A geological image segmentation method and system for oil and gas reservoir exploration based on the Petro Analytic Net deep learning model.

Abstract

This invention provides a geological image segmentation method and system for oil and gas reservoir exploration based on the PetroAnalyticNet deep learning model, belonging to the field of image processing. It addresses the problems of poor segmentation accuracy and low computational efficiency caused by the complexity of geological data and inherent noise in existing geological image segmentation processes. This invention includes separable dilated convolution, feature fusion, and attention mechanisms, which can efficiently capture multi-scale features while addressing the complexity and noise issues in geological data. Extensive experimental verification shows that this invention outperforms existing baseline methods in segmentation accuracy; furthermore, it employs a combination of advanced feature extraction techniques to more accurately distinguish geological structures, thereby better identifying key features such as oil-bearing strata. Ablation studies further emphasize the importance of each architectural component, demonstrating their collective contribution to overall performance.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a geological image segmentation method and system for oil and gas reservoir exploration based on the PetroAnalyticNet deep learning model. Background Technology

[0002] Geological image segmentation is one of the most critical components in accurately identifying and delineating subsurface geological structures. It locates potential oil and gas reservoirs by analyzing seismic and other geophysical imaging data. These images are typically generated by seismic exploration, which creates detailed maps of subsurface geological structures by sending sound waves into the Earth's subsurface and recording their reflections. Accurate segmentation of 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 about a disruptive revolution in the field of image analysis, including geological image segmentation. Convolutional Neural Networks (CNNs) have demonstrated particularly powerful 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 both accuracy and efficiency. This capability makes them naturally suitable for seismic and geological image segmentation tasks, which require the accurate identification of complex patterns and subtle features.

[0004] Among numerous CNN architectures, U-Net is widely popular due to its unique structure. This structure effectively captures both low-level and high-level features. U-Net has been applied in multiple fields, including geological imaging. Its precise pixel-level segmentation capabilities make it the preferred model for tasks requiring 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 in geological imaging still faces unique challenges. Seismic data is often noisy, and multiple sources of interference can obscure key geological features. Furthermore, 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, demands more specialized models to handle these complexities simultaneously and improve computational efficiency while maintaining high accuracy. Summary of the Invention

[0006] The technical problem to be solved by this invention is:

[0007] To address the problems of complex geological data and inherent noise in existing geological image segmentation processes, resulting in poor segmentation accuracy and low computational efficiency.

[0008] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0009] This invention provides a geological image segmentation method for oil and gas reservoir exploration based on the PetroAnalyticNet deep learning model, comprising the following steps: the PetroAnalyticNet deep learning model includes an encoder and a decoder.

[0010] The encoder comprises multiple convolutional blocks, each extracting different features from the geological input data through a set of filters, starting with basic features and progressively extracting them as the network deepens. After the convolutional layers, pooling operations condense the information into deeper abstractions to prepare feature maps. The convolutional blocks are separable dilated convolutional blocks, comprising dilated convolutions, separable convolutions, and residual connections. The dilated convolutions expand the network's receptive field. The separable convolutional blocks decompose the learning process into two layers: one for filtering features and the other for combining features. A series of residual blocks are stacked together, each performing an identity mapping to pass features from earlier layers to later layers.

[0011] At the bottleneck, the network reaches its deepest layer, serving as a 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 from 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, and uses pixel-by-pixel classification to distinguish reservoir and non-reservoir areas, thereby determining the location of oil and gas reservoirs.

[0014] Furthermore, the separable convolutional block decomposes the convolution operation into two stages: depthwise convolution and pointwise convolution, specifically including:

[0015] The first stage is depthwise convolution, where a single convolutional 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 It is the depth convolution output at position k, with m rows and n columns. 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 depthwise convolution:

[0019]

[0020] Where 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] Where O is the final output and I is the input.

[0024] Furthermore, the combination of depthwise convolution and dilated convolution extracts features through pointwise convolution and attention layers, including through 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 It is the depthwise convolution output of channel c, I c It is the input of channel c, K c Denotes the convolution kernel for channel c, (x · ,y · () are the coordinates within the kernel;

[0027] The input undergoes dilated convolution to enhance the receptive field and capture a larger spatial context:

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

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

[0030] The outputs of depthwise convolution and dilated convolution are combined and then compressed into a denser feature representation through pointwise convolution:

[0031]

[0032] Where P(x,y) is the pointwise convolution output, C is the number of channels, and W is the number of channels. c These are the weights of the 1×1 convolution kernel corresponding to channel c;

[0033] The attention layer follows pointwise convolution, and the final output is a feature map:

[0034] A(x,y)=σ(P(x,y)·W a )

[0035] Where A(x,y) is the output of the attention layer, and W a σ 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 processed through a set of 3x3 convolutional filters to extract features from the geological data. The mathematical expression for the convolution operation is:

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

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

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

[0041]

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

[0043] Introducing nonlinearity into the ReLU activation function:

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

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

[0046] Reduce the dimensionality of the feature map using point convolutions to prepare for residual connections:

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

[0048] Where P(x) is the output after point convolution, and W p These are the weights of a 1×1 convolution;

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

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

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

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

[0053] This invention discloses a geological image segmentation system for oil and gas reservoir exploration based on the PetroAnalytic Net deep learning model. The system has program modules corresponding to the above steps, and executes the steps in the above-described geological image segmentation method for oil and gas reservoir exploration based on the PetroAnalytic Net deep learning model when running.

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

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] This invention integrates separable dilatational 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 validation, the proposed model outperforms baseline methods in terms of precision, recall, F1 score, and intersection-over-union (IoU) ratio. These improvements demonstrate that the present invention can effectively address the complexity and diversity of geological structures, such as noise, irregular structures, and the need for high-resolution feature extraction.

[0058] Furthermore, the ablation study further demonstrates the importance of each component in the model architecture. The results confirm that the combination of separable dilated convolution, feature fusion strategies, and attention mechanisms significantly improves segmentation performance. Each component plays a crucial role in ensuring the model can efficiently process large-scale geological data, which is essential for practical applications in the oil and gas industry. Attached Figure Description

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

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

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

[0062] Figure 4 This is an architectural diagram of the convolutional block in an embodiment of the present invention;

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

[0064] Figure 6 This is a graph showing the accuracy curves of training and validation in various generations in the embodiments of the present invention.

[0065] Figure 7 This is a loss curve diagram of training and validation in each generation of embodiments of the present invention. Detailed Implementation

[0066] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0067] Specific Implementation Plan 1: Combining Figures 1 to 5 As shown, this invention provides a geological image segmentation method for oil and gas reservoir exploration based on the PetroAnalyticNet deep learning model, comprising:

[0068] Combination Figure 1 As shown, the PetroAnalyticNet deep learning model is based on an encoder-decoder framework and is optimized for geological image segmentation tasks; the architecture aims to capture complex subsurface features and transform these features into accurate segmentation maps through a series of convolutional operations.

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

[0070] At the bottleneck, the network reaches its deepest layer, where the most abstract form of the geological features is 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 layers in the decoder progressively increase the resolution of the feature map; at the same time, the features from the corresponding stages in the encoder are integrated through skip 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 and uses pixel-by-pixel classification to distinguish reservoir and non-reservoir areas, thereby determining the location of oil and gas reservoirs.

[0073] The Separable Dilated Convolutions Block (SDC block) is a key component of PetroAnalyticNet, designed to enhance the feature representation and extraction capabilities of geological images. This composite block combines dilated convolutions, separable convolutions, and residual connections, generating feature-rich outputs for subsequent processing. By using separable convolutions, the network maintains high efficiency while reducing computational burden, without sacrificing feature details. Dilated convolutions capture broader 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, allowing the PetroAnalyticNet deep learning model to prioritize regions of interest in subsurface images, improving the accuracy of segmentation outputs and enhancing its discriminative ability in complex geological backgrounds.

[0074] The first component in the SDC block is the dilated convolution block; it uses dilated convolution to expand the network's receptive field without increasing the number of parameters, thereby capturing greater contextual information from the input data; the dilation factor is carefully selected to ensure that the network can perceive a wide range of geological structures while maintaining resolution; this block is crucial for identifying features distributed over a large subsurface area, especially for detecting reservoirs with irregular shapes or sizes.

[0075] Next is the separable convolutional block; this block uses depthwise separable convolution to break down the learning process into two layers: one for filtering and the other for combining; this not only significantly reduces computational costs but also preserves the depth of feature processing; the separable convolutional block excels at optimizing 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 capable of performing an identity mapping to quickly transfer features from earlier layers to later layers; this setup helps mitigate the vanishing gradient problem and ensures that subtle geological features deep within the network are preserved; the residual blocks enhance the network’s learning ability by learning residuals (differences) rather than direct features, which is particularly effective for situations with low feature contrast in geological imaging.

[0077] The outputs of dilated convolutions and separable convolutional blocks are merged with the output streams of residual blocks. The integration of these parallel streams is achieved through element-wise addition, which combines broad contextual features, detailed local features, and residuals. The merged output is further propagated, carrying a rich feature set and the original characteristics of the input, ensuring that no key information is lost.

[0078] Separable convolution blocks are used for efficient processing of seismic data.

[0079] Combination Figure 2 As shown, by decomposing the convolution operation into two independent stages: depthwise convolution and pointwise convolution, it significantly reduces computational complexity while retaining essential feature extraction capabilities, which are crucial for geological analysis. Specifically, this includes:

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

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

[0082] Among them, D k,l,m,n It is the m-row, n-column depth convolution output at position k, where I represents the input seismic image and K represents the depth convolution kernel; each kernel K is specific to its corresponding input channel, thus enabling the concentrated extraction of features from different geological layers;

[0083] The second stage is point convolution, which uses a1*1 convolution to integrate the cross-channel information output by the depthwise 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, enhancing 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] Where O is the final output and I is the input, propagating the initial features along with the newly extracted features to support comprehensive analysis of subsurface structures;

[0089] Combination Figure 4As shown, a high-level convolutional block is provided, 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 a depthwise 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 It is the depthwise convolution output of channel c, I c It is the input of channel c, K c Represents the convolution kernel for channel c; (x · ,y · () are the coordinates within the kernel;

[0093] Meanwhile, the input undergoes dilated convolutions to enhance the receptive field and capture a 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 K represents the dilated convolution output. * It is an expanded kernel, r is the expansion rate, and a is the spacing between kernel elements, which allows the network to "skip" input pixels and cover more input areas;

[0096] Then, the outputs of depthwise convolution and dilated convolution are combined and compressed into a denser feature representation through pointwise convolution:

[0097]

[0098] Where P(x,y) is the pointwise convolution output, C is the number of channels, and W is the number of channels. c These are the weights of the 1×1 convolution kernel corresponding to channel c;

[0099] The attention layer follows pointwise convolution, focusing the model's capabilities on the most prominent features that contribute to accurate quality segmentation:

[0100] A(x,y)=σ(P(x,y)·W a )

[0101] Where A(x,y) is the output of the attention layer, and W a σ 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 more attention" to certain features than others.

[0102] The final output of this block is a feature map that is enhanced through deep and dilated convolutions for spatial context, pointwise convolutions for feature integration, and an attention mechanism for feature enhancement. This multifaceted approach ensures that the generated feature maps capture the complexity of geological structures in detail, facilitating accurate segmentation of them within 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 crucial 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 the accuracy and specificity of the task.

[0103] For residual blocks, combined with Figure 5 As shown, residual blocks are crucial for deep learning models in geological image segmentation. By adding the input (or shortcut connections) to the output of network layers, residual blocks alleviate the gradient vanishing problem, specifically including:

[0104] The input is first processed through a set of 3x3 convolutional filters to extract detailed features from the geological data. The mathematical expression for the convolution operation is:

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

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

[0107] Next, batch normalization is performed to standardize the feature maps to a mean of zero and a variance of one:

[0108]

[0109] in, This is the normalized output, μ is the average value, and σ is the normalized output. 2 is the variance, and ∈ 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] R(x) is the output after applying ReLU activation, which enhances the ability to model nonlinear geological features;

[0113] Ultimately, point convolutions reduce the dimensionality of the feature maps, preparing them for residual connections:

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

[0115] Where P(x) is the output after point convolution, and W p These are the weights of the 1×1 convolution, and R(x) is the input after ReLU activation;

[0116] The point convolution output is added 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 for maintaining gradient flow throughout the network, which is essential in complex geological environments.

[0119] This residual block can effectively preserve key features in complex geological environments, providing support for accurate segmentation of geological structures;

[0120] In the context of geological segmentation, this residual block helps preserve key subsurface features; it is essential for accurate segmentation of 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 when running.

[0122] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0123] Specific Implementation Scheme 3: The present invention provides a computer-readable storage medium storing a computer program configured to implement, when called by a processor, the steps of a geological image segmentation method for oil and gas reservoir exploration based on the Petro Analytic Net deep learning model.

[0124] The other combinations and connections in this implementation scheme are the same as in Specific Implementation Scheme 1.

[0125] Simulation Experiment

[0126] A. Dataset Description

[0127] This experiment utilizes a dataset designed for oil spill detection in port environments, employing RGB images captured by drones. The dataset contains 1268 high-resolution images collected over two years (September 2021 to September 2023), covering various conditions and activities in the port. These images encompass diverse scenes under varying lighting and weather conditions, ensuring a comprehensive representation of the potential challenges in oil spill detection. To maintain privacy and confidentiality, sensitive information in the images (such as identifiable text, signs, and faces) has been anonymized using image inpainting and blurring techniques. Each image is precisely labeled to distinguish between oil, water, and other elements (such as vessels and docks). The labeling process used the CVAT tool and was reviewed by multiple experts to ensure accuracy and reliability. The final dataset has been released on the Zenodo platform, divided into training, validation, and test sets in a 70 / 15 / 15 ratio, supporting various machine learning applications and providing a crucial resource for developing robust detection models that can significantly improve response strategies and reduce the environmental impact of ports.

[0128] Table 1 provides a detailed statistical overview of the datasets used, categorized into three main groups (oil, water, and others). The table is divided into three rows, presenting the statistical information of the datasets in a concise manner:

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

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

[0131] The third row shows the relative labeling percentage for each category, providing a distribution of the proportions of each category in the dataset.

[0132] Table 1: Overview of Datasets by Category

[0133]

[0134] B. Experimental Setup

[0135] The experimental setup employed an improved U-Net architecture equipped with an EfficientNet-B4 encoder, specifically optimized for high-resolution UAV RGB image segmentation tasks. Model training utilized the Adam optimizer with an initial learning rate of 0.001, incorporating binary cross-entropy loss and Dice loss as custom loss functions to effectively balance precision and recall. Training lasted 50 epochs with a batch size of 16, and extensive data augmentation techniques were employed to enhance the model's robustness under diverse environmental conditions.

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

[0137] The performance of the PetroAnalyticNet model was evaluated using several key metrics that provide insights into the model's accuracy and efficiency in geological image segmentation tasks. The definitions and mathematical formulas for each metric are as follows:

[0138] F1 score:

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

[0140]

[0141] Wherein, Precision represents the proportion of correctly predicted positive examples out of all predicted positive examples, and Recall represents the proportion of correctly predicted positive examples out of all actual positive examples.

[0142] Accuracy:

[0143] Precision measures the accuracy of positive example predictions, and its calculation formula is as follows:

[0144]

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

[0146] Recall rate:

[0147] Recall (or sensitivity) measures a model's ability to identify all relevant instances, and its formula is:

[0148]

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

[0150] Cross-well ratio:

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

[0152]

[0153] Where A represents the predicted segmented region and B represents the actual segmented region.

[0154] in conclusion

[0155] The performance of the OURS model across various categories was rigorously evaluated using standard segmentation metrics, and compared with baseline models. Table 2 presents the optimized results, detailing the model's enhancement capabilities across 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 across all performance metrics. Although the improvements are relatively modest, these enhancements reflect practical and achievable progress. The main improvements are as follows:

[0159] For the petroleum category, the F1 score improved by 0.04, thanks to a balanced improvement in precision and recall, indicating that the model is more accurate in identifying and segmenting petroleum features.

[0160] Water Category: Precision improved by 0.03, and recall reached a perfect score, 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, showed consistent improvement across all metrics, with both the F1 score and IoU increasing by 0.04, indicating that the model's ability to handle diverse features has been improved.

[0162] The significant advantage of the OURS model lies in its ability to improve precision without sacrificing recall, thereby enhancing F1 score and IoU. These improvements are attributed to the innovative architectural features of OURS, including advanced dilated convolution and a robust feature fusion strategy. Experimental results demonstrate that OURS is an efficient tool for geological image segmentation tasks in oil and gas exploration, providing more reliable and accurate analysis for the decision-making process.

[0163] The accuracy curves for training and validation are as follows: Figure 6 As shown, the model's performance changes over 50 training epochs. Initially, both training and validation accuracies are low, reflecting the model's early learning phase. As training progresses, the accuracy on both the training and validation sets steadily increases, reaching a high value at the end of training. Validation accuracy exhibits a similar trend, slightly lower than training accuracy, indicating that the model has good generalization ability on unseen data.

[0164] The loss curves for training and validation are as follows: Figure 7 As shown, the model's loss changes over 50 training epochs. The training loss continuously decreases, reflecting a gradual reduction in the model's error on the training data. Similarly, the validation loss also shows a decreasing trend, but slows down slightly in the later stages. This indicates that the model is learning well, but the slower rate of decrease in the validation loss may suggest that further fine-tuning or regularization could help prevent overfitting.

[0165] The accuracy and loss curves show no significant separation between the training and validation curves, indicating that the model is not overfitting and demonstrates good generalization ability on the validation set. The steady decrease in loss and the continuous improvement in accuracy suggest that the current learning rate is suitable for the training process.

[0166] Ablation Research

[0167] To determine the contribution of specific components in the OURS model to the overall performance, an ablation study was conducted. The study systematically removed key features from the complete 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 significantly contribute to the efficiency and accuracy of the model. When this component is removed, all performance metrics decrease.

[0169] Dilated convolutions enhance the model's context awareness, which is crucial for capturing a wider range of geological features. Removing this component resulted in a significant decrease in all metrics.

[0170] Table 3: Ablation Study Results

[0171]

[0172]

[0173] Feature fusion strategies improve the model's ability to integrate multi-scale information, thereby enhancing its overall accuracy. Performance degrades when this component is removed, highlighting its importance in multi-scale feature processing.

[0174] Attention mechanisms help models focus on relevant features, and their removal directly leads to a decrease in precision and recall, demonstrating the crucial 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 synergistic effect is crucial for achieving high accuracy and efficiency in geological image segmentation tasks. This ablation study provides clear evidence that the complex architecture of the OURS model can effectively address the complex requirements of geological segmentation.

[0176] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection 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, Includes the following steps: The PetroAnalyticNet deep learning model includes an encoder and a decoder. The encoder comprises multiple convolutional blocks, each extracting different features from the geological input data through a set of filters, starting with basic features and progressively extracting them as the network deepens. After the convolutional layers, pooling operations condense the information into deeper abstractions to prepare feature maps. The convolutional blocks are separable dilated convolutional blocks, comprising dilated convolutions, separable convolutions, and residual connections. The dilated convolutions expand the network's receptive field. The separable convolutional blocks decompose the learning process into two layers: one for filtering features and the other for combining features. A series of residual blocks are stacked together, each performing an identity mapping to pass features from earlier layers to later layers. At the bottleneck, the network reaches its deepest layer, serving as a 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 from 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, and uses pixel-by-pixel classification to distinguish reservoir and non-reservoir areas, thereby determining the location of oil and gas reservoirs.

2. The geological image segmentation method for oil and gas reservoir exploration based on the Petro Analytic Net deep learning model according to claim 1, characterized in that: The separable convolutional block decomposes the convolution operation into two stages: depthwise convolution and pointwise convolution, specifically including: The first stage is depthwise convolution, where a single convolutional 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 It is the depth convolution output at position k, with m rows and n columns. 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 depthwise convolution: Where 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 Where O is the final output and I is the input.

3. The geological image segmentation method for oil and gas reservoir exploration based on the Petro Analytic Net deep learning model according to claim 2, characterized in that: The combination of depthwise convolution and dilated convolution extracts features through pointwise convolution and attention layers. include, The input is a depthwise 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 It is the depthwise convolution output of channel c, I c It is the input of channel c, K c Denotes the convolution kernel for channel c, (x . y . () are the coordinates within the kernel; The input undergoes dilated convolution to enhance the receptive field and capture a larger spatial context: D d (x,y)=I(x,y)*K * (x . +r·a,y . +r·a) Among them, D d K represents the dilated convolution output. * It is an expanding kernel, r is the expansion rate, and a is the spacing between kernel elements; The outputs of depthwise convolution and dilated convolution are combined and then compressed into a denser feature representation through pointwise convolution: Where P(x, y) is the pointwise convolution output, C is the number of channels, and W is the number of channels. c These are the weights of the 1×1 convolution kernel corresponding to channel c; The attention layer follows pointwise convolution, and the final output is a feature map: A(x,y)=σ(P(x,y)·W a ) Where A(x, y) is the output of the attention layer, and W a σ 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 geological image segmentation method for 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 processed through a set of 3x3 convolutional filters to extract features from the geological data. The mathematical expression for the convolution operation is: F(x)=W d *x+b Where F(x) is the output feature map, W d This represents the weights of the 3×3 filters, where x is the input representing the surface image and b is the bias term; Then, batch normalization is performed to standardize the feature maps to a mean of zero and a variance of one: in, This is the normalized output, μ is the average value, and σ is the normalized output. 2 is the variance, and ∈ is a small constant for numerical stability; Introducing nonlinearity into the ReLU activation function: R(x) = max(0, x) Where R(x) is the output after applying ReLU activation; Reduce the dimensionality of the feature map using point convolutions to prepare for residual connections: P(x)=W p *R(x) Where P(x) is the output after point convolution, and W p These are the weights of a 1×1 convolution; Add the point convolution output to the original input: O(x) = P(x) + x O(x) is the final output of the residual block.

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

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

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

Images