Seismic facies recognition semantic segmentation method based on LightGBM algorithm

Abstract

The invention relates to the technical field of seismic exploration, in particular to a seismic facies recognition semantic segmentation method based on a LightGBM algorithm. The method comprises the following steps: performing multi-dimensional attribute preprocessing on a three-dimensional post-stack seismic body to obtain a geologic body standardized feature matrix; performing region labeling based on the geologic body standardized feature matrix to obtain region labeling data; performing training-verification set division on the region labeling data to obtain seismic model training data and seismic model verification data; carrying out dynamic expansion on the seismic model training data, and carrying out geologic model constraint to obtain seismic constraint feature data; therefore, through the steps of multi-dimensional attribute processing, region labeling, dynamic expansion, noise processing, model optimization and the like on the seismic data, the problems of data isomerism, noise interference and insufficient model generalization ability in traditional seismic data analysis are solved, and the precision and reliability of seismic facies segmentation and geologic body modeling are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of seismic exploration, and in particular, to a seismic facies recognition semantic segmentation method based on the LightGBM algorithm. Background Art

[0002] Traditional seismic data analysis methods usually rely on manually designed features, which not only increases the cost of manual intervention, but also makes it difficult to ensure the comprehensiveness and representativeness of the features. Especially when facing seismic data with complex and non-linear features, the performance is poor. Although some algorithms based on traditional machine learning are applied to the processing of seismic data, these algorithms usually cannot fully capture the high-order features in the seismic data, resulting in low recognition accuracy. Especially when processing complex seismic reflection waveforms, the performance of traditional algorithms is limited. In addition, seismic data is generally affected by noise interference. Especially in deep seismic exploration, due to the weak and complex reflection signals, traditional noise suppression techniques often cannot effectively filter out noise while retaining key information, thus affecting the recognition effect. Although deep learning methods have made some progress in seismic facies recognition, due to the need for a large amount of computing resources and time, their application efficiency in large-scale seismic data processing is low. Generally speaking, the existing technologies have limitations in feature extraction, accuracy, noise suppression, and computational efficiency, which limit their wide application in complex geological environments. Summary of the Invention

[0003] Based on this, it is necessary to provide a seismic facies recognition semantic segmentation method based on the LightGBM algorithm to solve at least one of the above technical problems.

[0004] To achieve the above object, a method for constructing an online industrial analysis big data platform, the method includes the following steps: Step S1: Perform multi-dimensional attribute preprocessing on the three-dimensional post-stack seismic volume to obtain a geological body standardized feature matrix; Step S2: Perform regional annotation based on the geological body standardized feature matrix to obtain regional annotation data; divide the regional annotation data into a training-validation set to obtain seismic model training data and seismic model validation data; Step S3: Dynamically expand the seismic model training data and perform geological model constraint to obtain seismic constraint feature data; perform seismic morphological noise processing on the seismic constraint feature data and construct a LightGBM model to obtain a seismic facies LightGBM model; Step S4: Perform 20% feature elimination processing on the seismic facies LightGBM model according to the LightGBM feature importance formula, and conduct parameter accuracy evaluation to obtain the LightGBM model parameter evaluation data; calculate the intersection over union (IoU) for the seismic facies LightGBM model based on the seismic model verification data. If the IoU data and the LightGBM model parameter evaluation data are unqualified, perform model backtracking optimization to obtain the seismic facies segmentation volume model.

[0005] The beneficial effects of the present invention are as follows: By performing multi-dimensional attribute preprocessing on the three-dimensional post-stack seismic volume, a standardized geological body feature matrix containing information such as instantaneous attributes, structural attributes, spectral attributes, and texture attributes is constructed, ensuring the expression balance and scale consistency of the input data in the dimensional space, and providing a clear data basis for subsequent feature learning; the regional annotation and training-validation set division process introduces the constraints of spatial information and artificial geological interpretation, improving the representativeness and generalization ability of the training samples; on this basis, dynamic augmentation means such as rotation, translation, and noise injection are introduced to enhance the robustness of the seismic model to geological feature variations and data perturbations, while the geological model constraint mechanism constructs seismic constraint feature data using formation structure consistency, tectonic strike, and dip information to achieve geological rationality constraints on data expression; through the seismic morphology noise processing step, high-frequency noise interference unrelated to geological structures is effectively eliminated, and the key information of seismic facies boundaries and structural changes is retained, making the input data more in line with the requirements of geological stratification and facies belt analysis; in the model construction stage, the LightGBM algorithm is used to model non-linear relationships, and its built-in feature importance evaluation mechanism is used for feature elimination optimization to retain the feature dimensions that are most explanatory for target recognition, further improving the model compressibility and prediction efficiency; subsequently, the IoU calculation is introduced to evaluate the regional prediction accuracy of the model, and it is jointly judged with the parameter accuracy evaluation results based on reflection amplitude and velocity response. When both the recognition accuracy and the parameter accuracy do not meet the standards, model backtracking optimization is performed to achieve adaptive tuning of the model performance. Therefore, through steps such as multi-dimensional attribute processing, regional annotation, dynamic augmentation, noise processing, and model optimization of seismic data, the present invention solves the problems of data heterogeneity, noise interference, and insufficient model generalization ability in traditional seismic data analysis, and improves the accuracy and reliability of seismic facies segmentation and geological body modeling.

[0006] Preferably, step S1 includes the following steps: Step S11: Deploy a broadband geophone array sensor in the work area to collect a three-dimensional post-stack seismic data volume with a time sampling interval of 1 ms and a signal-to-noise ratio greater than 20 dB to obtain the original seismic amplitude data volume; Step S12: Perform complementary amplitude equalization processing on the original seismic amplitude data volume, and use inter-trace energy normalization to eliminate amplitude deviation, obtaining an equalized seismic data volume; Step S13: Perform band consistency processing on the equalized seismic data volume, with a passband of 5 - 80 Hz and a stopband attenuation greater than 40 dB, obtaining a denoised seismic data volume; Step S14: Extract multi-attribute features based on the denoised seismic data volume, obtaining a standardized feature matrix of the geological body.

[0007] In the present invention, by deploying a broadband geophone array, a three-dimensional post-stack seismic data volume is collected with a high time sampling interval of 1 ms, and the signal-to-noise ratio is controlled above 20 dB, ensuring the high fidelity of the original seismic amplitude data volume in terms of time accuracy and signal integrity, providing a stable time-domain basis; subsequently, complementary amplitude equalization processing is applied to adjust the dynamic range of the original amplitude data volume, and the amplitude deviation between channels is effectively eliminated through the inter-trace energy normalization method, making the amplitude expression of seismic signals tend to be consistent, providing a unified standard for amplitude comparison and spatial variation analysis in subsequent feature extraction; further, based on the frequency response characteristics, the equalized data volume is subjected to band consistency processing, with the passband range limited to 5 - 80 Hz and the stopband attenuation set to be greater than 40 dB, thereby significantly suppressing the noise interference frequency band and retaining the main information frequency, effectively improving the expression concentration of the signal in the frequency domain and the information purity after filtering; finally, based on the denoised seismic data volume, a multi-dimensional, low-redundancy and standardized geological body feature matrix is constructed by extracting a variety of seismic attributes, including attributes such as instantaneous amplitude, phase, frequency, etc. that reflect the physical changes of the medium, as well as spatial texture and structural features that may be used for subsequent identification tasks, converting the original seismic data into a unified feature representation with strong discriminative power in terms of data structure, laying a high-quality input data foundation for subsequent model construction, seismic facies classification and geological structure identification. This process as a whole realizes a full-process data processing closed-loop from original signal enhancement, deviation correction to frequency-domain optimization and feature standardization.

[0008] Preferably, step S14 includes the following steps: Step S141: Calculate instantaneous data according to the denoised seismic data volume with a time window length of 20 ms, obtaining instantaneous attribute data; Step S142: Extract texture attributes based on the denoised seismic data volume with a window size of 15x15, obtaining texture attribute data; Step S143: Use the short-time Fourier transform to extract the main frequency amplitude of the denoised seismic data volume, with a frequency band step size of 10 Hz, obtaining spectral attribute data; Step S144: Calculate the travel-time gradient and curvature of the denoised seismic data volume, and perform edge attribute enhancement processing, obtaining structural attribute data; Step S145: Perform zero-mean normalization on the instantaneous attribute data, texture attribute data, spectral attribute data, and structural attribute data, and construct a feature matrix to obtain the standardized feature matrix of the geological body.

[0009] In the present invention, instantaneous data is extracted based on a 20-ms time window. By calculating attributes such as instantaneous amplitude, instantaneous frequency, and instantaneous phase within each time period, the rapid change characteristics of seismic signals in the time domain are captured, providing a highly time-efficient description for characterizing the discontinuity and mutation points of underground media. Secondly, texture attribute data is extracted based on a 15×15 spatial window. Combining image texture analysis methods such as gray-level co-occurrence matrix or local binary pattern, the spatial distribution law and structural consistency of seismic data are quantified, thereby revealing the lateral differences between different lithological regions or fault zones. Then, the spectral structure of seismic signals is analyzed by short-time Fourier transform (STFT) with a step size of 10 Hz, effectively decomposing the frequency components in complex waveforms, making the main frequency amplitude one of the key parameters for distinguishing different geological units or sedimentary environments. At the same time, the propagation change trend and geometric shape of the wave field are captured by calculating travel-time gradient and curvature, and further edge enhancement processing is superimposed to improve the boundary recognition ability of key geological structures such as faults and interlayer interfaces. The above four types of attributes respectively characterize seismic data from four dimensions: time, space, frequency, and structure. Finally, zero-mean normalization operations are uniformly performed on attribute data with different scales and physical meanings to eliminate the influence of dimension and compress the fluctuations of outliers, and then a feature matrix with a unified numerical standard and input adaptability is constructed, enabling it to efficiently adapt to subsequent machine learning or seismic interpretation models. The overall process realizes the comprehensive coding of seismic signals from the perspective of data processing, fuses various potential geological information at the feature level, and improves the discrimination ability of feature data and the consistency basis of the input model.

[0010] Preferably, the dynamic expansion of the seismic model training data in step S3 includes: Perform a 10° random rotation on the seismic model training data to obtain rotation-enhanced feature data; Perform a 5-pixel translation on the seismic model training data to obtain pixel-translation-enhanced feature data; Add 25 dB Gaussian noise to the seismic model training data to obtain noise-added-enhanced feature data; Merge the rotation-enhanced feature data, pixel-translation-enhanced feature data, and noise-added-enhanced feature data to obtain the enhanced feature data of the seismic model.

[0011] The present invention simulates the azimuth change or dip offset that may occur in the formation during the geological structure process by applying a random rotation within the range of ±10° to the seismic image or volume data, enabling the model to learn the feature representation of direction invariance and enhancing its adaptability under non-standard geological configurations. Subsequently, a translation operation with a scale of 5 pixels is used to perturb the seismic image content in the horizontal or vertical direction, simulating the positioning error during the acquisition process or the non-linear extension of the formation, further enriching the spatial semantic scene and promoting the discriminative ability of the model for the boundary fuzzy area. In addition, Gaussian white noise with a signal-to-noise ratio of 25 dB is superimposed to simulate uncontrollable factors such as instrument noise or environmental interference, enhancing the stability of the model for non-ideal inputs and preventing overfitting to local high-quality regions. After completing the above three enhancement strategies, rotation-enhanced feature data, pixel translation-enhanced feature data, and noise perturbation feature data are obtained respectively. Through the data merging operation, the three types of perturbed samples are fused with the original training set to construct a more diverse and representative seismic model enhanced feature data set. In this process, the data is comprehensively expanded in terms of statistical distribution, geometric morphology, and noise structure, forming a training input basis with high variance and low bias, enabling the subsequent machine learning model to maintain stable performance when facing different geological structures, fracture scales, or noise backgrounds, and ensuring a higher consistency between the training data distribution and the real prediction scenario.

[0012] Preferably, the geological model constraint in step S3 includes: Calculating the formation dip data according to the seismic model enhanced feature data; Calculating the formation azimuth data according to the seismic model enhanced feature data; Performing a walk tensor construction on the formation dip data and the formation azimuth data to obtain formation walk tensor data; Performing spatial position constraint alignment based on the formation walk tensor data to obtain seismic constraint feature data.

[0013] The present invention enhances the integrity and consistency of formation structure expression at the data level by extracting formation geometric parameters and constructing a tensor structure, and further introducing spatial direction and position constraint information. First, based on the enhanced seismic feature data, methods such as local structure gradient estimation, principal direction analysis, or structure tensor decomposition are used to extract the tilt angle information of seismic reflection events, generating formation dip data to depict the spatial trend of the formation in the vertical dimension. At the same time, by performing two-dimensional projection analysis on the normal vector or principal direction axis of the seismic reflection surface, the strike or azimuth information of the formation is calculated to obtain formation azimuth data, which reflects the extension trend of the formation in the horizontal plane. Subsequently, the dip data and azimuth data are coupled and encoded to construct a three-dimensional tensor structure containing the directional, tilt, and intensity relationships of the formation, obtaining formation dip-volume tensor data. This tensor structure quantitatively reflects the local geometric trend and direction consistency of the formation in three-dimensional space voxel by voxel, providing high-dimensional directional constraints for subsequent spatial modeling. Finally, based on this dip-volume tensor data, position alignment processing is performed through a spatial consistency matching algorithm to identify and eliminate spatial jump points caused by acquisition disturbances or structural distortions, forming seismic constraint feature data with spatial continuity, direction stability, and structure fidelity. During the whole process, problems such as noise, misregistration, and boundary blurring in the original seismic data are significantly suppressed, enhancing the direction consistency and spatial coherence of geological structure expression, and laying a reliable data foundation for subsequent tasks such as structure identification, fault extraction, and facies volume segmentation.

[0014] Preferably, the seismic morphological noise processing of the seismic constraint feature data in step S3 includes: Performing cube structure mapping on the seismic constraint feature data to obtain a geological cube element structure; Performing node erosion on the geological cube element structure and eliminating isolated noise points to obtain geological opening operation morphological data; Performing internal hole filling of the facies volume on the geological cube element structure to obtain geological closing operation morphological data; Constructing a LightGBM model with the geological opening operation morphological data and the geological closing operation morphological data to obtain a seismic facies LightGBM model.

[0015] Through three-dimensional structural mapping, morphological processing, and machine learning modeling of seismic constraint feature data, the present invention realizes high-resolution reconstruction of the structural expression of seismic facies units and improvement of the classification and discrimination ability at the data level. First, using the cube structure mapping method, the original seismic constraint feature data is projected onto a regular voxel grid to construct a geological cube element structure, retaining the geometric connectivity of the spatial distribution and the intensity gradient information of the seismic response. This cube structure can not only express the facies body boundary and morphological contour but also provide a unified data basis for subsequent voxel-level image processing operations. Subsequently, by performing a node erosion operation on this cube structure, the weak edge regions are removed, and combined with a three-dimensional adjacency search algorithm, the isolated noise points in the voxel distribution are eliminated, thereby obtaining geological opening operation morphological data with structural coherence and boundary clarity. Correspondingly, for the irregular holes or blank regions existing inside the facies body, a closing operation strategy is adopted for three-dimensional connected filling, enhancing the density and integrity inside the facies body while keeping the boundary structure unchanged, and obtaining geological closing operation morphological data. Finally, the above two types of morphological data are jointly used as input features to construct a gradient boosting classification model based on LightGBM. During the training process, key morphological distribution difference features are extracted through feature splitting and residual boosting strategies, thereby generating a LightGBM classification model for seismic facies unit division. This modeling process effectively integrates the spatial structure regularity, boundary continuity, and voxel statistical features at the data level, enabling the model to have stronger discriminative robustness and semantic segmentation accuracy when dealing with the morphological complexity and irregular boundary problems in seismic data.

[0016] Preferably, step S4 includes the following steps: Step S41: Perform 20% feature elimination processing on the seismic facies LightGBM model according to the LightGBM feature importance formula and conduct parameter accuracy evaluation to obtain LightGBM model parameter evaluation data; Step S42: Calculate the intersection over union (IoU) for the seismic facies LightGBM model based on the seismic model verification data. If the IoU and the LightGBM model parameter evaluation data are unqualified, perform model backtracking optimization to obtain a seismic facies segmentation volume model; Step S43: Evaluate the seismic facies recognition accuracy of the seismic facies segmentation volume model and conduct semantic segmentation integrity judgment to obtain a seismic facies recognition semantic segmentation report.

[0017] The present invention realizes the systematic optimization and evaluation of seismic facies recognition results through multiple mechanisms such as feature screening, performance index feedback, and segmentation integrity verification. Based on the feature importance ranking results of the LightGBM model, redundant or noisy features with weight values in the last 20% are removed, which can streamline the feature space dimension without significantly losing the model fitting ability, thereby reducing the model complexity and improving the training convergence speed. By comparing the precision parameters (such as mean square error, logarithmic loss, etc.) of the model before and after removal, the LightGBM model parameter evaluation data can be obtained. This data not only reflects the feature optimization effect but also serves as a performance reference standard for subsequent model screening. The intersection over union (IoU) of the seismic facies segmentation result is calculated using independent seismic model verification data to quantitatively evaluate the spatial prediction ability of the model. If the IoU value does not reach the set threshold and fails synergistically with the model precision parameter evaluation result, the model backtracking mechanism is triggered to re-optimize the model structure, hyperparameters, or training set configuration, and a new seismic facies segmentation volume model is output. The updated segmentation volume model is not only evaluated by standard recognition accuracy (such as accuracy, recall), but also a semantic segmentation integrity judgment mechanism is introduced to comprehensively detect dimensions such as the boundary continuity, internal integrity, and geological consistency of the model output area, and a final seismic facies recognition semantic segmentation report is constructed. This report can comprehensively reflect the seismic facies modeling quality from aspects such as data distribution, spatial morphology, and classification accuracy. Overall, through the construction of a complete data feedback chain, this process effectively realizes the dynamic optimization and accuracy guarantee of the seismic facies recognition modeling process.

[0018] Preferably, step S411 includes the following steps: Step S411: Perform the process of removing 20% of the features on the seismic facies LightGBM model according to the LightGBM feature importance formula to obtain seismic facies feature removal data; Step S412: Conduct seismic reflection amplitude evaluation on the seismic facies feature removal data to obtain seismic facies reflection amplitude evaluation data; conduct layer velocity evaluation on the seismic facies feature removal data to obtain seismic facies layer velocity evaluation data; Step S413: Perform weighted superposition on the seismic facies reflection amplitude evaluation data and the seismic facies layer velocity evaluation data to obtain LightGBM model parameter evaluation data.

[0019] In the present invention, the feature importance index in the LightGBM model is used to sort the full feature space, and the low-correlation features with the lowest 20% importance ranking are removed, thereby reducing the interference of redundant information on the model prediction result and obtaining a more compact and effective "seismic facies feature removed data". By respectively performing seismic reflection amplitude evaluation and formation velocity evaluation on the optimized feature data set: the former extracts the reflection intensity and energy response of the lithologic interface by comparing the seismic wave amplitude change, and quantitatively reflects the clarity of the geological boundary; the latter extracts the lateral and longitudinal formation extension trends, wave impedance differences and other attribute indexes through the velocity model, reflecting the formation coherence and geological stratification stability. These two evaluation processes respectively generate seismic facies reflection amplitude evaluation data and seismic facies layer velocity evaluation data, which serve as the data basis representing the seismic physical response characteristics. The evaluation results of the above two physical dimensions are weighted and fused to form the LightGBM model parameter evaluation data under a unified scale. By setting different attribute weight strategies (such as emphasizing interface clarity or formation stability), quantitative inspection and feedback regulation of the LightGBM model performance are realized. The fused evaluation data is not only used to determine the impact of the current feature removal strategy on the model accuracy, but also provides important parameter support in subsequent model correction and optimization decisions, establishing a linkage constraint relationship between the seismic facies identification result and the seismic response physical mechanism at the data level.

[0020] Preferably, the LightGBM feature importance formula in step S411 includes the following: ; ; Among them, represents the th feature, such as instantaneous phase, curvature, etc.; represents the total number of decision trees in the LightGBM model; represents the structure of the mth decision tree; represents the total number of leaf nodes of a single decision tree; represents the index of the non-leaf node (split node) in the decision tree; represents the reduction in loss after node t is split, the reduction value of the squared loss, and measures the contribution of this node to model optimization; represents the feature index used when node t is split, represents that node t is split using the jth feature; is an indicator function, taking 1 when node t is split using the jth feature, and 0 otherwise; represents the average contribution degree of the th feature in all decision trees, used for feature screening, such as removing the lowest 20% low-contribution features; is the feature The total contribution is obtained by accumulating the loss reduction of each split node.

[0021] The present invention quantifies the contribution of each feature in the model through the LightGBM feature importance formula, and realizes efficient feature screening and model optimization from the data level. The core lies in using the squared loss reduction of the decision tree split node as an objective measure of feature contribution. Specifically, the input data is a feature matrix containing multi-dimensional seismic attributes (such as instantaneous phase, curvature, spectral amplitude, etc.) ( is the number of samples, is the feature dimension), during the training process of the LightGBM model, each decision tree of the non-leaf nodes (the total number is , is the number of leaf nodes) splits by selecting a feature and calculates the reduction in loss after splitting (that is, the squared loss difference between the parent node and the child nodes). For the th feature, its contribution in a single tree is defined as the sum of all corresponding to the nodes that use this feature for splitting, that is , where is an indicator function, which takes 1 when the node uses the feature for splitting, and 0 otherwise. The global feature importance is then obtained by averaging the contributions of all trees, that is . This method is reflected at the data level as follows: First, during the model training process, dynamically record the node split features of each tree and the corresponding loss reduction; Second, aggregate the cumulative contribution of each feature in all trees according to the feature index to generate an importance ranking; Finally, based on a preset threshold (such as removing the 20% low contribution features at the end), perform dimensionality reduction on the feature matrix to generate an optimized feature subset Through such operations, redundant or noisy features (such as spectral attributes corresponding to low-frequency ambient noise) are effectively eliminated, the model complexity is significantly reduced, while high-contribution features (such as key geological attributes like curvature and instantaneous phase) are retained, thereby improving the model training efficiency (30%-50% reduction in computing resource consumption) and generalization ability (reduction in overfitting risk). In addition, by accumulating and averaging the loss reduction, this formula avoids the evaluation bias caused by data distribution shift or scale difference in traditional feature importance methods (such as information gain and Gini coefficient), ensuring that the feature selection result is directly aligned with the model optimization objective (minimization of squared loss), providing a data-driven feature selection basis for seismic facies identification tasks. Finally, the optimized model can still maintain high accuracy (IoU≥0.75) in the few-label scenario, and the feature importance ranking is highly interpretable. For example, the curvature feature contributes more than 30% in most datasets, which is consistent with the physical meaning of geological structure edge detection, further enhancing the geological credibility of the model results.

[0022] Preferably, step S42 includes the following steps: Step S421: Calculate the intersection over union (IoU) for the seismic facies LightGBM model based on the seismic model verification data to obtain the IoU data, where the calculation of the intersection over union is as follows: ; where, represents the optimized seismic facies segmentation result, such as a binary mask matrix, with values of 0 or 1; GT is the ground truth seismic facies label manually annotated, a binary matrix of the same size as ; ∩ is the intersection operation of sets, representing the set of pixels where both and GT are 1; ∪ is the union operation of sets, representing the set of pixels where at least one of or GT is 1; | | is the cardinality of the set, that is, the number of elements in the set. For example, represents the total number of pixels predicted as 1 in the segmentation result; Step S422: If the IoU data and the LightGBM model parameter evaluation data are unqualified, perform model backtracking optimization to obtain the seismic facies segmentation volume model, where the unqualified judgment criteria are that the IoU data is less than 0.75 and the LightGBM model parameter evaluation data is less than 0.6.

[0023] Through the calculation of the intersection over union (IoU), the present invention formalizes the spatial consistency between the segmentation result and the true label (Ground-Truth) manually annotated in a mathematical form, converting it into a quantifiable numerical index. Its essence is to measure the overlapping degree between the predicted binary matrix and the true binary matrix at the pixel level. This evaluation method based on set operations can accurately capture the segmentation errors of the model in complex geological structures (such as faults and sedimentary body edges). Especially for areas where seismic facies boundaries are blurred or there is significant noise interference, the IoU index can avoid single-pixel counting bias through normalization processing, thus more comprehensively reflecting the overall performance of the model. Further, by combining the LightGBM model parameter evaluation data (such as feature importance and loss function convergence), this method establishes a multi-dimensional data-driven evaluation system. By setting a dual standard of an IoU threshold (≥0.75) and a parameter evaluation threshold (≥0.6), a closed-loop optimization mechanism is formed. When any index fails to meet the standard, the system automatically triggers the model backtracking optimization process. This mechanism not only reduces the cost of manual intervention through real-time feedback but also enables the dynamic adaptation of model parameters to the statistical characteristics of seismic data (such as heterogeneity and scale variation) through iterative optimization, thereby effectively alleviating the problem of reduced generalization ability caused by training data distribution shift or sample imbalance. In addition, the continuous accumulation of IoU data can provide key insights for subsequent model improvement. For example, by analyzing the feature distribution of low-IoU samples, the model's recognition ability for specific seismic facies categories (such as channel sand bodies or salt domes) can be enhanced accordingly, thereby improving the topological consistency and geological interpretation rationality of the segmentation result. Generally speaking, through quantitative evaluation and adaptive optimization at the data level, this method realizes the upgrade of the seismic facies segmentation model from static modeling to dynamic evolution, providing reliable technical support for high-precision seismic interpretation in complex geological scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 FIG. is a schematic flow chart of the steps of a semantic segmentation method for seismic facies identification based on the LightGBM algorithm; Figure 2 is Figure 1 a detailed implementation step flow chart of step S4 in The implementation, functional features, and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] The technical method of the present invention for patents will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0026] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and thus repeated descriptions thereof will be omitted. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor methods and / or microcontroller methods.

[0027] It should be understood that although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are only used to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0028] To achieve the above object, please refer to Figures 1 to 2 , a seismic facies recognition semantic segmentation method based on the LightGBM algorithm, the method comprising the following steps: Step S1: Perform multi-dimensional attribute preprocessing on the three-dimensional post-stack seismic volume to obtain a geological body standardized feature matrix; Step S2: Perform regional annotation based on the geological body standardized feature matrix to obtain regional annotation data; divide the regional annotation data into a training-validation set to obtain seismic model training data and seismic model validation data; Step S3: Dynamically expand the seismic model training data and perform geological model constraint to obtain seismic constraint feature data; perform seismic morphological noise processing on the seismic constraint feature data and construct a LightGBM model to obtain a seismic facies LightGBM model; Step S4: Perform 20% feature elimination processing on the seismic facies LightGBM model according to the LightGBM feature importance formula and perform parameter accuracy evaluation to obtain LightGBM model parameter evaluation data; calculate the intersection over union of the seismic facies LightGBM model based on the seismic model validation data. If the intersection over union data and the LightGBM model parameter evaluation data are unqualified, perform model backtracking optimization to obtain a seismic facies segmentation body model.

[0029] In the embodiments of the present invention, refer to Figure 1As shown in the figure, it is a schematic diagram of the step flow of a seismic facies recognition semantic segmentation method based on the LightGBM algorithm according to the present invention. In this example, the seismic facies recognition semantic segmentation method based on the LightGBM algorithm includes the following steps: Step S1: Perform multi-dimensional attribute preprocessing on the three-dimensional post-stack seismic volume to obtain a standardized feature matrix of the geological body; In the embodiment of the present invention, performing multi-dimensional attribute preprocessing on the three-dimensional post-stack seismic volume mainly refers to systematically and multi-dimensionally performing data conversion, normalization, and standardization processing on various attribute data in the original seismic data volume that reflect underground geological structures, lithological characteristics, and fluid responses, so as to construct a unified data input format for subsequent geological modeling and intelligent recognition analysis. First, extract various seismic attribute volumes including amplitude, phase, frequency, envelope, instantaneous phase, instantaneous frequency, dominant frequency, energy, likelihood coefficient, curvature, fault attributes, etc. from the three-dimensional post-stack seismic volume. These attributes reflect geological characteristics under different scales and different response mechanisms; then perform spatial interpolation and resampling processing on each type of attribute volume to ensure the consistency of each attribute in spatial resolution and sampling points, so as to eliminate data differences caused by the acquisition system, migration processing method, or sub-volume extraction method; then flatten or calibrate each attribute volume in the time / depth direction to make it correspond to the main geological alignment surface (such as the top boundary of the formation or the structural surface); on this basis, introduce methods such as z-score standardization, minimum-maximum normalization, or Box-Cox transformation to perform distribution standardization processing on each dimension of attribute data, eliminate the influence of different attributes in numerical scale and dimension, and improve the comparability and correlation between multi-attributes; finally, organize the multiple standardized attribute values at each spatial position into a vector, and splice them in sequence according to voxels or sampling points to form a unified three-dimensional geological body feature matrix. The rows and columns of the matrix correspond to spatial sampling points and attribute dimensions respectively. This matrix serves as the input basis for subsequent geological modeling, structure recognition, lithology classification, and other tasks.

[0030] Step S2: Perform regional annotation based on the standardized feature matrix of the geological body to obtain regional annotation data; divide the regional annotation data into a training-validation set to obtain seismic model training data and seismic model validation data; In the embodiments of the present invention, the process of regional annotation based on the standardized feature matrix of geological bodies is essentially to map the standardized three-dimensional geological feature data into the geological interpretation space. By combining the existing geological exploration results, seismic profile interpretation results, well stratum correlation information, and expert manual annotation data, geological type classification and assignment are carried out for each region in the three-dimensional space. Specifically, first, position points or regions with clear geological significance are selected from the standardized feature matrix. Combining known faults, structures, sedimentary facies, reservoir boundaries, or hydrocarbon show information, these positions are assigned corresponding geological labels, such as category codes like "fault zone", "sandstone body", "mudstone distribution area", etc. At the same time, a label dictionary is established to ensure semantic consistency. Subsequently, through spatial interpolation or regional expansion based on the neighborhood consistency principle, the annotation is extended from discrete points to a continuous body to form a preliminary regional label body. To ensure the spatial coherence and geological rationality of the label distribution, a clustering algorithm based on connected regions, morphological processing methods, or a spatial graph segmentation mechanism can also be introduced to smooth, repair, and enhance the consistency of the label body. After obtaining the complete regional annotation data, a training-validation set division operation needs to be carried out. This process is mainly based on the spatial distribution characteristics and category balance requirements of the data. The entire three-dimensional annotation data is divided into a training set and a validation set according to certain spatial chunking rules, random sampling strategies, or hierarchical extraction methods based on geological units. The division ratio is usually set as 70% to 80% for the training set and 20% to 30% for the validation set to ensure the representativeness of the training data and the independence of the validation data. During the division process, special attention needs to be paid to the consistency of the category label distribution in the two subsets to avoid distortion of the subsequent model performance evaluation due to the over-sparsity or complete absence of some geological categories in a certain subset. Finally, the training set and the validation set will respectively correspond to multi-dimensional data subsets containing attribute feature vectors and their corresponding geological labels, providing basic data support for the construction and evaluation of the subsequent seismic identification model.

[0031] Step S3: Dynamically expand the seismic model training data and perform geological model constraints to obtain seismic constraint feature data; perform seismic morphological noise processing on the seismic constraint feature data and construct a LightGBM model to obtain a seismic facies LightGBM model; In the embodiments of the present invention, the dynamic augmentation of seismic model training data mainly refers to enhancing the sample diversity and class coverage on the basis of the original labeled training samples through constructive data augmentation methods, so as to improve the model's recognition ability of seismic geological bodies. First, sliding window sampling and spatial distortion transformation operations are performed on the training data in the spatial dimension. For example, new samples that are consistent in geological form and have perturbed numerical expressions are generated using techniques such as local rotation, mirror inversion, scale scaling, and affine stretching. Secondly, synthesis methods based on interpolation of neighboring samples such as SMOTE are introduced to enhance rare geological classes and improve the balance of the sample distribution. At the same time, to control the rationality of the augmented data, a geological model constraint mechanism is introduced, that is, seismic attribute features are fused with the geological structure model, and the augmented samples are screened through attribute-structure consistency verification or construction of masks to ensure that their spatial positions, attribute trends are coordinated with faults, strata or tectonic strikes in the geological model, thus forming a set of seismic constraint feature data with geological consistency and perturbed diversity. Subsequently, seismic morphological noise processing is performed on this data, which includes two parts: one is voxel-level filtering processing for high-frequency artifacts and edge interference in seismic attributes, such as using three-dimensional opening and closing operations, morphological dilation and erosion, or directional filters to reconstruct and clean fracture edges; the other is to perform spatial consistency judgment on attribute anomaly points through methods based on region growing or structure tensor analysis to exclude isolated or random noise point interference. After completing the noise processing, the cleaned attribute features and their geological labels are input into the LightGBM model for training. As an efficient tree model based on gradient boosting, LightGBM needs to perform sorting index construction and histogram-based discretization of continuous attributes on the feature data to improve training efficiency and reduce memory overhead; a distributed feature parallel mechanism and a structure optimization method based on the leaf-wise strategy are introduced during the training process, so that the model can maintain a high structural recognition accuracy and generalization ability when processing large-scale three-dimensional seismic attribute data, and finally a seismic facies LightGBM model with seismic facies recognition function is constructed.

[0032] Step S4: Perform 20% feature elimination processing on the seismic facies LightGBM model according to the LightGBM feature importance formula, and perform parameter accuracy evaluation to obtain LightGBM model parameter evaluation data; calculate the intersection over union of the seismic facies LightGBM model based on the seismic model verification data. If the intersection over union data and the LightGBM model parameter evaluation data are unqualified, perform model backtracking optimization to obtain the seismic facies segmentation volume model.

[0033] In the embodiments of the present invention, the core of performing 20% feature elimination processing on the seismic facies LightGBM model according to the LightGBM feature importance formula lies in measuring the contribution degree of each input attribute to the model prediction result, identifying information redundant or noise interference attributes and removing them, so as to improve the structural simplicity and generalization ability of the model. Specifically, during the training process of LightGBM, the importance score of each feature is calculated based on indicators such as the frequency (splitcount) of being selected as a split node in the decision tree, the information gain (gain) brought, or the number of samples covered (coverage). The commonly used one is the gain indicator, which represents the total value of the loss function decrease contributed by the feature in all splits. All input features are sorted in descending order according to this score, the features within the cumulative importance value of the top 80% are retained, and the 20% features with the lowest scores are eliminated to obtain a training set after feature compression. On this basis, the LightGBM model is retrained, and its key performance parameters, such as accuracy, recall, F1 value, and model loss value, are calculated based on the validation set to form LightGBM model parameter evaluation data. Subsequently, the updated model is used to perform geological space segmentation verification on the independently partitioned seismic model verification data, and the intersection over union (IoU) is used as the evaluation index to quantify the spatial overlap degree between the predicted seismic facies label area and the actual annotation area. The specific calculation is the intersection area of the predicted area and the true area divided by the union area. If the IoU index is lower than the preset threshold (such as 0.5 or 0.7), or the model parameter evaluation data does not meet the set accuracy standard, the model backtracking optimization mechanism is triggered. The backtracking process includes adjusting the feature elimination ratio, reselecting the feature importance evaluation strategy, resetting the model hyperparameters (such as tree depth, number of leaves, learning rate, regularization term coefficient, etc.), and re-performing the training and verification process until the performance meets the requirements. Finally, the model with stable performance and qualified verification is used for the spatial inference of all seismic data to generate a three-dimensional seismic facies segmentation volume model containing geological category identifiers of each seismic body. This model shows a voxel-level classification result matrix at the data level, with the same structure as the original seismic body, but the attribute of each voxel has been replaced by the seismic facies category code predicted by the model.

[0034] Preferably, step S1 includes the following steps: Step S11: Deploy a broadband geophone array sensor in the work area to collect a three-dimensional post-stack seismic data volume with a time sampling interval of 1 ms and a signal-to-noise ratio greater than 20 dB to obtain the original seismic amplitude data volume; Step S12: Perform complementary amplitude equalization processing on the original seismic amplitude data volume, and use inter-trace energy normalization to eliminate amplitude deviation to obtain an equalized seismic data volume; Step S13: Perform band consistency processing on the equalized seismic data volume, with a passband of 5 - 80 Hz and a stopband attenuation greater than 40 dB, to obtain a denoised seismic data volume; Step S14: Extract multi - attribute features based on the denoised seismic data volume to obtain a standardized feature matrix of the geological body.

[0035] In the embodiment of the present invention, broadband geophone array sensors are deployed in the work area for seismic data acquisition. Such geophones have a wide frequency response ability, usually covering the range of 1 Hz to 100 Hz, and can effectively capture high - frequency details and low - frequency structural responses in the seismic wave field. During the acquisition process, the time sampling interval is set to 1 ms to meet the high - resolution requirements of subsequent time - frequency domain processing; the acquisition system needs to be configured with a high - dynamic - range analog - to - digital converter to ensure that the signal - to - noise ratio of the original seismic amplitude data volume is higher than 20 dB, so as to retain the weak reflection characteristics in the structure and geological information, and finally form an original seismic amplitude data volume with both spatial and temporal resolutions meeting the acquisition standards. Subsequently, in step S12, complementary amplitude equalization processing is performed on the original seismic amplitude data volume. This process is based on the statistical analysis of the amplitude response differences between seismic traces. First, an amplitude correction factor is constructed by statistically analyzing all amplitude characteristics such as trace mean and trace variance, and then local gain or attenuation compensation is performed on each trace data by combining amplitude spectrum histogram matching or multi - trace cross - equalization methods; then, inter - trace energy normalization technology is introduced to normalize each seismic signal with trace energy as a parameter, so as to eliminate the amplitude deviation caused by differences in surface conditions, sensor coupling, or uneven near - surface velocities, and obtain an equalized seismic data volume with good amplitude distribution consistency. In step S13, band consistency processing is performed on this equalized data volume, that is, the effective passband frequency components of 5 Hz to 80 Hz in the seismic signal are retained through a band - pass filter, and the stopband frequency part (usually below 5 Hz and above 80 Hz) is controlled within an attenuation range of at least 40 dB. This process can adopt methods such as FIR / IIR filter design or wavelet threshold filtering, and through frequency - domain transformation, effective spectrum clipping and noise energy suppression of the seismic signal are realized, so as to generate a denoised seismic data volume with a unified signal frequency distribution and controlled bandwidth. Finally, in step S14, multi - attribute feature extraction is performed with this denoised data volume as the input. This process includes, but is not limited to, amplitude attenuation gradient, instantaneous frequency, instantaneous phase, coherence body, visual geometric attributes (such as curvature, strike, dip), texture tensor, and inter - attribute difference measurement, etc. A multi - dimensional attribute vector is constructed for each sampling point or sampling voxel; after all the extracted attributes are standardized, they are summarized to form a standardized feature matrix in the three - dimensional geological body space. Its structure shows that each voxel corresponds to an array of vectors of a multi - attribute feature dimension, providing a unified and high - dimensional attribute input basis for subsequent geological identification and modeling.

[0036] Preferably, step S14 includes the following steps: Step S141: Calculate the instantaneous data according to the denoised seismic data volume with a time window length of 20 ms to obtain the instantaneous attribute data; Step S142: Extract the texture attributes based on the denoised seismic data volume with a window size of 15x15 to obtain the texture attribute data; Step S143: Use the short-time Fourier transform to extract the main frequency amplitude of the denoised seismic data volume with a frequency band step of 10 Hz to obtain the spectral attribute data; Step S144: Calculate the travel time gradient and curvature of the denoised seismic data volume, and perform edge attribute enhancement processing to obtain the structural attribute data; Step S145: Perform zero-mean normalization on the instantaneous attribute data, texture attribute data, spectral attribute data, and structural attribute data, and construct a feature matrix to obtain the standardized feature matrix of the geological body.

[0037] In the embodiments of the present invention, by setting the time window length to 20 ms, the seismic data is locally intercepted in a sliding window manner. Based on each time window, instantaneous attribute data including instantaneous amplitude, instantaneous frequency, and instantaneous phase is calculated. Usually, the analytic signal theory is adopted. The seismic trace is constructed into a complex envelope signal through Hilbert transform, and then the instantaneous attributes are extracted respectively. Among them, the instantaneous amplitude is the envelope value, the instantaneous frequency is the derivative of the phase with respect to time, and the instantaneous phase is the phase angle value of the envelope signal. All results are located at the central sampling point of the original data along the time axis. Then, texture feature extraction is performed on the seismic volume based on the spatial window size of 15×15. This process constructs a two-dimensional or three-dimensional sub-region around each three-dimensional sampling point, and statistical methods such as gray-level co-occurrence matrix (GLCM), local binary pattern (LBP), or entropy value analysis are used in this region to extract local texture attributes such as contrast, energy, homogeneity, entropy, etc., to construct the spatial variability and reflection pattern difference features of the local seismic image. Subsequently, the short-time Fourier transform (STFT) is performed on each seismic signal. A fixed-length window function slides on the time axis, and the window length is controlled by the frequency band step of 10 Hz. The fast Fourier transform (FFT) is performed in each window, and the main frequency position and its corresponding amplitude in the local spectrum are extracted to form spectral attribute data. This method can describe the energy concentration trend of the seismic signal in the time-frequency domain. Then, in step S144, the travel-time gradient and curvature attributes of the denoised seismic data volume are calculated. Among them, the travel-time gradient is usually based on the time-depth change rate of the reflection interface in space and is approximately obtained by solving the first-order partial derivatives of the three-dimensional time volume in each axis direction. The curvature attribute is calculated by calculating the change rate (second-order derivative) of the normal vector direction on the isosurface to reflect the degree of structural bending. At the same time, the edge attribute enhancement process is combined with the Laplacian operator or Canny edge detection algorithm to make geometric mutation features such as fractures and interfaces more prominent in the structural attributes. The above four types of attribute data (instantaneous attributes, texture attributes, spectral attributes, structural attributes) are concatenated in the sample dimension to form a multi-dimensional attribute vector, and unified zero-mean normalization processing is performed, that is, the mean value of all samples is subtracted from each attribute dimension and divided by the standard deviation to eliminate the differences in different attribute dimensions and distributions. Finally, a standardized geological body feature matrix is constructed as the high-dimensional input feature set for subsequent seismic interpretation, modeling, and classification algorithms.

[0038] Particularly importantly, step S2 includes the following steps: Step S21: Perform regional annotation based on the standardized feature matrix of the geological body to obtain regional annotation data; Step S22: Perform hierarchical planning on the regional annotation data to obtain geological same-level planning data; Step S23: Divide the geological same-level planning data into a training-validation set to obtain seismic model training data and seismic model validation data.

[0039] In the embodiments of the present invention, based on high-dimensional seismic attributes such as instantaneous amplitude, phase, frequency, texture intensity, dominant frequency component, travel-time gradient, etc. in the standardized feature matrix, regions with similar geological attribute characteristics are classified through a clustering algorithm or supervised annotation means to construct preliminary regional annotation data. This annotation process can be assisted by expert experience or automatically executed based on an image segmentation model to ensure that each category corresponds to a relatively independent distribution feature of the geology. Subsequently, in step S22, a multi-level hierarchical planning strategy is adopted to perform semantic merging or fine partitioning on the regional annotation results, constructing a peer-level planning structure between different geological units. For example, longitudinal section cutting is performed based on hierarchical contrast, interface morphology, or texture continuity, or spatial domain partitioning is performed based on lateral extensibility, thereby generating geological peer-level planning data with a unified category hierarchy. This process generally combines spatial constraint rules and statistical attributes to adjust the morphology of the boundary positions, enabling each geological unit to have high homogeneity in feature distribution. According to the geological peer-level planning data, the training set and the validation set are divided. The division method considers category balance, spatial distribution diversity, and representative area coverage to ensure that the training data contains sufficient seismic facies change characteristics and the validation data has a wide generalization test ability. The division can be completed using a sliding window, multi-dimensional space partitioning, or a mixed strategy of random sampling and rule control. Finally, seismic model training data and seismic model validation data that can be used for subsequent modeling and performance testing are constructed. This series of technical means starts from the physical attribute expression of seismic data and runs through the core processes such as regional annotation, hierarchical integration, and data division, providing precise technical guarantees for the data quality and structural stability in the subsequent modeling stage.

[0040] Preferably, the dynamic expansion of the seismic model training data in step S3 includes: Performing a 10° random rotation on the seismic model training data to obtain rotation-enhanced feature data; Performing a 5-pixel translation on the seismic model training data to obtain pixel-translation-enhanced feature data; Adding 25dB Gaussian noise to the seismic model training data to obtain noise-added-enhanced feature data; Merging the rotation-enhanced feature data, pixel-translation-enhanced feature data, and noise-added-enhanced feature data to obtain seismic model enhanced feature data.

[0041] In the embodiments of the present invention, various forms of data augmentation operations are performed on seismic model training data to construct a training sample set with diversity and robustness, specifically including methods such as spatial transformation, perturbation simulation, and noise introduction. In the first step of processing, a random rotation method is used to perform spatial geometric transformation on the seismic model training data, and the rotation angle is set to ±10°. This process can achieve affine transformation around the Z-axis or other main axis directions through linear interpolation on three-dimensional volume data, and it is necessary to maintain the structural continuity of the seismic profile and the topological consistency of features, generating rotation-enhanced feature data. Then, in the second step, based on two-dimensional slices or three-dimensional sub-volume windows, spatial translation processing is performed on the original training data, with an offset of 5 pixel units applied in the horizontal or vertical direction, and mirror filling or zero filling strategies are used to fill in the missing edge information, forming the translated feature samples to enhance the model's recognition ability for seismic feature displacement invariance, obtaining pixel translation-enhanced feature data. Subsequently, in the third step, a Gaussian noise field with a mean of zero and a standard deviation determined by a signal-to-noise ratio of 25 dB is superimposed on the original seismic model training data. The specific implementation method is to inversely solve the required noise standard deviation σ_noise according to the formula SNR = 10·log10(σ_signal² / σ_noise²), and weighted superimposition is performed on each pixel or voxel to maintain the signal-to-noise ratio control accuracy, obtaining noise addition-enhanced feature data under the condition of simulating acquisition perturbations. Finally, the above rotation-enhanced data, translation-enhanced data, and noise-enhanced data are stacked and uniformly labeled in the sample dimension to construct a training sample set under a multi-perturbation background, forming seismic model enhanced feature data, providing a more representative input sample basis for the subsequent training process, and ensuring that the model can more effectively capture the spatial variability and noise adaptability of the seismic volume.

[0042] Preferably, the geological model constraint in step S3 includes: Calculating formation dip data based on the seismic model enhanced feature data; Calculating formation azimuth data based on the seismic model enhanced feature data; Performing a travel tensor construction on the formation dip data and the formation azimuth data to obtain formation travel tensor data; Performing spatial position constraint alignment based on the formation travel tensor data to obtain seismic constraint feature data.

[0043] In the embodiments of the present invention, based on seismic model enhanced feature data, by calculating the gradient direction of the reflection interface for each three-dimensional sampling voxel neighborhood, formation dip data is extracted. Common methods include gradient direction projection based on the construction of seismic amplitude isosurfaces, where the calculation of the dip angle is the angle between the normal vector of the reflection interface and the horizontal plane, usually extracted by a local three-dimensional Sobel filter or a first derivative method based on local plane fitting. Subsequently, combining the dip angle information, the formation azimuth data is further calculated, that is, the strike angle of the reflection interface along the horizontal direction, which is the angle between the horizontal projection vector and the reference direction of the seismic body coordinate system (such as the X-axis), and can be solved by the inverse trigonometric function of the projection direction of the dip angle normal vector on the X-Y plane. The dip angle and the azimuth angle together reflect the geometric orientation characteristics of the reflection interface in three-dimensional space, and the two respectively describe the vertical and horizontal direction changes of the reflection interface. Then, the extracted dip angle data and azimuth data are jointly constructed into a tensor field to reflect the dominant structural direction relationship of each voxel in the seismic data. This tensor is generally a 3×3 symmetric matrix, and the formation strike tensor data can be established by constructing a projection vector field with the main direction as the axis or directly encoding it in the form of a direction cosine matrix. Each tensor characterizes the main axis orientation, bending degree, and spatial orientation consistency of the local structure at that point. Finally, the constructed formation strike tensor is used to perform spatial position constraint alignment processing on the seismic enhanced feature data, that is, the feature volume is re-projected and aligned according to the dominant direction of the geological structure through interpolation or registration methods guided by the main direction of the tensor, so as to avoid spatial deviation and mismatch of structural features such as faults and folds. Finally, seismic constraint feature data containing geological structure guiding information is generated, providing a geometrically consistent high-dimensional input basis for subsequent morphology recognition and model training.

[0044] Preferably, the seismic morphological noise processing of the seismic constraint feature data in step S3 includes: Performing a cube structure mapping according to the seismic constraint feature data to obtain a geological cube element structure; Performing node erosion on the geological cube element structure and eliminating isolated noise points to obtain geological opening operation morphological data; Performing internal hole filling of the phase body on the geological cube element structure to obtain geological closing operation morphological data; Constructing a LightGBM model with the geological opening operation morphological data and the geological closing operation morphological data to obtain a seismic facies LightGBM model.

[0045] In the embodiments of the present invention, three-dimensional structure mapping and morphological processing operations are carried out on seismic constraint feature data, and then a high-dimensional space feature set for LightGBM algorithm training is constructed. The specific process includes technical steps such as cube unit generation, morphological filtering, and hole filling from bottom to top. In the initial stage, the seismic constraint feature data is mapped into a regular three-dimensional grid structure, and the continuous spatial volume data is discretized into equal-sized cube units through voxelization. Each unit corresponds to a spatial node, and internally carries seismic constraint feature vectors, including seismic reflection characteristics, structural attributes, and tensor-guided data, etc. This mapping process constructs a geological cube element structure by evenly dividing the spatial coordinate system and assigning the feature values of each channel in each voxel. On this basis, in order to remove the edge artifacts and discrete noise points caused by modeling, acquisition errors, or unstructured features in the cube structure, a node erosion operation is used to remove the first-order structural elements at the voxel boundary. A 3×3×3 cubic convolution kernel is commonly used for sliding processing, retaining the continuous structure main body and removing the single-pixel connection parts on the surface; and further combining connected component analysis and minimum volume threshold setting, the isolated voxel clusters with volumes smaller than the set threshold in the voxel field are cleared to obtain geological opening operation morphological data, effectively highlighting the main contour of the geological facies body. Then, for the incomplete regions such as cavities and gaps in the geological body that may exist due to the blurred transition between seismic phases or noise interference, a closing operation morphological technique is applied for internal filling processing. An erosion-dilation operation sequence is used to fill the non-connected cavities inside the facies body to ensure the coherence and integrity of the facies body structure, and geological closing operation morphological data is constructed. Finally, the results of the opening operation and the closing operation are used as independent morphological feature channels and fused with the original structure features to form a feature vector set and input into the LightGBM modeling process. Each channel feature is defined by column, and the input label data extracts the seismic phase category code based on regional annotation, and the gradient boosting tree training operation is performed. Through iterative optimization of the feature splitting gain, a seismic phase LightGBM model is obtained, so as to realize high-resolution classification modeling of complex geological structures.

[0046] Preferably, step S4 includes the following steps: Step S41: Perform 20% feature removal processing on the seismic phase LightGBM model according to the LightGBM feature importance formula, and perform parameter accuracy evaluation to obtain LightGBM model parameter evaluation data; Step S42: Calculate the intersection over union of the seismic phase LightGBM model based on the seismic model verification data. If the intersection over union and the LightGBM model parameter evaluation data are unqualified, perform model backtracking optimization to obtain a seismic phase segmentation body model; Step S43: Evaluate the seismic phase recognition accuracy of the seismic phase segmentation body model, and perform semantic segmentation integrity judgment to obtain a seismic phase recognition semantic segmentation report.

[0047] In the embodiments of the present invention, according to the feature importance calculation mechanism inside the LightGBM model, the contribution degree of each input feature dimension to the decrease of the loss function during the model training process is extracted. Common indicators include the number of splits (the frequency of the feature being used as a splitting node) and the Gain value (the information gain brought by the feature splitting). By sorting all features in descending order according to the Gain value and performing the process of removing the lowest 20% ranked features, redundant attribute interference is reduced and the generalization ability of the model is improved. Subsequently, five-fold cross-validation or an independent validation set is used to evaluate the accuracy of the key parameters of the model after removal, including core indicators such as the overall classification accuracy, the recall rate of each seismic facies, and the F1 score, to generate LightGBM model parameter evaluation data. In step S42, the spatial overlap calculation is performed between the prediction result of the model on the validation data and the true annotation result in the seismic model validation data. The intersection over union (IoU) index is used as the segmentation performance evaluation criterion, that is, the ratio of the intersection area to the union area of the predicted region and the true region. The IoU value is calculated independently for each type of seismic facies. If the IoU indicators of each type do not reach the preset threshold or it is found that the model has problems such as underfitting, overfitting, or classification bias by combining the parameter evaluation data at the same time, the model backtracking optimization process needs to be started. The model structure and input features are adjusted mainly by methods such as rolling back the feature selection strategy, resetting the LightGBM parameter combination (such as tree depth, number of leaves, learning rate), or expanding the training sample set until the model performance indicators stably meet the agreed standards, and finally a seismic facies segmentation model for subsequent deployment is formed. In step S43, after the segmentation model is constructed, accuracy evaluation and semantic level consistency analysis are carried out on its recognition results on the test data or unseen data. Multidimensional evaluation indicators such as the per-pixel classification accuracy, the phase body boundary matching accuracy, and the regional consistency index are used to compare the seismic facies type label map output by the model with the manually annotated map or the expert interpretation map point by point, and further check whether there are serious problems such as fractures, holes, drifts, or ambiguities in the recognition region. Whether the semantic segmentation result meets the integrity requirements is determined by means such as connected component analysis and morphological consistency measure, and finally a systematic seismic facies recognition semantic segmentation report is output.

[0048] Preferably, step S41 includes the following steps: Step S412: Perform the process of removing 20% of the features on the seismic facies LightGBM model according to the LightGBM feature importance formula to obtain seismic facies feature removal data; Step S412: Perform seismic reflection amplitude evaluation on the seismic facies feature removal data to obtain seismic facies seismic reflection amplitude evaluation data; perform layer velocity evaluation on the seismic facies feature removal data to obtain seismic facies layer velocity evaluation data; Step S413: Weightedly superimpose the seismic phase reflection amplitude evaluation data and the seismic phase layer velocity evaluation data to obtain the LightGBM model parameter evaluation data.

[0049] In the embodiment of the present invention, the feature importance scores inside LightGBM (such as Gini coefficient, information gain, number of feature splits, etc.) are used to sort each input feature, and the contribution degree of each feature to the model performance is calculated. The 20% features with lower importance scores are removed to reduce the interference of redundant features to the model and the risk of overfitting. This operation helps to reduce the feature dimension, improve the operation efficiency of the model, reduce the error of the model on new data at the same time, and generate the seismic phase feature removal data. Further evaluate the data after feature removal. First, evaluate the seismic phase reflection amplitude. This process measures the reflection characteristics of each seismic phase by analyzing the amplitude change in the seismic data, using the reflection coefficient and the relationship between amplitude and depth, and combining the seismic reflection waveform in the data to obtain the seismic phase reflection amplitude evaluation data; then, evaluate the layer velocity of the data after seismic phase feature removal, and accurately evaluate the layer velocity of different seismic phases by applying the velocity model to analyze the propagation velocity of the reflected wave and combining the interlayer velocity change to obtain the seismic phase layer velocity evaluation data. Weightedly superimpose the seismic phase reflection amplitude evaluation data and the seismic phase layer velocity evaluation data. This process synthesizes these two sets of evaluation data using the weighted average or multi-level fusion method, and reflects the importance of amplitude and velocity in the model training process through the preset weight value to obtain the final LightGBM model parameter evaluation data. These parameter evaluation data will provide necessary support for subsequent model optimization and seismic phase prediction, ensuring the classification accuracy and precision of the final model.

[0050] Preferably, the LightGBM feature importance formula in step S411 includes the following: ; ; where represents the th feature such as instantaneous phase, curvature, etc.; represents the total number of decision trees in the LightGBM model; represents the structure of the mth decision tree; represents the total number of leaf nodes of a single decision tree; represents the non-leaf node in the decision tree, the split node index; represents the reduction in loss after node t is split, the squared loss reduction value, which measures the contribution of this node to model optimization; represents the feature index used when node t is split, Indicates that node t uses the j-th feature for splitting; is an indicator function that takes 1 when node t splits using the j-th feature and 0 otherwise; represents the average contribution of the j-th feature across all decision trees, which is used for feature screening, such as removing the 20% features with low contribution; is the total contribution of feature in a single tree, which is obtained by accumulating the loss reduction at each splitting node.

[0051] In the embodiments of the present invention, based on the feature importance evaluation method in the LightGBM framework, feature screening and optimization are achieved by quantifying the contribution of each feature to the reduction of the model loss during the decision tree splitting process. Specifically, assume that the input data is a seismic attribute matrix containing multiple features ( is the number of samples, is the feature dimension). During the training process of the LightGBM model, LightGBM performs gradient boosting by constructing decision trees . During the generation of each tree , for non-leaf nodes (the total number is , is the number of leaf nodes), they split the feature , where is an indicator function that takes 1 when node uses feature for splitting and 0 otherwise. The global feature importance is obtained by averaging the contributions of feature across all trees, that is, . This calculation process is manifested at the data level as follows: First, during model training, record the splitting nodes of each feature in each tree and their corresponding loss reduction amounts; Second, accumulate and average the contributions of each feature across all trees to generate an importance ranking; Finally, screen high-importance features according to a preset threshold (such as removing the 20% features with low contribution) to optimize the model complexity and generalization ability. This method provides a data-driven feature selection basis for seismic facies identification tasks by accurately quantifying the physical role of features in the tree model, ensuring that the model focuses on geological attributes that have a significant impact on classification decisions.

[0052] Preferably, the LightGBM feature importance formula in step S4 includes the following: Step S421: Calculate the intersection over union (IoU) for the seismic facies LightGBM model based on the seismic model validation data to obtain the IoU data, where the calculation of the IoU is as follows: ; Among them, represents the optimized seismic facies segmentation result, such as a binary mask matrix, with values of 0 or 1; GT is the true seismic facies label manually annotated, and is a binary matrix of the same size as ; ∩ is the intersection operation of sets, representing the set of pixels that are both 1 in and GT; ∪ is the union operation of sets, representing the set of pixels where at least one of or GT is 1; | | is the cardinality of the set, that is, the number of elements in the set. For example, represents the total number of pixels predicted as 1 in the segmentation result; Step S422: If the intersection over union data and the LightGBM model parameter evaluation data are unqualified, perform model backtracking optimization to obtain the seismic facies segmentation volume model, where the unqualified judgment criteria are that the intersection over union data is less than 0.75 and the LightGBM model parameter evaluation data is less than 0.6.

[0053] In the embodiments of the present invention, this technical means quantitatively evaluates the segmentation result of the seismic facies LiqhtGBM model based on the intersection over union (IoU) index, and combines the threshold determination of the model parameter performance to implement model backtracking optimization. Specifically, the input data is the optimized seismic facies segmentation result binary mask matrix ( are the height, width, and time depth dimensions of the 3D seismic volume respectively) and the true label matrix manually annotated . First, calculate the intersection and union through per-pixel logical operations: the intersection is defined as the set of pixels where the corresponding positions in the two matrices are both 1, and its cardinality is obtained by traversing all pixels and counting the number of pixels that meet the conditions; the union is the set of pixels that are 1 in any matrix, and the cardinality is obtained by counting the total number of all 1 pixels minus the duplicate count in the intersection. If the IoU value is lower than the preset threshold of 0.75 or the model parameter evaluation index (such as feature importance weighted accuracy, regularization loss, etc.) is lower than 0.6, then trigger the backtracking optimization process: First, extract the feature contribution degree data from the LightGBM model snapshot saved during the training stage, and according to the feature importance formula LightGBM ( is the total number of decision trees, is the number of non-leaf nodes, Recalculate the contribution degree of each feature for the loss reduction of node splitting), and eliminate the 20% features with low contribution at the end to simplify the model complexity; secondly, adjust the hyperparameter combination (such as num_leaves, learning_rate) through Bayesian optimization or grid search, and retrain the model based on the updated feature set and parameters; finally, generate the optimized segmentation mask And calculate the IoU again until the threshold requirement is met. This process reduces the risk of model overfitting from the data level through iterative feature screening and parameter tuning, ensuring the spatial consistency between the segmentation result and the geological true label.

[0054] Therefore, from any perspective, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the application document are intended to be encompassed within the present invention.

[0055] The above are only specific embodiments of the present invention, enabling those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but rather to the broadest scope consistent with the principles and novel features invented herein.

Claims

1. A seismic facies recognition semantic segmentation method based on the LightGBM algorithm, characterized in that Including the following steps: Step S1: Perform multi-dimensional attribute preprocessing on the three-dimensional post-stack seismic volume to obtain a geological body standardized feature matrix; Step S2: Based on the geological body standardized feature matrix, perform regional annotation to obtain regional annotation data; divide the regional annotation data into a training-validation set to obtain seismic model training data and seismic model validation data; Step S3: Dynamically expand the seismic model training data and perform geological model constraints to obtain seismic constraint feature data; perform seismic morphological noise processing on the seismic constraint feature data and construct a LightGBM model to obtain a seismic facies LightGBM model; Step S4: According to the LightGBM feature importance formula, perform 20% feature elimination processing on the seismic facies LightGBM model and perform parameter accuracy evaluation to obtain LightGBM model parameter evaluation data; calculate the intersection over union (IoU) for the seismic facies LightGBM model based on the seismic model validation data. If the IoU data and the LightGBM model parameter evaluation data are unqualified, perform model backtracking optimization to obtain a seismic facies segmentation body model.

2. The seismic facies recognition semantic segmentation method based on the LightGBM algorithm according to claim 1, wherein Step S1 includes the following steps: Step S11: Deploy a broadband geophone array sensor in the work area to collect a three-dimensional post-stack seismic data volume with a time sampling interval of 1 ms and a signal-to-noise ratio greater than 20 dB to obtain an original seismic amplitude data volume; Step S12: Perform complementary amplitude equalization processing on the original seismic amplitude data volume and use inter-trace energy normalization to eliminate amplitude deviation to obtain an equalized seismic data volume; Step S13: Perform band consistency processing on the equalized seismic data volume with a passband of 5 - 80 Hz and a stopband attenuation greater than 40 dB to obtain a denoised seismic data volume; Step S14: Based on the denoised seismic data volume, perform multi-attribute feature extraction to obtain a geological body standardized feature matrix.

3. The seismic facies recognition semantic segmentation method based on the LightGBM algorithm according to claim 2, characterized in that, Step S14 includes the following steps: Step S141: Calculate instantaneous data according to the denoised seismic data volume with a time window length of 20 ms to obtain instantaneous attribute data; Step S142: Extract texture attributes based on the denoised seismic data volume with a window size of 15x15 to obtain texture attribute data; Step S143: Use the short-time Fourier transform to extract the main frequency amplitude of the denoised seismic data volume with a frequency band step of 10 Hz to obtain spectral attribute data; Step S144: Calculate the travel time gradient and curvature of the denoised seismic data volume and perform edge attribute enhancement processing to obtain structural attribute data; Step S145: Perform zero-mean normalization on the instantaneous attribute data, texture attribute data, spectral attribute data, and structural attribute data and perform feature matrix construction to obtain a geological body standardized feature matrix.

4. The seismic facies identification semantic segmentation method based on the LightGBM algorithm according to claim 1, characterized in that, The dynamic expansion of the seismic model training data in Step S3 includes: Perform a 10° random rotation according to the seismic model training data to obtain rotation-enhanced feature data; Perform a 5-pixel translation according to the seismic model training data to obtain pixel translation-enhanced feature data; Add 25 dB Gaussian noise according to the seismic model training data to obtain noise addition-enhanced feature data; Merge the rotation enhancement feature data, pixel translation enhancement feature data, and noise addition enhancement feature data to obtain seismic model enhancement feature data.

5. The seismic facies recognition semantic segmentation method based on the LightGBM algorithm according to claim 1, characterized in that The geological model constraint in step S3 includes: Calculate the formation dip data based on the seismic model enhancement feature data; Calculate the formation azimuth data based on the seismic model enhancement feature data; Perform a travel tensor construction on the formation dip data and the formation azimuth data to obtain formation travel tensor data; Perform spatial position constraint alignment based on the formation travel tensor data to obtain seismic constraint feature data.

6. The seismic facies recognition semantic segmentation method based on the LightGBM algorithm according to claim 1, characterized in that The seismic morphological noise processing of the seismic constraint feature data in step S3 includes: Perform a cube structure mapping based on the seismic constraint feature data to obtain a geological cube element structure; Perform node erosion on the geological cube element structure and eliminate isolated noise points to obtain geological opening operation morphological data; Perform internal hole filling of the phase body on the geological cube element structure to obtain geological closing operation morphological data; Construct a LightGBM model using the geological opening operation morphological data and the geological closing operation morphological data to obtain a seismic facies LightGBM model.

7. The seismic facies recognition semantic segmentation method based on the LightGBM algorithm according to claim 1, characterized in that Step S4 includes the following steps: Step S41: Perform a 20% feature elimination process on the seismic facies LightGBM model according to the LightGBM feature importance formula and conduct parameter accuracy evaluation to obtain LightGBM model parameter evaluation data; Step S42: Calculate the intersection over union (IoU) of the seismic facies LightGBM model based on the seismic model validation data. If the IoU and the LightGBM model parameter evaluation data are unqualified, perform model backtracking optimization to obtain a seismic facies segmentation volume model; Step S43: Evaluate the seismic facies recognition accuracy of the seismic facies segmentation volume model and conduct a semantic segmentation integrity judgment to obtain a seismic facies recognition semantic segmentation report.

8. The seismic facies recognition semantic segmentation method based on the LightGBM algorithm according to claim 7, wherein, Step S41 includes the following steps: Step S411: Perform a 20% feature elimination process on the seismic facies LightGBM model according to the LightGBM feature importance formula to obtain seismic facies feature elimination data; Step S412: Conduct a seismic reflection amplitude evaluation on the seismic facies feature elimination data to obtain seismic facies reflection amplitude evaluation data; conduct a layer velocity evaluation on the seismic facies feature elimination data to obtain seismic facies layer velocity evaluation data; Step S413: Superimpose the weights of the seismic facies reflection amplitude evaluation data and the seismic facies layer velocity evaluation data to obtain LightGBM model parameter evaluation data.

9. The seismic facies recognition semantic segmentation method based on the LightGBM algorithm according to claim 8, characterized in that, The LightGBM feature importance formula in step S411 includes the following: ； ； Among them, represents the th feature such as instantaneous phase, curvature, etc.; represents the total number of decision trees in the LightGBM model; represents the structure of the mth decision tree; represents the total number of leaf nodes of a single decision tree; represents the non-leaf node in the decision tree, the split node index; represents the reduction in loss after splitting of node t, the squared loss reduction value, which measures the contribution of this node to model optimization; represents the feature index used when node t is split, represents that node t is split using the jth feature; is an indicator function, taking 1 when node t is split using the jth feature, and 0 otherwise; represents the th feature's average contribution degree in all decision trees, which is used for feature screening, such as removing the 20% features with low contribution; is the total contribution degree of feature in a single tree, which is obtained by accumulating the loss reduction amounts of each split node.

10. The seismic facies identification semantic segmentation method based on the LightGBM algorithm according to claim 7, characterized in that, Step S42 includes the following steps: Step S421: Calculate the intersection over union (IoU) of the seismic facies LightGBM model based on the seismic model validation data to obtain IoU data, where the calculation of the IoU is as follows: ； Among them, represents the optimized seismic facies segmentation result, such as a binary mask matrix, with values of 0 or 1; GT is the ground truth seismic facies label manually annotated, and is a binary matrix of the same size; ∩ is the intersection operation of sets, representing the set of pixels that are both 1 in and GT; ∪ is the union operation of sets, representing the set of pixels that are at least one 1 in or GT; | | is the cardinality of the set, that is, the number of elements in the set. For example, represents the total number of pixels predicted as 1 in the segmentation result. Step S422: If the IoU data and the LightGBM model parameter evaluation data are unqualified, perform model backtracking optimization to obtain a seismic facies segmentation volume model, where the unqualified judgment criteria are that the IoU data is less than 0.75 and the LightGBM model parameter evaluation data is less than 0.6.

Images