Petri Analytic Net deep learning model-based geological image segmentation method and system in oil and gas reservoir exploration

Abstract

The invention provides a geological image segmentation method and system in oil and gas reservoir exploration based on a PetroAnalytic Net deep learning model, and belongs to the field of image processing. The problems that in the existing geological image segmentation process, geological data are complex, inherent noise exists, segmentation precision is poor, and calculation efficiency is low are solved. The method comprises separable expansion convolution, feature fusion and an attention mechanism, multi-scale features can be efficiently captured, and meanwhile the problems of complexity and noise in geological data are solved. Through wide experimental verification, the method is superior to an existing baseline method in segmentation accuracy; and the advanced feature extraction technology combination is adopted, so that the geological structure can be distinguished more accurately, and key features of oil-bearing strata and the like can be identified better. Ablation studies further emphasize the importance of each architecture component and show their collective contribution to the overall performance.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a geological image segmentation method and system in oil and gas reservoir exploration based on a PetroAnalyticNet deep learning model. Background Art

[0002] Geological image segmentation is one of the most critical components in accurately identifying and delineating underground geological structures. It locates potential oil and gas reservoirs by analyzing seismic and other geophysical imaging data. These images are usually generated by seismic surveys, which create detailed maps of underground geological structures by sending sound waves into the earth's subsurface and recording their reflections. Accurately segmenting these images helps geophysicists identify important features such as faults, folds, and stratigraphic boundaries, which often indicate the presence of oil and gas reservoirs.

[0003] The rise of machine learning, especially deep learning, has brought disruptive changes to the field of image analysis, including geological image segmentation. Convolutional neural networks (CNNs) have shown strong capabilities in pattern recognition and image classification tasks. CNNs can automatically learn and extract relevant features from raw image data, outperforming traditional image processing techniques in terms of accuracy and efficiency. This capability makes them naturally suitable for seismic and geological image segmentation tasks, which require accurate recognition of complex patterns and subtle features.

[0004] Among the many CNN architectures, U-Net is popular because of its unique structure. This structure can effectively capture low-level and high-level features. U-Net is used in many fields including geological imaging. Its accurate pixel-level segmentation ability makes it the preferred model for tasks that require detailed boundary detection, such as delineating fault lines and stratigraphic interfaces.

[0005] However, despite the success of CNNs and U-Nets in image segmentation tasks, their application to geological imaging faces unique challenges. Seismic data is often noisy, with multiple sources of interference that can obscure key geological features. In addition, geological structures are highly variable in scale, texture, and orientation, making it difficult for standard CNNs to maintain consistent performance across different regions or datasets. This variability, coupled with the inherent noise in seismic data, requires more specialized models to simultaneously handle these complexities and improve computational efficiency while maintaining high accuracy. Summary of the invention

[0006] The technical problems to be solved by the present invention are:

[0007] In order to solve the problems of complex geological data and inherent noise in the existing geological image segmentation process, resulting in poor segmentation accuracy and low computational efficiency.

[0008] The present invention adopts the following technical solutions to solve the above technical problems:

[0009] The present invention provides a geological image segmentation method in oil and gas reservoir exploration based on the PetroAnalytic Net deep learning model, comprising the following steps: the PetroAnalyticNet deep learning model comprises an encoder and a decoder,

[0010] The encoder includes a plurality of convolution blocks, each of which extracts different features of geological input data through a set of filters, starting from basic features and gradually extracting as the network deepens; after the convolution layer, the information is condensed into a deeper level of abstraction to prepare a feature map through a pooling operation; wherein the convolution block is a separable dilated convolution block, and the separable dilated convolution block includes dilated convolution, separable convolution and residual connection; the dilated convolution is used to expand the receptive field of the network; the separable convolution block decomposes the learning process into two layers, one for filtering features and the other for combining features; a series of residual blocks are stacked together, and each block performs an identity mapping to pass the features of the early layer to the later layer;

[0011] At the bottleneck, the network reaches its deepest layer, which serves as the transition point between the encoder and decoder;

[0012] The decoder reconstructs the segmentation map from the compressed feature representation. The upsampling layer in the decoder gradually increases the resolution of the feature map and integrates the features of the corresponding stages in the encoder through skip connections.

[0013] The final output of the segmentation map is generated by the last layer of the decoder; this output layer generates a binary image through a set of filters, using pixel-by-pixel classification to distinguish reservoir from non-reservoir areas, thereby determining the location of oil and gas reservoirs.

[0014] Furthermore, the separable convolution block decomposes the convolution operation into two stages, namely, depth convolution and point convolution, specifically including:

[0015] The first stage is a depthwise convolution, where a single convolution filter is applied to each input channel:

[0016] D k,l,m,n =I k,l,m *K l,n

[0017] Among them, D k,l,m,n is the depth convolution output of m rows and n columns at position k, I represents the input seismic image, and K represents the depth convolution kernel;

[0018] The second stage is point convolution, which uses a1*1 convolution to integrate the cross-channel information output by the depth convolution:

[0019]

[0020] Among them, W is a 1×1 convolution kernel and M is the number of channels;

[0021] The final output combines the results of depthwise convolution and pointwise convolution:

[0022] O k,l =P k,l +I k,l

[0023] Among them, O is the final output and I is the input.

[0024] Furthermore, the depthwise convolution and dilated convolution are combined to extract features through point-by-point convolution and attention layers, including, through the depthwise convolution input, where the spatial features of each channel are extracted independently:

[0025] D c (x, y) = I c (x, y)*k c (x · ,y · )

[0026] Among them, D c is the depthwise convolution output of channel c, I c is the input of channel c, K c represents the convolution kernel of channel c, (x · ,y · ) are the coordinates within the nucleus;

[0027] The input undergoes a dilated convolution, which increases the receptive field to capture larger spatial context:

[0028] D d (x, y) = I (x, y) * K * (x · +r·a,y · +r·a)

[0029] Among them, D d represents the dilated convolution output, K * is the dilated kernel, r is the dilation rate, and a is the spacing between kernel elements;

[0030] The outputs of depthwise convolution and dilated convolution are combined and passed through point-wise convolution to compress the channel information into a denser feature representation:

[0031]

[0032] Among them, P(x,y) is the point-by-point convolution output, C is the number of channels, and W c is the weight of the 1×1 convolution kernel corresponding to channel c;

[0033] The attention layer follows point-by-point convolution and the final output is a feature map:

[0034] A(x,y)=σ(P(x,y)·W a )

[0035] Among them, A(x,y) is the output of the attention layer, W a is the weight matrix of the attention layer, and σ is the sigmoid activation function, which scales the output between 0 and 1.

[0036] Furthermore, the residual block adds the input to the output of the network layer, specifically including:

[0037] The input is passed through a set of 3*3 convolution filters to extract features from the geological data. The mathematical expression of the convolution operation is:

[0038] F(x)=W d *x+b

[0039] Among them, F(x) is the output feature map, W d represents the weights of 3×3 filters, x is the input representing the surface image, and b is the bias term;

[0040] Then batch normalization is performed to standardize the feature map to have a mean of zero and a variance of one:

[0041]

[0042] in, is the normalized output, μ is the mean value, σ 2 is the variance, ε is a small constant for numerical stability;

[0043] Introducing the ReLU activation function into nonlinearity:

[0044] R(x)=max(0,x)

[0045] Where R(x) is the output after applying ReLU activation;

[0046] Reduce the dimension of the feature map through point convolution to prepare for residual connection:

[0047] P(x)=W p *R(x)

[0048] Among them, P(x) is the output after point convolution, W p is the weight of the 1×1 convolution;

[0049] Add the point convolution output to the original input:

[0050] O(x)=P(x)+x

[0051] Among them, O(x) is the final output of the residual block.

[0052] Furthermore, the basic feature is a rock edge.

[0053] The present invention provides a geological image segmentation system for oil and gas reservoir exploration based on the PetroAnalytic Net deep learning model. The system has a program module corresponding to the above steps, and executes the steps in the geological image segmentation method for oil and gas reservoir exploration based on the PetroAnalytic Net deep learning model during operation.

[0054] The present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program is configured to implement the steps of a geological image segmentation method in oil and gas reservoir exploration based on a Petro Analytic Net deep learning model when called by a processor.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] The present invention integrates separable dilated convolution, feature fusion strategy and attention mechanism, which enhances the model's ability to capture multi-scale geological features, while improving segmentation accuracy and computational efficiency. The ability to accurately locate and identify oil and gas reservoirs can significantly reduce the risks and costs associated with drilling operations.

[0057] Through comprehensive experimental verification, the model outperforms the baseline method in terms of precision, recall, F1 score and intersection over union (IoU). These improvements show that the present invention can effectively cope with the complexity and diversity of geological structures, such as noise, irregular structures, and the need for high-resolution feature extraction.

[0058] In addition, ablation studies further demonstrate the importance of each component in the model architecture. The results confirm that the combination of separable dilated convolutions, feature fusion strategies, and attention mechanisms significantly improves segmentation performance. Each component plays a key role in ensuring that the model can efficiently process large-scale geological data, which is critical for practical applications in the oil and gas industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a framework diagram of a geological image segmentation method in oil and gas reservoir exploration based on the PetroAnalyticNet deep learning model in an embodiment of the present invention;

[0060] Figure 2 is an architectural diagram of a DSA block in an embodiment of the present invention;

[0061] Figure 3 is an architecture diagram of a separable CONV block in an embodiment of the present invention;

[0062] Figure 4 is an architecture diagram of a convolution block in an embodiment of the present invention;

[0063] Figure 5 is a structural diagram of a residual block in an embodiment of the present invention;

[0064] Figure 6 It is a curve diagram of the accuracy of training and verification in various generations in the embodiment of the present invention;

[0065] Figure 7 Graph showing the loss of training and validation for each generation in the embodiments of the present invention. DETAILED DESCRIPTION

[0066] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0067] Specific implementation plan 1: Combine Figures 1 to 5 As shown, the present invention provides a geological image segmentation method in oil and gas reservoir exploration based on the PetroAnalyticNet deep learning model, comprising:

[0068] Combination Figure 1 As shown in the figure, the PetroAnalyticNet deep learning model is based on an encoder-decoder framework optimized for geological image segmentation tasks; this architecture is designed to capture complex subsurface features and convert them into accurate segmentation maps through a series of convolution operations;

[0069] The encoder includes multiple convolutional blocks for feature extraction; each convolutional block extracts different features of geological input data through a set of filters, starting with basic features (such as rock edges) and gradually extracting more complex patterns as the network deepens; after the convolutional layer, the spatial dimension is reduced through pooling operations, condensing the information and preparing feature maps for deeper abstraction;

[0070] At the bottleneck, the network reaches its deepest layer, where the most abstract forms of geological features are stored; this critical node serves as a transition point between the encoder and decoder, ensuring that high-level features are ready for the detailed reconstruction process;

[0071] The decoder corresponds to the encoder and reconstructs the segmentation map from the compressed feature representation. The upsampling layer in the decoder gradually increases the resolution of the feature map. At the same time, the features of the corresponding stage in the encoder are integrated through jump connections, which helps to preserve the spatial details required for segmentation.

[0072] The final output of the segmentation map is generated by the last layer of the decoder; this output layer generates a binary image through a set of filters, using pixel-by-pixel classification to distinguish reservoir from non-reservoir areas, thereby determining the location of oil and gas reservoirs;

[0073] Among them, the Separable Dilated Convolutions Block (SDC block) is a key component of PetroAnalyticNet, which is designed to improve the feature representation and extraction capabilities of geological images. This composite block combines dilated convolution, separable convolution and residual connection, and the generated feature-rich output is used for subsequent processing. By using separable convolution, the network maintains high efficiency while reducing the computational burden without sacrificing feature details. Dilated convolution is used to capture a wider context, enabling the network to effectively segment geological structures with different sizes and appearances. The embedded attention mechanism further optimizes the focus on key geological features, enabling the Petro Analytic Net deep learning model to prioritize the areas of interest in underground images, improve the accuracy of segmentation output, and enhance its ability to distinguish in complex geological backgrounds.

[0074] The first component in the SDC block is the dilated convolution block; it uses dilated convolutions to enlarge the network’s receptive field while not increasing the number of parameters, thereby capturing greater contextual information from the input data; the dilation factor is carefully chosen to ensure that the network can perceive a wide range of geological structures while maintaining resolution; this block is critical for identifying features distributed over a large area in the subsurface, especially for detecting reservoirs with irregular shapes or sizes;

[0075] Then comes the separable convolution block; this block uses depth-wise separable convolution to decompose the learning process into two layers: one for filtering and the other for combining; this not only significantly reduces the computational cost, but also preserves the depth of feature processing; the separable convolution block excels at optimizing the features extracted by dilated convolution, providing a more detailed understanding of local features in geological data;

[0076] A series of residual blocks are stacked together, each of which performs identity mapping to quickly pass features from early layers to later layers; this setting helps alleviate the gradient vanishing problem and ensures that subtle geological features deep in the network are preserved; residual blocks improve the network's learning ability by learning residuals (differences) rather than direct features, which is particularly effective for low feature contrast in geological imaging;

[0077] The outputs of the dilated convolution and separable convolution blocks are merged with the output stream of the residual block. These parallel streams are integrated through element-wise addition, integrating broad contextual features, detailed local features, and residuals. The merged output is further passed on, carrying a rich feature set and the original characteristics of the input, ensuring that no key information is lost.

[0078] For the separable convolution block (Separable ConvolutionBlock), it is used to efficiently process seismic data.

[0079] Combination Figure 2 As shown in the figure, by decomposing the convolution operation into two independent stages: depth convolution and point convolution, the computational complexity is significantly reduced while retaining the necessary feature extraction capabilities, which are crucial for geological analysis, including:

[0080] The first stage is a depthwise convolution, where a single convolution filter is applied independently to each input channel; the mathematical representation of this operation is:

[0081] D k,l,m,n =I k,l,m *K l,n

[0082] Among them, D k,l,m,n is the output of the deep convolution with m rows and n columns at position k, I represents the input seismic image, and K represents the deep convolution kernel; each kernel K is specific to its corresponding input channel, so that features can be extracted from different geological layers in a concentrated manner;

[0083] The second stage is point convolution, which uses a1*1 convolution to integrate the cross-channel information output by the depth convolution; its mathematical expression is:

[0084]

[0085] W is a 1×1 convolution kernel, and M is the number of channels. This operation combines channel-level features into a new feature space, which enhances the detection capability of complex geological structures.

[0086] The final output combines the results of depthwise convolution and pointwise convolution, providing an efficient and powerful mechanism for analyzing geological features. Its output is expressed as:

[0087] O k,l =P k,l +I k,l

[0088] Among them, O is the final output and I is the input, which propagates the initial features together with the newly extracted features, supporting a comprehensive analysis of the subsurface structure;

[0089] Combination Figure 4As shown in Figure 1, an advanced convolution block is provided, which is designed for efficient feature extraction in geological image segmentation. It combines depthwise convolution and dilated convolution, followed by pointwise convolution and attention layers, specifically including:

[0090] The input is passed through a deep convolution, where the spatial features of each channel are extracted independently:

[0091] D c (x, y) = I c (x, y)*K c (x · ,y · )

[0092] Among them, D c is the depthwise convolution output of channel c, I c is the input of channel c, K c represents the convolution kernel of channel c; (x · ,y · ) are the coordinates within the nucleus;

[0093] At the same time, the input undergoes a dilated convolution, which enhances the receptive field to capture larger spatial context without increasing the number of parameters:

[0094] D d (x, y) = I (x, y) * K * (x · +r·a,y · +r·a)

[0095] Among them, D d represents the dilated convolution output, K * is the dilated kernel, r is the dilation rate, and a is the spacing between kernel elements, allowing the network to "skip" input pixels and cover more input area;

[0096] The outputs of the depthwise convolution and the dilated convolution are then combined and passed through a point-wise convolution to compress the channel information into a denser feature representation:

[0097]

[0098] Among them, P(x,y) is the point-by-point convolution output, C is the number of channels, and W c is the weight of the 1×1 convolution kernel corresponding to channel c;

[0099] The attention layer follows the point-wise convolution to focus the model’s power on the most salient features that contribute to accurate geological segmentation:

[0100] A(x,y)=σ(P(x,y)·W a )

[0101] Among them, A(x,y) is the output of the attention layer, W a is the weight matrix of the attention layer, and σ is a sigmoid activation function that scales the output between 0 and 1, effectively allowing the network to "pay attention" to some features more than others;

[0102] The final output of this block is a feature map that is refined through deep and dilated convolutions for spatial context, point-wise convolutions for feature integration, and an attention mechanism for feature enhancement. This multifaceted approach ensures that the resulting feature map captures the complexity of geological structures in detail, facilitating their accurate segmentation from the surrounding rock matrix. This block is particularly well suited for geological segmentation because it combines detailed local feature extraction with broader contextual understanding, which is critical for identifying subsurface features such as hydrocarbon reservoirs. The added attention mechanism ensures that the model prioritizes the most important features for segmentation, thereby improving accuracy and specificity of the task.

[0103] For the residual block, combined Figure 5 As shown in Figure 1, the residual block is crucial for deep learning models in geological image segmentation. By adding the input (or shortcut connection) to the output of the network layer, the residual block alleviates the gradient vanishing problem, including:

[0104] The input first passes through a set of 3*3 convolution filters to extract detailed features in the geological data. The mathematical expression of the convolution operation is:

[0105] F(x)=W d *x+b

[0106] Among them, F(x) is the output feature map, W d represents the weights of 3×3 filters, x is the input representing the surface image, and b is the bias term;

[0107] Next, we perform batch normalization to standardize the feature map to have a mean of zero and a variance of one:

[0108]

[0109] in, is the normalized output, μ is the mean value, σ 2 is the variance, ∈ is a small constant for numerical stability;

[0110] Then, the ReLU activation function introduces nonlinearity, enhancing the model’s ability to capture complex geological patterns:

[0111] R(x)=max(0,x)

[0112] Among them, R(x) is the output after applying ReLU activation, which enhances the ability to model nonlinear geological features;

[0113] Finally, point convolution reduces the dimensionality of the feature map in preparation for the residual connection:

[0114] P(x)=W p *R(x)

[0115] Among them, P(x) is the output after point convolution, W p is the weight of the 1×1 convolution, and R(x) is the input after ReLU activation;

[0116] Add the point convolution output to the original input via a shortcut connection:

[0117] O(x)=P(x)+x

[0118] Where O(x) is the final output of the residual block and x is the original input. This step is crucial to maintain the gradient flow throughout the network, which is critical in complex geological environments.

[0119] This residual block can effectively retain key features in complex geological environments and provide support for accurate segmentation of geological structures;

[0120] In the context of geological segmentation, this residual block helps preserve key subsurface features; it is necessary for accurately segmenting geological structures; the integration of batch normalization and ReLU activation not only refines the feature map but also enhances the network's ability to effectively distinguish different geological layers and strata.

[0121] Specific implementation scheme 2: The present invention provides a geological image segmentation system for oil and gas reservoir exploration based on the Petro Analytic Net deep learning model. The system has program modules corresponding to the above steps, and executes the steps in the above-mentioned geological image segmentation method for oil and gas reservoir exploration based on the Petro Analytic Net deep learning model during operation.

[0122] The other combinations and connection relationships of this embodiment are the same as those of the first embodiment.

[0123] Specific implementation scheme three: The present invention provides a computer-readable storage medium, which stores a computer program, and the computer program is configured to implement the steps of a geological image segmentation method in oil and gas reservoir exploration based on the Petro Analytic Net deep learning model when called by a processor.

[0124] The other combinations and connection relationships of this embodiment are the same as those of the first embodiment.

[0125] Simulation experiment

[0126] A. Dataset Description

[0127] This experiment used a dataset designed for oil spill detection in port environments using RGB images captured by drones. The dataset contains 1,268 high-resolution images, collected over a two-year period (September 2021 to September 2023), covering a variety of conditions and activities in ports. The images contain a variety of scenes under different lighting and weather conditions, ensuring a comprehensive representation of the potential challenges of oil spill detection. To maintain privacy and confidentiality, sensitive information contained in the images (such as recognizable text, logos, and faces) was anonymized using image restoration and blurring techniques. Each image was accurately annotated to distinguish between areas of oil, water, and other elements (such as ships and docks). The annotation process used the CVAT tool and was reviewed by multiple experts to ensure the accuracy and reliability of the annotations. The final dataset has been published on the Zenodo platform and is divided into training, validation, and test sets in a 70 / 15 / 15 ratio, supporting a variety of machine learning applications and providing a key resource for the development of robust detection models that can significantly improve response strategies and reduce the impact of port environments.

[0128] Table 1 provides a detailed statistical description of the dataset used by three main categories (oil, waters, and others). The table is divided into three rows to present the statistical information of the dataset in a concise way:

[0129] The first row lists the total number of images containing each category, showing the composition and diversity of the dataset.

[0130] The second row shows the total number of annotated pixels for each category, reflecting the scope of detailed annotation and the focus of image segmentation work.

[0131] The third row shows the relative annotation ratio (percentage) of each category, providing the proportion distribution of each category in the dataset.

[0132] Table 1: Overview of datasets by category

[0133]

[0134] B. Experimental Setup

[0135] The experimental setup uses a modified U-Net architecture with an EfficientNet-B4 encoder optimized for high-resolution drone RGB image segmentation tasks. The model is trained using the Adam optimizer with an initial learning rate of 0.001 and a combination of Binary Cross-Entropy and Dice Loss as custom loss functions to effectively balance the precision and recall of the model. The training process lasts for 50 cycles with a batch size of 16 and incorporates a large number of data augmentation techniques to improve the robustness of the model under different environmental conditions.

[0136] The experiment was run in a high-performance computing environment, specifically configured with NVIDIA Tesla V100 GPU and Intel Xeon E5-2698v4 CPU, equipped with 256GB memory and 2TB SSD storage to ensure efficient processing power. The experiment was based on TensorFlow 2.4 and Keras frameworks and ran on Ubuntu 20.04 LTS system. Data processing relied on OpenCV, PIL, NumPy and Pandas, building a powerful and scalable environment for performing detailed segmentation tasks in the port environment.

[0137] The performance of the PetroAnalyticNet model is evaluated through several key metrics that provide insight into the accuracy and efficiency of the model in geological image segmentation tasks. The following are the definitions and mathematical formulas for each metric:

[0138] F1 score:

[0139] The F1 score is the harmonic mean of precision and recall, which is used to comprehensively measure the accuracy of the model. It is defined as:

[0140]

[0141] Among them, Precision represents the proportion of correctly predicted positive examples to all predicted positive examples, and Recall represents the proportion of correctly predicted positive examples to all actual positive examples.

[0142] Accuracy:

[0143] Precision measures the accuracy of positive example predictions and is calculated as:

[0144]

[0145] Among them, TP is the number of correctly predicted positive examples, and FP is the number of incorrectly predicted positive examples.

[0146] Recall:

[0147] Recall (or sensitivity) measures the model's ability to identify all relevant instances and is given by:

[0148]

[0149] Where FN is the number of misclassified negative examples.

[0150] Well-to-well ratio:

[0151] IoU (also known as the Jaccard index) measures the degree of overlap between the predicted segmentation and the true segmentation. Its calculation formula is:

[0152]

[0153] Among them, A represents the predicted segmentation area, and B represents the real segmentation area.

[0154] in conclusion

[0155] The performance of the OURS model on each category was rigorously evaluated using standard segmentation metrics and compared with baseline models. Table 2 shows the optimized results, detailing the model’s enhanced capabilities on different categories.

[0156] Table 2: Performance comparison of BASELINE and OURS by category

[0157]

[0158] As shown in Table 2, the OURS model significantly outperforms the baseline model in all performance metrics. Although the improvements are relatively modest, these improvements reflect practical and achievable progress. The following are the main improvements:

[0159] Oil category: The F1 score increased by 0.04, thanks to the balanced improvement in precision and recall, indicating that the model is more accurate in identifying and segmenting oil features.

[0160] Water category: The precision rate increased by 0.03, and the recall rate reached full marks, indicating that the model’s ability to distinguish water areas in complex geological environments has reached an excellent level.

[0161] Other categories: The “Other” category, which includes diverse geological features, shows consistent improvements in all metrics, with both F1 score and IoU increasing by 0.04, indicating that the model’s ability to handle diverse features has improved.

[0162] The significant advantage of the OURS model is that it can improve precision without sacrificing recall, thereby improving F1 score and IoU. These improvements are attributed to the innovative architectural features of OURS, including advanced dilated convolutions and powerful feature fusion strategies. Experimental results show that OURS is an efficient tool for geological image segmentation tasks in the field of oil and gas exploration, providing more reliable and accurate analysis for the decision-making process.

[0163] The training and validation accuracy curves are as follows: Figure 6 The figure shows the performance change of the model during 50 training cycles. In the initial stage, both the training and validation accuracy are low, reflecting the initial learning stage of the model. As the training progresses, the accuracy of both the training and validation sets increases steadily, and at the end of the training, the training accuracy reaches a high value. The validation accuracy shows a similar trend, slightly lower than the training accuracy, indicating that the model has good generalization ability for unseen data.

[0164] The loss curves for training and validation are as follows Figure 7 Figure 1 shows the change in loss of the model over 50 training cycles. The training loss continues to decrease, reflecting the gradual decrease in the error of the model on the training data. Similarly, the validation loss also shows a downward trend, but it slows down slightly in the later stages. This shows that the model is learning well, but the slowdown in the rate of decrease in validation loss may indicate that further fine-tuning or regularization may help prevent overfitting.

[0165] From the accuracy and loss curves, we can see that there is no obvious separation between the training and validation curves, which means that the model is not overfitting and has good generalization ability on the validation set. The steady decline in loss and the continuous improvement in accuracy indicate that the current learning rate is suitable for the training process.

[0166] Ablation studies

[0167] To determine the contribution of specific components in the OURS model to the overall performance, an ablation study was performed. The study systematically removed key features from the full model configuration and observed the impact on various segmentation metrics. The results in Table 3 highlight the key role of each component:

[0168] Separable convolutions contribute significantly to the efficiency and accuracy of the model. When this component is removed, all performance metrics degrade.

[0169] Dilated convolutions enhance the context-awareness of the model and are essential for capturing a wider range of geological features. After removing this component, all metrics drop significantly.

[0170] Table 3: Ablation study results

[0171]

[0172]

[0173] The feature fusion strategy improves the model’s ability to integrate multi-scale information and improves the overall accuracy of the model. When this component is removed, the performance decreases, indicating its importance in multi-scale feature processing.

[0174] The attention mechanism helps the model focus on relevant features, and its removal directly leads to a decrease in precision and recall, indicating the key role of this mechanism in geological image segmentation tasks.

[0175] These findings highlight the importance of each component in the complete OURS model, confirming that their synergy is crucial to achieving high accuracy and efficiency in geological image segmentation tasks. This ablation study provides clear evidence that the complex architecture of the OURS model is able to effectively cope with the complex demands of geological segmentation.

[0176] Although the present invention is disclosed as above, the protection scope of the present invention is not limited thereto. Those skilled in the art may make various changes and modifications without departing from the spirit and scope of the present invention, and these changes and modifications will fall within the protection scope of the present invention.

Claims

1. A geological image segmentation method for oil and gas reservoir exploration based on the Petro Analytic Net deep learning model, characterized in that: The following steps are involved: The PetroAnalyticNet deep learning model includes an encoder and a decoder. The encoder includes a plurality of convolution blocks, each of which extracts different features of geological input data through a set of filters, starting from basic features and gradually extracting as the network deepens; after the convolution layer, the information is condensed into a deeper level of abstraction to prepare a feature map through a pooling operation; wherein the convolution block is a separable dilated convolution block, and the separable dilated convolution block includes dilated convolution, separable convolution and residual connection; the dilated convolution is used to expand the receptive field of the network; the separable convolution block decomposes the learning process into two layers, one for filtering features and the other for combining features; a series of residual blocks are stacked together, and each block performs an identity mapping to pass the features of the early layer to the later layer; At the bottleneck, the network reaches its deepest layer, which serves as the transition point between the encoder and decoder; The decoder reconstructs the segmentation map from the compressed feature representation. The upsampling layer in the decoder gradually increases the resolution of the feature map and integrates the features of the corresponding stages in the encoder through skip connections. The final output of the segmentation map is generated by the last layer of the decoder; this output layer generates a binary image through a set of filters, using pixel-by-pixel classification to distinguish reservoir from non-reservoir areas, thereby determining the location of oil and gas reservoirs.

2. The method for geological image segmentation in oil and gas reservoir exploration based on the Petro Analytic Net deep learning model according to claim 1, characterized in that: The separable convolution block decomposes the convolution operation into two stages, namely depth convolution and point convolution, specifically including: The first stage is a depthwise convolution, where a single convolution filter is applied to each input channel: D k,l,m,n =I k,l,m *K l,n Among them, D k,l,m,n is the depth convolution output of m rows and n columns at position k, I represents the input seismic image, and K represents the depth convolution kernel; The second stage is point convolution, which uses a1*1 convolution to integrate the cross-channel information output by the depth convolution: Among them, W is a 1×1 convolution kernel and M is the number of channels; The final output combines the results of depthwise convolution and pointwise convolution: O k,l =P k,l +I k,l Among them, O is the final output and I is the input.

3. The method for geological image segmentation in oil and gas reservoir exploration based on the Petro Analytic Net deep learning model according to claim 2, characterized in that: The depth convolution is combined with the dilated convolution to extract features through point-by-point convolution and attention layer. include, The input is passed through a deep convolution, where the spatial features of each channel are extracted independently: D c (x,y)=I c (x,y)*K c (x . ,y . ) Among them, D c is the depthwise convolution output of channel c, I c is the input of channel c, K c represents the convolution kernel of channel c, (x . ,y . ) are the coordinates within the nucleus; The input undergoes a dilated convolution, which increases the receptive field to capture larger spatial context: D d (x,y)=I(x,y)*K * (x . +r·a,y . +r·a) Among them, D d represents the dilated convolution output, K * is the dilated kernel, r is the dilation rate, and a is the spacing between kernel elements; The outputs of depthwise convolution and dilated convolution are combined and passed through point-wise convolution to compress the channel information into a denser feature representation: Among them, P(x, y) is the point-by-point convolution output, C is the number of channels, and W c is the weight of the 1×1 convolution kernel corresponding to channel c; The attention layer follows point-by-point convolution and the final output is a feature map: A(x,y)=σ(P(x,y)·W a ) Among them, A(x, y) is the output of the attention layer, W a is the weight matrix of the attention layer, and σ is the sigmoid activation function, which scales the output between 0 and 1.

4. The method for geological image segmentation in oil and gas reservoir exploration based on the Petro Analytic Net deep learning model according to claim 3, characterized in that: The residual block adds the input to the output of the network layer, specifically including: The input is passed through a set of 3*3 convolution filters to extract features from the geological data. The mathematical expression of the convolution operation is: F(x)=W d *x+b Among them, F(x) is the output feature map, W d represents the weights of 3×3 filters, x is the input representing the surface image, and b is the bias term; Then batch normalization is performed to standardize the feature map to have a mean of zero and a variance of one: in, is the normalized output, μ is the mean value, σ 2 is the variance, ∈ is a small constant for numerical stability; Introducing the ReLU activation function into nonlinearity: R(x)=max(0,x) Where R(x) is the output after applying ReLU activation; Reduce the dimension of the feature map through point convolution to prepare for residual connection: P(x)=W p *R(x) Among them, P(x) is the output after point convolution, W p is the weight of the 1×1 convolution; Add the point convolution output to the original input: O(x)=P(x)+x Among them, O(x) is the final output of the residual block.

5. The method for geological image segmentation in oil and gas reservoir exploration based on the Petro Analytic Net deep learning model according to claim 4, characterized in that: The base feature is a rock edge.

6. A geological image segmentation system for oil and gas reservoir exploration based on the Petro Analytic Net deep learning model, characterized by: The system has a program module corresponding to the steps of any one of claims 1 to 5 above, and executes the steps of the geological image segmentation method in oil and gas reservoir exploration based on the Petro Analytic Net deep learning model when running.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which is configured to implement the steps of the geological image segmentation method in oil and gas reservoir exploration based on the PetroAnalytic Net deep learning model described in any one of claims 1 to 5 when called by a processor.

Images