the multi-scale feature extraction capability of authigenic minerals and the ability to model the overall structure of large targets using global context information. Each residual module contains two... The encoder consists of 3×3 convolutional layers, each followed by a batch normalization layer and a ReLU activation function. For the second convolutional layer, its output is element-wise added to the features of the short-circuit connection branches before applying ReLU activation to achieve residual fusion, preserving effective features from shallow layers while enhancing the expressive power of deep features. The encoder's 4-level residual structure outputs feature maps at each level, which are then input to the corresponding level of the decoder via skip connections. The decoder uses bilinear interpolation for upsampling and fuses encoded features through skip connections. In the final stage, the decoder adjusts the channel dimensions using 1×1 convolutions to achieve the segmentation of the target authigenic minerals.

[0021] The design of the loss function and the selection of evaluation indicators in step (2) above are as follows:

[0022] During model training, the loss function and evaluation metrics jointly supervise the model training direction and drive parameter optimization. This step uses a composite loss function as the loss function to supervise model training, quantifying the deviation between the model's prediction results and the true labels. The closer the loss function value is to 0, the smaller the error between the model's prediction results and the true labels, and the better the fitting effect. The composite loss function is a combination of the Cross Entropy Loss function and the Dice loss function, specifically:

[0023] ;

[0024] Where Loss is the total loss function, Loss CEIt is the cross-entropy loss function, Loss Dice This is the Dice loss function. To measure the model's segmentation ability, the Dice score coefficient and pixel accuracy (PA) are chosen as evaluation metrics for the network model. The Dice score coefficient measures the degree of overlap between the predicted region and the ground truth region. The closer the value is to 1, the higher the overlap between the model's predicted region and the ground truth region, and the better the pixel segmentation matching effect. Pixel accuracy is the proportion of correctly classified pixels out of the total pixels, reflecting the overall pixel-level accuracy. The closer the value is to 1, the more pixels the model correctly predicts, and the better the overall prediction effect. The specific definitions of the evaluation metrics are as follows:

[0025] ;

[0026] ;

[0027] Wherein, TP is the number of pixels predicted as the current category and actually belonging to the current category, FP is the number of pixels predicted as the current category but actually belonging to all other categories, FN is the number of pixels predicted as all other categories but actually belonging to the current category, and TN is the number of pixels actually belonging to all other categories and also predicted as all other categories.

[0028] Compared with the prior art, this application has the following advantages:

[0029] This invention establishes a method for identifying authigenic minerals in shale using a U-Net model based on residual structures. Firstly, this method uses an image dataset composed of backscattered images, elemental surface distribution images, and cathodoluminescence images from the same field of view as input. This dataset fully reflects the microscopic morphological features, elemental composition, and cathodoluminescence characteristics of different authigenic minerals, providing multiple information sources for feature extraction of target authigenic minerals and increasing the accuracy of target authigenic mineral identification. Secondly, this invention uses residual modules to replace the convolutional layers in the traditional U-Net network, enhancing the network model's ability to extract multi-scale features of target authigenic minerals and its ability to model the overall structure of large targets based on global context information. Compared with existing technologies, this method can segment fine-grained minerals from the surrounding matrix, significantly improving the efficiency of identifying target authigenic minerals in large batches of fine-grained samples. This provides preliminary microscopic evidence for subsequent research work such as exploring the enrichment mechanism of shale oil and gas and predicting high-quality reservoirs. Attached Figure Description

[0030] Figure 1 Flowchart of the U-Net shale authigenic mineral identification method based on residual structure of the present invention;

[0031] Figure 2 The backscattering-elemental distribution-catholuminescence image group in the embodiments of the present invention;

[0032] Figure 3 The color mask image in the embodiments of the present invention;

[0033] Figure 4 : Overall network structure diagram in this embodiment of the invention;

[0034] Figure 5 : Residual module structure diagram in an embodiment of the present invention;

[0035] Figure 6 The backscattered-elemental distribution-catholuminescence image group acquired in this embodiment of the invention;

[0036] Figure 7 : The predicted distribution map of authigenic minerals in shale in the embodiments of the present invention;

[0037] Figure 8 : A diagram showing the content of different types of authigenic minerals identified in the embodiments of this invention. Detailed Implementation

[0038] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. 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.

[0039] Example 1: The flowchart of a U-Net shale authigenic mineral identification method based on residual structure is as follows: Figure 1 As shown, the specific steps are as follows:

[0040] Step (1): Acquire sample images:

[0041] Shale samples were prepared into blocks measuring 1.0 cm × 1.0 cm × 2.5 cm, and the sample surfaces were polished with argon ions. A microscopic observation field was selected, and continuous backscattered images, elemental distribution images, and cathodoluminescence images were acquired at the same magnification for each local sub-field. Each image was 1024 × 768 pixels. In the elemental distribution image, red, green, blue-violet, and cyan represent the elements silicon, potassium, sodium, and calcium, respectively, indicating the spatial distribution of common minerals such as quartz, potassium feldspar, sodium feldspar, and carbonate minerals. Due to the small field of view of a single image, the acquired sub-field backscattered images, elemental distribution images, and cathodoluminescence images were stitched together to form three high-resolution backscattered images, elemental distribution images, and cathodoluminescence images required for the detection field. These three images from each field were then combined into an image group, such as... Figure 2 As shown, each image group includes the mineral microstructure, material composition, and cathodoluminescence information within the field of view to be detected.

[0042] Step (2): Construct a self-generated mineral identification network model:

[0043] 1) Constructing and partitioning the image dataset:

[0044] This embodiment focuses on mineral identification of authigenic quartz, authigenic albite, authigenic calcite, authigenic dolomite, authigenic chlorite, and euhedral pyrite. Based on the distinct microstructure, mineral composition, and cathodoluminescence intensity of each of these authigenic minerals, a labeling tool (Labelme) is used to delineate and label the effective regions of these six target authigenic minerals in the image. Boundaries are distinguished based on morphology and luminescence characteristics to ensure that the labeled outlines closely match the actual edges of the minerals. During labeling, the category name and pixel-level contour coordinates of each authigenic mineral in a single image are recorded, generating corresponding pixel-level color mask images, such as... Figure 3 As shown, a category mapping dictionary is constructed to establish the correspondence between the RGB color values ​​of the mask image and the category index, ensuring that the pixels of the mask image accurately correspond to the mineral category and the model training index. Each image group and its corresponding mask image are cropped into 512×512 pixel image patches. The 200 image patch dataset is randomly divided into training, validation, and test sets in a 7:2:1 ratio.

[0045] 2) Construct a U-Net neural network model based on residual structure:

[0046] The U-Net model structure based on residual structure used in this embodiment is as follows: Figure 4 As shown, the overall structure retains the U-Net encoder-decoder symmetrical architecture. The encoder module uses a residual network, replacing the conventional convolutional stacking structure of traditional U-Net. The model uniformly inputs a 5-channel image group of 512×512 pixels. In the initial stage of the encoder, the input image is first processed by a 7×7 convolution, then batch normalization and ReLU activation function, and finally 3×3 max pooling is used to achieve initial downsampling, reducing the output feature map space size to 1 / 4 of the original size. At the same time, the number of feature channels is uniformly adjusted to 64 to complete the basic feature extraction of different types of authigenic minerals.

[0047] After initial feature extraction, a four-level residual structure is used instead of conventional convolutional modules as the core part of the model for extracting deep semantic features. Each level of the residual structure consists of two residual modules (BasicBlock), progressively mining different intrinsic high-level semantic features. The structure of a single residual module is as follows: Figure 5As shown, the internal structure includes two 3×3 convolutional layers. After each convolutional operation, a batch normalization layer and a ReLU activation function are sequentially connected. For the second convolutional layer, its output is element-wise added to the features of the short-circuit connection branches, and then ReLU activation is applied to achieve residual fusion. The batch normalization layer can standardize the feature data, effectively alleviating the gradient vanishing problem; the ReLU activation function significantly improves the model's feature learning ability and training convergence efficiency by introducing non-linear mapping. The short-circuit connection branches within each residual module adopt a differentiated mapping strategy depending on whether downsampling is required. When downsampling is performed, its skip connections complete feature map size reduction and channel dimension expansion through a 1×1 convolutional layer. When there is no downsampling, its skip connections directly use identity mapping, passing the input features to the fusion node as is. Neither of the two residual modules in the first stage performs downsampling, and all their skip connections use identity mapping. As the network depth increases, the residual structure gradually halves the feature map size and doubles the number of channels, completing multi-level feature extraction and feature map scale compression for the encoder.

[0048] The decoder employs an upsampling structure that is completely symmetrical to the encoder's four-level downsampling, and restores the feature map scale through bilinear interpolation. After each bilinear interpolation operation, the deep semantic features obtained by the decoder's upsampling are concatenated with the shallow features of the same layer encoder passed through skip connections, slicing them by channel dimension. This ensures that both local detail information and deep semantic discriminative ability are taken into account, significantly improving the model's local fine-grained segmentation performance. The decoder terminal adjusts the channel dimension through a 1×1 convolutional layer while maintaining the feature map spatial size unchanged. The model ultimately outputs pixel-level segmentation feature maps for 6 authigenic minerals and other substances (background category), totaling 7 categories, enabling automated fine-grained segmentation of multiple authigenic mineral categories.

[0049] 3) Design the loss function and select evaluation metrics:

[0050] The cross-entropy loss function is used to constrain the class classification of each pixel, while the Dice loss function helps alleviate the bias problem caused by class imbalance. A composite loss function, combining the cross-entropy and Dice loss functions, is used as the loss function for model training, jointly supervising model training and guiding the direction of model parameter updates. The Dice coefficient and pixel accuracy are chosen as evaluation metrics for the network model, measuring model performance from two dimensions: region overlap and pixel-wise classification.

[0051] 4) Training a U-Net shale authigenic mineral model based on residual structure:

[0052] The backscattered-elemental-catholuminescence image set and its corresponding mask image are input into the constructed residual-based U-Net model for training. During model training, to expand sample diversity and mitigate the risk of overfitting on small datasets, the training set is randomly flipped horizontally, vertically, rotated, or its brightness is adjusted to improve the model's generalization ability and segmentation reliability for complex samples.

[0053] Step (3): Identify authigenic minerals in shale images using the U-Net model based on residual structure.

[0054] Acquire backscattered images, elemental distribution maps, and cathodoluminescence images of the area to be detected, such as... Figure 6 As shown. Image groups are generated and cropped into 512×512 image patches. The image patch dataset is input into a trained U-Net model based on residual structure to segment authigenic quartz, authigenic albite, authigenic calcite, authigenic dolomite, authigenic chlorite, and euhedral pyrite from the image patch dataset. The predicted results of the image patch dataset are then stitched together in situ to obtain the spatial distribution map of the above six authigenic minerals in the detected region. The output result is shown in the figure. Figure 7 As shown. From Figure 7 As can be seen from the above, the method used in this invention can effectively extract the crystal morphology and cathodoluminescence characteristics of different types of authigenic minerals at the micron level, segment the particle edges of fine-grained authigenic minerals, and distinguish authigenic minerals from clay matrix.

[0055] Based on the output images, the contents of different authigenic minerals in the shale of the example show certain differences. The contents of different types of authigenic minerals are as follows: Figure 8 As shown, authigenic quartz and authigenic albite are the most abundant minerals. The formation of authigenic minerals fills pores, blocks pore throats, and reduces permeability. Authigenic quartz, albite, and calcite are non-conductive and brittle minerals, while pyrite is highly conductive. The formation of these authigenic minerals leads to certain changes in the electrical and mechanical properties of the reservoir. Identifying these authigenic minerals allows for the interpretation and quantification of changes in the pore structure, electrical properties, and mechanical characteristics of reservoir rocks from a microscopic perspective, providing integrated geological-rock physics evidence for well logging, seismic analysis, and engineering evaluation, and enabling accurate prediction of reservoir parameters.

Claims

1. A U-Net method for identifying authigenic minerals in shale based on residual structures, characterized in that, Includes the following steps; Step (1): Acquire sample images: Rock samples were prepared into block-shaped samples measuring 1.0cm × 1.0cm × 2.5cm. The sample surfaces were argon-ion polished for observation under a scanning electron microscope. For each analytical field of view, backscattered images, elemental distribution images, and cathodoluminescence images were acquired at the same magnification, and these three types of images were combined to form an image set. Each image set includes information on the microstructure, material composition, and cathodoluminescence characteristics of the rock, providing a basis for the identification of authigenic minerals. Step (2): Construct a self-generated mineral identification network model: 1) Constructing and partitioning the image dataset; 2) Construct a U-Net neural network model based on residual structure; 3) Design the loss function and select evaluation indicators; 4) Train a U-Net shale authigenic mineral model based on residual structure; Based on the model built above, the image dataset formed by combining backscattered-elemental distribution-catholuminescence images and the corresponding label images are input into the U-Net model based on residual structure for training. The training process of the model is monitored by loss function and validation set evaluation index. Finally, the segmentation accuracy of the model is tested using test set data to see if it meets the requirements. Step (3): Identify authigenic minerals based on the U-Net model with residual structure. Backscattered images, elemental distribution maps, and cathodoluminescence images of the region to be detected are acquired, forming image groups and cropping them into image patches. These patches are then input into a U-Net model based on residual structure to predict the target authigenic minerals in the image patches, resulting in segmented images of the target authigenic minerals. The image patches are then stitched together to obtain the authigenic mineral picking results for each image group.

2. The method for identifying authigenic minerals in shale based on residual structure according to claim 1, characterized in that: The specific steps for constructing and partitioning the image dataset in step (2) are as follows: Based on the microstructure, material composition, and cathodoluminescence image information of each image group, different authigenic minerals are manually labeled with different category names to generate mask images. A mapping dictionary is constructed to realize the mapping relationship between the RGB color values ​​of the mask images and the category index. Each image group and its corresponding mask image are cropped into 512×512 pixel image blocks. The image block dataset is randomly divided into training set, validation set, and test set as the data basis for model training and evaluation.

3. The method for identifying authigenic minerals in shale based on residual structure according to claim 1, characterized in that: The specific steps in step (2) of building the U-Net neural network model based on the residual structure are as follows: The U-Net model based on residual structure uses the traditional U-Net network structure. The encoder initially consists of one 7×7 convolutional layer and one 3×3 max-pooling layer, downsampling the input image size to 1 / 2 and 1 / 4 of its original size, respectively. The convolutional layer adjusts the number of channels to 64. The encoder path then uses a 4-level residual structure instead of the conventional convolutional layers. Each residual structure consists of two residual modules (BasicBlock) to enhance the multi-scale feature extraction capability of authigenic minerals and the ability to model the overall structure of large targets using global context information. Each residual module contains two... The encoder consists of 3×3 convolutional layers, each followed by a batch normalization layer and a ReLU activation function. For the second convolutional layer, its output is element-wise added to the features of the short-circuit connection branches before applying ReLU activation to achieve residual fusion, preserving effective features from shallow layers while enhancing the expressive power of deep features. The encoder's 4-level residual structure outputs feature maps at each level, which are then input to the corresponding level of the decoder via skip connections. The decoder uses bilinear interpolation for upsampling and fuses encoded features through skip connections. In the final stage, the decoder adjusts the channel dimensions using 1×1 convolutions to achieve the segmentation of the target authigenic minerals.

4. The method for identifying authigenic minerals in shale based on residual structure according to claim 1, characterized in that: The design of the loss function and the selection of evaluation indicators in step (2) are as follows: During model training, the loss function and evaluation metrics jointly supervise the model training direction and drive parameter optimization. This step uses a composite loss function as the loss function to supervise model training, which is used to quantify the deviation between the model prediction results and the true labels. The closer the loss function value is to 0, the smaller the error between the model's prediction and the true label, and the better the fit. The composite loss function is a combination of the Cross Entropy Loss function and the Dice loss function, specifically: ; Where Loss is the total loss function, Loss CE It is the cross-entropy loss function, Loss Dice This is the Dice loss function. To measure the model's segmentation ability, the Dice score and pixel accuracy (PA) are chosen as evaluation metrics for the network model. The Dice score measures the overlap between the predicted region and the ground truth region; the closer the value is to 1, the higher the overlap between the predicted region and the ground truth region, and the better the pixel segmentation matching effect. Pixel accuracy is the proportion of correctly classified pixels out of the total pixels, reflecting the overall pixel-level accuracy. The closer the value is to 1, the more pixels the model correctly predicts, and the better the overall prediction effect. The specific definitions of the evaluation metrics are as follows: ; ; Wherein, TP is the number of pixels predicted as the current category and actually belonging to the current category, FP is the number of pixels predicted as the current category but actually belonging to all other categories, FN is the number of pixels predicted as all other categories but actually belonging to the current category, and TN is the number of pixels actually belonging to all other categories and also predicted as all other categories.

Images