Method, device, system and storage medium for enhancing geoscientific table data based on deep learning

Abstract

The present invention belongs to the technical fields of resource exploration, geological mapping and environmental monitoring, and discloses a method, device, system and storage medium for enhancing geological table data based on deep learning. The method retains the physical properties and multi-scale correlations of continuous feature vectors in the data set as conditional inputs; and constructs a classification voter based on random forest, SVM and XGBoost to achieve the prediction and completion of discrete features, effectively addressing the problem of insufficient discrete labels for small samples. In order to systematically evaluate the performance of the model, a multi-dimensional evaluation system was also constructed. The experimental results using multiple sets of core analysis data as an example show that compared with the currently optimal CTGAN model, the data distribution generated by ICG-GAN is closer to the original data, and significant performance improvements are achieved in all six application indicators.

Description

Technical Field

[0001] The present invention relates to the technical fields of resource exploration, geological mapping and environmental monitoring, and in particular to a method, device, system and storage medium for enhancing geological table data based on deep learning. Background Art

[0002] The widespread application of artificial intelligence and big data technologies in fields such as mineral resource evaluation, oil and gas exploration, seismic data interpretation and imaging, and environmental monitoring has provided new analytical paradigms and technical approaches for geoscience research. However, due to constraints such as exploration costs and geological conditions, obtaining sample data with clear geological attributes or characteristics (i.e., labeled data) often requires manual labeling, which is inefficient and leads to a scarcity of labeled data, severely restricting the generalization ability and prediction accuracy of models. Therefore, how to effectively expand geoscience labeled datasets and improve the performance and stability of data-driven models has become a key scientific and engineering challenge that needs to be addressed urgently.

[0003] Geological data mainly includes two types: image data and tabular data. This paper focuses on the amplification of tabular data. Tabular data is usually stored in matrix form, with rows representing different samples and columns corresponding to various geological features. It includes both discrete features (such as lithology type and geological classification labels) and continuous features (such as physical parameters and geochemical element content). The notable characteristics of geological tabular data are that the sample size is usually small, and it has multi-scale characteristics and strong correlation (including spatial correlation and feature correlation).

[0004] In response to the scarcity of geological tabular data, existing technologies have proposed a variety of data augmentation strategies. For example, new samples are created near the center of a known mineral deposit through location perturbation; some technologies have proposed a window-based augmentation method (WBDA) to generate more geologically constrained label samples by gradually subdividing the initial window. In the field of seismic data, a translation-based augmentation strategy is adopted to generate additional seismic characteristics and porosity label data by moving a fixed-size sampling window along the well trajectory. Some scholars have proposed a random reduction and increase (RDI) algorithm, combined with a physical constraint closed loop (PC-CL) strategy, to improve the accuracy of seismic impedance inversion in a semi-supervised manner. These methods maintain the spatial distribution characteristics of geological data to a certain extent, but are still limited in improving sample diversity and have failed to effectively solve the problem of data imbalance.

[0005] To address data imbalance, existing technologies have introduced augmentation methods suitable for unbalanced datasets. For example, the ADASYN method was used to generate new synthetic samples near minority class samples, effectively balancing the lithologic distribution in drilling datasets. To address the overfitting problem in coal seam wettability prediction models, researchers used the SMOTE method to augment small, unbalanced datasets. However, these traditional data augmentation methods struggle to maintain strong correlations in geological data, and their interpolation strategies are prone to introducing noise, reducing the authenticity and applicability of the generated data.

[0006] With the continuous development of machine learning and deep learning technologies, various new data generation strategies are constantly emerging. For example, an improved variational autoencoder (VAE) model is used to model random noise in geophysical data, thereby generating simulated data with characteristics consistent with real noise. The Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) method, combined with a Pearson correlation coefficient screening strategy, constructs more representative training samples. Existing technologies have proposed the Geo-TabGAN (Geological Tabular Data Generative Adversarial Network) model for geoscientific tabular data augmentation. Through a local augmentation strategy, it effectively alleviates the problem of uneven distribution of generated labeled samples. These deep learning-based generative models demonstrate significant advantages in improving the flexibility and high fidelity of data augmentation, but they still face certain challenges when dealing with multi-scale features, especially mixed tabular data containing discrete and continuous variables.

[0007] With the advancement of machine learning and deep learning technologies, various new data generation strategies are constantly emerging. For example, an improved variational autoencoder (VAE) model is used to model random noise in geophysical data, generating simulated data with characteristics consistent with real noise. Based on the improved complete ensemble empirical mode decomposition (ICEEMDAN) method, combined with a Pearson correlation coefficient screening strategy, more representative training samples are constructed. Using StyleGAN2 and StyleGAN2-ADA, a controllable geophysical model generation method is constructed. The generated results cover complex geological structures (such as faults and sedimentary bedding), providing an effective data augmentation tool for geophysical inversion tasks. A semi-supervised seismic facies classification framework based on conditional GANs has been proposed. Using LoGANv2 to synthesize labeled seismic facies images, classification accuracy is significantly improved under small sample conditions, validating the practicality of generative adversarial networks for geoscientific image data augmentation. Some researchers have proposed the Geo-TabGAN model for geoscientific tabular data augmentation, and its local augmentation strategy effectively alleviates the problem of uneven distribution of generated labeled samples. These methods based on deep generative models have shown significant advantages in improving the flexibility and high fidelity of data augmentation, providing diverse solutions to the data scarcity problem in the field of geoscience. However, they still face challenges when dealing with multi-scale features, especially mixed tabular data containing discrete and continuous variables.

[0008] In recent years, generative adversarial networks (GANs) have demonstrated advantages in handling the relationship between discrete and continuous features. For example, by incorporating a classifier into discriminator training, they effectively address the generation of mixed tabular data. CTGANs, by applying conditional generative adversarial networks to augment tabular data, improve the realism of generated data. However, these methods still have limitations when applied directly to geoscientific data.

[0009] Geological tabular data do not have large-scale standardized data sets. Geological tabular data usually have the characteristics of small sample size, significant multi-scale characteristics, strong spatial correlation and correlation between features. Therefore, there is an urgent need to develop data generation strategies for geological tabular data. Summary of the Invention

[0010] The present invention is provided to solve the above-mentioned problems existing in the prior art. Therefore, a method, device, system and storage medium for enhancing geological table data based on deep learning are needed, which retains the physical properties and multi-scale correlations by taking the continuous feature vectors in the data set as conditional inputs; and constructs a classification voter based on random forest, SVM and XGBoost to achieve the prediction and completion of discrete features, effectively addressing the problem of insufficient discrete labels for small samples. In order to systematically evaluate the performance of the model, the present invention constructs a multi-dimensional evaluation system. The experimental results using multiple sets of core analysis data as an example show that compared with the current optimal CTGAN model, the data distribution generated by ICG-GAN is closer to the original data, and significant performance improvements are achieved in all six application indicators.

[0011] According to a first embodiment of the present invention, a method for enhancing geospatial data based on deep learning is provided, the method comprising:

[0012] Establish a data enhancement model; wherein, the data enhancement model includes a generator, a supervised classifier and a discriminator; the generator outputs simulated continuous feature data in response to an input random noise vector and conditional information, wherein the conditional information is a vector obtained by preprocessing the continuous variables of the geoscientific table data, and the continuous variables of the geoscientific table data include magnetic susceptibility, apparent resistivity and element content; the supervised classifier classifies the simulated continuous feature data to obtain corresponding category labels, and the simulated continuous feature data and its corresponding category labels are combined into sample pairs, and the discrete label completion process is realized by denormalizing to the original scale to obtain generated samples; the discriminator is used to output a probability estimate of the data source based on the real data sample or the generated sample and the conditional information;

[0013] Training the data enhancement model to obtain a trained data enhancement model;

[0014] A multi-dimensional evaluation system is constructed to comprehensively evaluate the practical utility of augmented data from two aspects: data similarity and downstream prediction task performance; wherein the augmented data is obtained based on the trained data augmentation model.

[0015] Furthermore, the supervised classifier includes a plurality of base classifiers, and the supervised classifier classifies the simulated continuous feature data to obtain corresponding category labels in the following manner:

[0016] Each base classifier outputs a probability distribution for all categories for the input simulated continuous feature data;

[0017] Based on the output of each base classifier, the class label is determined.

[0018] Furthermore, based on the output of each base classifier, the category label is determined by the following formula:

[0019] ;

[0020] Where, represents the category label, is the number of classifiers participating in the voting, Indicates the Classifiers for categories The predicted probability of X represents the input sample. k Indicates the number of categories.

[0021] Furthermore, when training the data augmentation model, a binary cross entropy loss function is used to model the adversarial target between the generator and the discriminator. The learning rate is decayed to 90% of the initial value after each set number of training rounds. In each mini-batch, the discriminator is updated first and then the generator is updated to maintain a dynamic balance of the training process. The weight decay parameter is set and L2 regularization is introduced to suppress overfitting of the discriminator. The loss value is recorded once after each set number of training rounds to monitor the training process.

[0022] Furthermore, the multi-dimensional evaluation system includes two dimensions: qualitative evaluation and quantitative evaluation.

[0023] Furthermore, the qualitative evaluation includes: visual analysis using single-feature distribution histograms and dual-feature group marginal plots; wherein the single-feature distribution histogram intuitively presents the similarities and differences in feature distribution of different data sets; the dual-feature group marginal plots include a central main plot, an upper box plot, and a normal distribution curve on the right. The central main plot displays the scatter distribution of two key features and their linear regression fitting curves, which are used to characterize the feature correlation and distribution patterns of different data sets; the upper box plot displays the distribution characteristics, dispersion degree, and potential outliers of the horizontal axis features, so as to quickly identify the median, interquartile range, and distribution symmetry of the data; the normal distribution curve on the right describes the central tendency and dispersion degree of the vertical axis features.

[0024] Furthermore, the quantitative evaluation includes: comparing the performance of the original dataset and the augmented dataset on multiple classifiers using multiple evaluation indicators; the multiple evaluation indicators include confusion matrix, accuracy, precision, recall rate, F1 score, Matthews correlation coefficient and area under the ROC curve.

[0025] According to a second technical solution of the present invention, a device for enhancing geospatial data based on deep learning is provided, the device comprising a processor configured to:

[0026] A model building module is configured to establish a data enhancement model; wherein the data enhancement model includes a generator, a supervised classifier, and a discriminator; the generator outputs simulated continuous feature data in response to an input random noise vector and conditional information, wherein the conditional information is a vector obtained by preprocessing the continuous variables of the geoscientific table data, and the continuous variables of the geoscientific table data include magnetic susceptibility, apparent resistivity, and element content; the supervised classifier classifies the simulated continuous feature data to obtain corresponding category labels, and the simulated continuous feature data and their corresponding category labels are combined into sample pairs, and the discrete label completion process is implemented by denormalizing to the original scale to obtain generated samples; the discriminator is used to output a probability estimate of the data source based on the real data sample or the generated sample and the conditional information;

[0027] A model training module is configured to train the data enhancement model to obtain a trained data enhancement model;

[0028] The model evaluation module is configured to construct a multi-dimensional evaluation system to comprehensively evaluate the actual utility of the augmented data from two aspects: data similarity and downstream prediction task performance; wherein the augmented data is obtained based on the trained data augmentation model.

[0029] According to the third technical solution of the present invention, a deep learning-based geological table data enhancement system is provided, the system comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the method described above.

[0030] According to a fourth technical solution of the present invention, a non-transitory computer-readable storage medium storing instructions is provided. When the instructions are executed by a processor, the method described above is executed.

[0031] The deep learning-based geospatial data enhancement method, device, system, and storage medium according to various solutions of the present invention have at least the following technical effects:

[0032] 1) This paper proposes a tabular data augmentation scheme based on an improved conditional generative adversarial network (ICG-GAN) for geoscientific data with small sample sizes and complex coupled features. Compared to existing tabular data generation models (such as CTGAN), ICG-GAN optimizes its network structure and conditional settings, making it more adaptable to the multi-source nature, feature correlations, and physical property constraints of geoscientific data. Validation experiments on core analysis data demonstrate that this method achieves statistical distributions closer to real data and significantly improves downstream prediction tasks using various classifiers, such as balanced random forests, random forests, support vector machines (SVMs), and XGBoost. This improvement is reflected not only in single metrics such as accuracy, precision, recall, F1 score, and Matthews correlation coefficient (MCC), but also in a comprehensive, multi-dimensional, qualitative and quantitative evaluation system.

[0033] 2) The advantage of ICG-GAN lies in its ability to effectively generate synthetic data with realistic statistical properties even with small sample sizes, leveraging the physical constraints and feature correlations inherent in geoscientific data. This property is crucial for geoscientific tasks with high data volumes and complex distributions, such as geological mapping, mineral resource prediction, and environmental monitoring. By increasing both the quantity and quality of data, ICG-GAN paves the way for building robust data-driven models with good generalization capabilities, enabling researchers to obtain reliable and practical predictions even in data-scarce environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 A network structure diagram of the data enhancement model ICG-GAN according to an embodiment of the present invention is shown;

[0035] Figure 2 A schematic diagram of a histogram comparison of each feature according to an embodiment of the present invention is shown, including the original dataset, the dataset generated by CTGAN, and the dataset generated by ICG-GAN;

[0036] Figure 3 Figure 1 shows group marginal plots for four feature pairs for three datasets according to an embodiment of the present invention: (a) and (b) show geophysical parameter feature pairs; (c) and (d) show geochemical parameter feature pairs. In each group marginal plot, the main plot is a scatter plot with two key features as the coordinate axes, overlaid with a linear regression fit curve. Above the main plot is a boxplot of the feature on the horizontal axis, and to the right is a probability distribution curve (normal distribution fit result) for the feature on the vertical axis.

[0037] Figure 4Figure 3 shows performance comparisons of generated data and original data using three mainstream classification models: (a1–a3) Balanced Random Forest, (b1–b3) Random Forest, (c1–c3) Support Vector Machine (SVM), and (d1–d3) XGBoost. In each group, the first data set is the original data, the second data set is generated using CTGAN, and the third data set is generated using the method proposed in this invention.

[0038] Figure 5 Radar charts according to an embodiment of the present invention are shown, showing the performance of three data sets under different classification models: (a) Balanced Random Forest; (b) Random Forest; (c) SVM; (d) XGBoost;

[0039] Figure 6 The ROC-AUC curves according to an embodiment of the present invention are shown, comparing the prediction performance of three data sets under four classification models. DETAILED DESCRIPTION

[0040] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] In recent years, the application of artificial intelligence in the Earth sciences has become increasingly widespread, but the scarcity of labeled data has severely hampered its effectiveness. While existing data augmentation methods based on conditional generative adversarial networks (CGANs) have been successfully applied in fields such as finance and medicine, they fail to fully consider the multi-scale characteristics and strong correlations of geoscience data, making them difficult to directly apply to the field.

[0042] Based on this, an embodiment of the present invention provides a deep learning-based geospatial data augmentation method. This method employs an improved conditional generative adversarial network (ICG-GAN) model specifically for augmenting geospatial data. The method includes three key design features: (1) using continuous features as conditional inputs to preserve the physical properties and inherent correlations of geospatial data; (2) introducing a classification voting mechanism based on the "integration concept" to generate more consistent discrete labels for each set of continuous features, thereby improving the label reliability of the generated data; and (3) constructing a multi-dimensional evaluation system to comprehensively evaluate the practical utility of augmented data from the perspectives of data similarity and downstream prediction task performance.

[0043] Specifically, the method comprises the following steps:

[0044] S100. Establish a data enhancement model; wherein, the data enhancement model includes a generator, a supervised classifier and a discriminator; the generator outputs simulated continuous feature data in response to an input random noise vector and conditional information, wherein the conditional information is a vector obtained by preprocessing the continuous variables of the geological table data, and the continuous variables of the geological table data include magnetic susceptibility, apparent resistivity and element content; the supervised classifier classifies the simulated continuous feature data to obtain corresponding category labels, and the simulated continuous feature data and its corresponding category labels are formed into sample pairs, and through denormalization to the original scale, a discrete label completion process is implemented to obtain generated samples; the discriminator is used to output a probability estimate of the data source based on the real data sample or the generated sample and the conditional information.

[0045] Geoscientific tabular data usually consists of two types of variables: continuous variables and discrete variables Among them, this embodiment will include samples, each sample has A dataset with features Defined as continuous variables; whereas information representing categories or labels is considered discrete variables .

[0046] Conditional Generative Adversarial Network (cGAN) is a type of model that introduces conditional information (such as labels, specific attributes, or contextual information) to guide the generation process based on the traditional GAN ​​(Generative Adversarial Network) framework. Generative Adversarial Networks (CGANs) consist of two parts: the generator and the discriminator.

[0047] Generator: Generator Receive noise vector ( is the noise dimension) and conditional information (i.e., the feature vector of the training data), and output the generated data The mathematical expression is:

[0048] ;

[0049] in, The generator parameters.

[0050] Discriminator: Discriminator Receive continuous variables or data generated by the generator and condition information , the probability estimate of the output data source . Its mathematical expression is:

[0051] ;

[0052] in, is the discriminator parameter. When the input is generated data , the discriminator will try to judge it as a false sample; when the input is real data , the discriminator tries to identify it as a real sample.

[0053] The training goal of CGAN is to make the samples generated by the generator look the same as real data to the discriminator, while the discriminator continuously improves its ability to distinguish between true and false. The optimization objective function of adversarial training is:

[0054] ;

[0055] in, It is a minimax game. G represents the generator, whose goal is to minimize the value function V; D represents the discriminator, whose goal is to maximize the value function V. The two are continuously optimized through adversarial training; E represents the mathematical expectation, which is used to average the values ​​of random variables according to their probability distribution. P data (X,C) represents a continuous variable and condition information The joint probability distribution of describes the probability pattern of the two in real scenarios; Denotes the discriminator D for continuous variables Combined condition information After that, the probability output is determined to be "real data" (the value is between 0 and 1, the closer to 1, the more the discriminator thinks it is real data), That is, taking the logarithm of the judgment probability, which is used to construct the objective function, so that the discriminator can judge the real data as true as much as possible; Represents the probability distribution of the noise vector Z, which is used to initialize the generation process of the generator; G ( Z , C ) indicates that the generator G receives the noise vector Z and conditional information After that, the generated samples; D ( G ( Z , C ), C ) represents the sample generated by the discriminator G ( Z , C ) Combined with condition information After that, the probability output is determined to be the real data. It means that the discriminator is asked to identify the generated samples as fake as much as possible (maximizing this expectation, the corresponding discriminator must distinguish between true and false; and the generator must minimize the part corresponding to this expectation, making it difficult for the discriminator to distinguish). Through this confrontation, the generator is pushed to generate more and more realistic samples.

[0056] Although conditional generative adversarial networks (CGANs) were initially applied primarily to image data generation, their theoretical framework suggests that this generative mechanism can be extended to augmenting tabular data. This paper, through three key innovations: optimizing the model structures of the generator and discriminator, improving the representation of conditional information, and proposing a discrete feature completion strategy, enables CGANs to effectively handle the generation of structured tabular data in geoscience.

[0057] The structure of the improved conditional generative adversarial network (i.e., data enhancement model ICG-GAN) is as follows: Figure 1 As shown, mainly by the generator (Generator) and Discriminator The generator consists of four fully connected layers. Its input is the concatenation of 100-dimensional random noise and conditional information. The input passes through hidden layers of dimensions 128, 256, and 512, respectively. ReLU is used as the activation function, and the final output is a synthetic sample with the same dimensions as the conditional information. The discriminator has a similar structure, also consisting of four fully connected layers. Its input is the concatenation of sample data and conditional information. The hidden layers have dimensions 512, 256, and 128, respectively. LeakyReLU activation function (negative slope of 0.2) is used, and the final output is a Sigmoid function, which determines the probability of the sample being real or generated.

[0058] Traditional conditional generative adversarial networks (cGANs) usually use discrete features as conditional information. However, due to the complex feature correlations in geological data, using discrete features as conditional inputs is difficult to fully express the intrinsic relationships between samples in a multi-scale and complex geological environment. Therefore, this embodiment uses the continuous variables of the geological data as inputs. As conditional information Input generation model. It should be noted that continuous variables Rather than raw continuous variables (such as magnetic susceptibility, apparent resistivity, and elemental content), these variables are "conditional information" obtained through preprocessing operations such as outlier removal, normalization, or standardization. Using continuous features as conditional inputs aims to fully exploit the physical and chemical properties inherent in geoscience data and the high correlations between these properties, thereby improving the ability of generated samples to approximate the true distribution in multivariate space. The specific implementation process is as follows:

[0059] ;

[0060] This conditional setting enables the generator to refer to the characteristic distribution and correlation of the original data when generating new samples, thereby better preserving the physical feature constraints of the geoscientific data. Of course, in order to more effectively maintain the physical feature constraints of the geoscientific data, the present invention introduces regularization strategies such as batch normalization and weight decay into the generator network to prevent abnormal deviations in the generated data during the training process.

[0061] Tabular data usually contains columns of mixed types, that is, there are both discrete variables and continuous variables. This mixed characteristic makes it difficult for the generative adversarial network to capture their respective distribution patterns and the correlation between variables at the same time, thus affecting the quality and consistency of the generated data. In addition, in the scenario of small sample label data, if the number of samples of a certain category in the discrete column is too small, the traditional cGAN model that uses discrete variables as conditional input is difficult to effectively learn the statistical characteristics of the category. The present invention uses a classification voting mechanism to complete the discrete labels of the continuous feature data generated by cGAN. This process is essentially to use a supervised classifier constructed from real samples to infer the "latent category" of the generated samples. Specifically, the generator Receive a random noise vector With a conditional message ( is the continuous feature of the real sample), output the simulated continuous feature data , the calculation process is as follows:

[0062] ;

[0063] Since the generated data only contains Lack of corresponding category labels In this embodiment, a supervised classifier is used to predict the label. In the classifier design, the soft voting strategy in ensemble learning is adopted to perform weighted average of the prediction results of multiple base classifiers according to the category probability, and finally generate the category label. Assume that each base classifier For input samples, i.e. continuous variables Output a probability distribution over all classes:

[0064] ;

[0065] Then the final prediction result of the voting classifier for the sample can be expressed as:

[0066] ;

[0067] in, is the number of classifiers participating in the voting, Indicates the Classifiers for categories In this embodiment, three commonly used supervised learning models are selected as base classifiers: Random Forest (RF), Support Vector Machine (SVM) and XGBoost, and they are considered equally important, that is, they have the same voting weights.

[0068] After the voting classifier is trained, the present invention applies it to the continuous feature samples generated by the generator, that is, to generate data . Used to predict its corresponding category label . Thus, a generated sample pair with a complete structure (continuous features + discrete labels) is constructed , the calculation process is as shown in formula (8). This method not only maintains the diversity of generated samples in the feature space, but also combines the discriminative ability of the original label information, significantly improving the usability of generated data.

[0069] ;

[0070] Where, h Represents a trained classifier.

[0071] Finally, the continuous feature part of the generated sample will be denormalized to the original scale according to formula (9), and the discrete label completion process will be completed.

[0072] ;

[0073] Where InverseTransformer represents the inverse transformation function / operation, Represents data normalized to the original scale.

[0074] S200: Train the data enhancement model to obtain a trained data enhancement model.

[0075] During training, ICG-GAN uses the Binary Cross Entropy (BCE) loss function to model the adversarial objective between the generator and the discriminator. BCE is widely used in binary classification tasks and effectively measures the difference in discrimination probability between generated and real samples, helping to accelerate the convergence of the discriminator and improve the stability of generator training. For the optimizer, the Adam optimizer with an adaptive learning rate mechanism is used. This method exhibits good stability and convergence performance when dealing with non-convex optimization problems, and is therefore widely adopted in generative adversarial networks. However, in actual training, traditional GANs are still susceptible to problems such as unstable generated distributions and mode collapse.

[0076] In order to improve the training stability and efficiency in small sample scenarios, the present invention introduces the StepLR learning rate scheduling strategy, which decays the learning rate to 90% of the initial value every 500 epochs, so as to take into account both rapid convergence in the early stage of training and fine-tuning in the later stage. In each mini-batch, the discriminator is updated first, and then the generator is updated to maintain the dynamic balance of the training process; in addition, the weight decay parameter is set to 1e-4, and L2 regularization is introduced to suppress overfitting of the discriminator. The mini-batch size of the network is set to 64, and the loss value is recorded once every 500 epochs to monitor the training process. Through the coordinated optimization of network structure and training strategy, ICG-GAN demonstrates good training stability and high-quality sample diversity in the task of generating geoscientific tabular data, alleviating the common training instability and mode collapse problems of traditional conditional generative adversarial networks in small sample geoscientific data.

[0077] S300. Construct a multi-dimensional evaluation system to comprehensively evaluate the practical utility of the augmented data from two aspects: data similarity and downstream prediction task performance; wherein the augmented data is obtained based on the trained data augmentation model.

[0078] This embodiment constructs a complete evaluation system to comprehensively evaluate the quality of generated data from both qualitative and quantitative dimensions (see Table 1).

[0079] Quantitative evaluation verifies the effectiveness of data augmentation methods from the perspective of application, primarily by comparing the performance of the original and augmented datasets on multiple classifiers. This study selected four machine learning classifiers with extensive application in geoscience: Balanced Random Forest (BRF), Random Forest (RF), Support Vector Machine (SVM), and XGBoost. Each classifier has its own strengths: BRF effectively addresses class imbalance through balanced sampling; RF offers strong feature selection and noise immunity; SVM is adaptable to high-dimensional feature spaces and exhibits excellent generalization performance; and XGBoost, combining the strengths of decision trees and gradient boosting, has achieved outstanding performance in numerous machine learning competitions. To comprehensively evaluate model performance, the six evaluation metrics listed in Table 1 were used. These metrics range from [0 to 1], with values ​​closer to 1 indicating better predictive performance. This comprehensive comparison of multiple metrics allows for a comprehensive assessment of the impact of data augmentation on classification performance. Furthermore, k-fold cross-validation was used for model training and testing, effectively reducing the randomness of the evaluation results.

[0080] Table 1 Evaluation metrics for generated data

[0081]

[0082] The following embodiments of the present invention will be combined with a specific implementation case to fully illustrate the feasibility and progress of the present invention.

[0083] This example uses two sets of borehole data, both collected from different locations within a mining area. The first set of data comes from two deep scientific research boreholes, KY15-03-01 and KY14-02-01, located between the main and eastern mining areas, respectively. The drilling depths are 1762.75 m and 2022.18 m, respectively, and contain a total of nine characteristic variables. The second set of data includes four boreholes located in different areas: TK13-4, TK8-2, TK13-3, and TK7-1, and contains a total of six characteristic variables. The physical properties and chemical element contents in the data used in this example are the result of comprehensive processing of multiple measurements. The data labeling information was jointly completed by experts in the fields of geology, well logging, and core analysis. To ensure traceability of the labeling work, the entire labeling process is recorded in detail, including time, work area, instrument model, measurement personnel, verification personnel, specimen number, drilling depth, core name, lithologic description reference, and labeling version, to facilitate subsequent verification and revision.

[0084] The first set of core analysis data comes from two scientific research deep boreholes (KY15-03-01 and KY14-02-01), collecting a total of 690 core samples. Each sample contains nine characteristic variables, including three geophysical parameters (magnetic susceptibility, polarizability, and resistivity) and six elemental contents (Mg, Al, Si, Fe, Ni, and Nb), along with lithologic labels. The sample distribution of lithologic types is as follows: slate (41.2%), (magnetic) mineralized dolomite (22.6%), magnetite (14.9%), dolomite (2.3%), quartz sandstone (4.2%), and dolomite-slate interbeds (14.8%). Selected statistical information is provided in Table 2. This dataset was selected for its coverage of typical geoscientific information, its multi-source, multi-dimensional, and multi-scale properties, and its integration of geophysical and geochemical characteristics, making it highly representative and applicable.

[0085] Table 2. Partial core analysis data set

[0086] magnetic susceptibility Polarizability Resistivity Mg Al Si Fe Ni Nb <![CDATA[Class Don't > 3.07 21.8285 122303.345 31824 83474 227998 67984 53 102 1 9.18 62.935 149.77 15868 82472 351379 25930 26 84 1 10.83 155.457 0.415 8801 74532 286398 37598 23 87 1 28.81 53.1315 1005.49 0 19586 456890 18853 15 7 1 10.61 72.337 77.085 84013 8743 35113 118374 18 14 2 80.89 52.6265 3405.36 67864 4129 21154 96099 35 175 2 67.05 57.1405 1820.89 80245 7155 25069 86097 32 284 2 38.46 44.7215 706.45 72617 5335 15095 102921 0 46 2 24009 124.3465 43.23 50336 13581 42638 121809 25 29 3 32801 59.988 0.85 74834 16785 52195 176772 0 152 3 57632 64.749 62.255 54971 13203 40801 182376 0 128 3 36368 121.825 16.815 26742 15355 50794 180548 75 185 3 7.32 20.2655 486.075 36416 6388 35097 188115 32 99 4 2.61 6.669 1268.965 81273 9649 26526 60407 23 26 4 0.87 13.204 5346.04 7862 7911 448785 2409 0 0 5 1.98 61.599 82.515 0 18107 486542 4222 10 7 5 1.06 30.6135 92.665 0 36625 418936 30182 83 172 5 837 32.647 66.15 51320 23906 81470 48334 34 394 6 259 18.0165 104.22 86948 1195 38500 55125 22 197 6 190 12.306 88.115 50791 17649 79167 46792 23 214 6 243 16.8575 1187.22 84261 24043 175799 67281 29 491 6

[0087] The experimental environment of this embodiment is configured as follows: the hardware platform is a Lenovo P920 graphics workstation equipped with an Intel Xeon Gold 6226R CPU (2.90GHz), 64GB of memory and an NVIDIA RTX A2000 GPU; the software environment is based on Python 3.8 and the PyTorch deep learning framework. The model uses the Adam optimizer, the initial learning rate of the generator and the discriminator are both set to 0.0002, the batch size is 128, the training rounds are 10,000, and the potential space dimension is 100. In order to improve the performance of the model, the StepLR learning rate scheduling strategy is introduced, and binary cross entropy is selected as the loss function. In terms of data preprocessing, the Min-Max normalization method is used to standardize the data, and the distribution consistency of the generated data and the original data is ensured by feature accuracy matching. In terms of data set division, the main experiment uses a 7:3 ratio of training set to test set.

[0088] Figure 2The distribution comparisons of nine key features (including magnetic susceptibility, polarizability, resistivity, and elemental contents (Mg, Al, Si, Fe, Ni, and Nb)) between real data, CTGAN-generated data, and the proposed ICG-GAN-generated data are presented. Overall, the ICG-GAN-generated data are closer to the real distribution in terms of morphology and range. For most features (such as magnetic susceptibility, polarizability, Mg, Al, Si, Fe, Ni, and Nb), ICG-GAN closely reproduces the skewness and kurtosis characteristics of the real data. For example, the tail behavior and overall morphology of elemental features such as Al, Fe, Ni, and Nb show higher consistency between ICG-GAN and the real data. In contrast, the CTGAN-generated data tend to be concentrated and deviate from the real distribution. In particular, for polarizability and the distribution of Mg and Al, the CTGAN data are overly concentrated in a small range of values. Furthermore, for some features (such as resistivity), the CTGAN-generated data exhibit unusual peaks and irrational distributions, while the ICG-GAN effectively avoids these. Overall, these comparative results show that ICG-GAN exhibits superior performance in terms of approximating data distribution and preserving the characteristics of real samples.

[0089] Figure 3 The four sets of marginal plots in the figure show the comparison of the original data (red), CTGAN generated data (blue) and ICG-GAN generated data (green) in two-dimensional statistics. Figure 3 The four sub-graphs (a), (b), (c), and (d) are analyzed:

[0090] Figure 3 (a) shows the marginal plots of magnetic susceptibility and resistivity. It can be seen that the lateral distribution range of the original data and the ICG-GAN-generated data is narrower than that of the CTGAN-generated data. The original data has several high-resistivity outliers in the low-susceptibility region, which is consistent with the ICG-GAN-generated data, while the outliers in the CTGAN-generated data are more dispersed. Linear regression fitting of the resistivity and magnetic susceptibility of different datasets shows that the correlation between the ICG-GAN-generated data and the original data is closer. The box plot above the main figure shows that the magnetic susceptibility distribution of the CTGAN-generated data is more discrete, with large differences and more outliers; while the ICG-GAN-generated data is more similar to the original data in terms of dispersion and concentration. The normal distribution plot on the right shows that the means of the three datasets are not much different, but the variance of both generated data is smaller than that of the original data.

[0091] Figure 3Panel (b) shows the correlation between magnetic susceptibility and polarizability. The original data exhibit a clear positive correlation trend for this feature combination, with some data points exhibiting a high degree of diffusion along the magnetic susceptibility axis, demonstrating a large linear expansion. In contrast, the data generated by CTGAN differ significantly from the original data, particularly in the high magnetic susceptibility and low polarizability regions, where points not present in the original data appear in the generated data, causing the regression curve of magnetic susceptibility and polarizability to deviate from the original relationship. In contrast, the data points generated by ICG-GAN are more densely distributed in the low magnetic susceptibility and low polarizability ranges, consistent with the mean and variation trend of the real data. A certain number of high polarizability points can also be reproduced in the high magnetic susceptibility range, thus maintaining the diversity and authenticity of the overall distribution. In terms of marginal distribution, the magnetic susceptibility box plot of the ICG-GAN-generated data shows a similar data concentration range and discrete points to the original data; the generated polarizability distribution curve maintains the overall shape of the original data. However, the polarizability distribution curve generated by CTGAN is obviously offset, and its magnetic susceptibility box plot shows a large number of abnormal points that do not match the real data. The length of the box plot also indicates that the generated data has a high dispersion and large differences between the data.

[0092] Figure 3 (c) shows the relationship between magnesium and silicon. The original data showed a clear negative correlation trend in this feature combination. However, the data generated by CTGAN differed significantly from the original data, especially in the high-silicon region, where discrete points that did not exist in the original data appeared, causing the regression curve of magnesium and silicon to deviate significantly from the original relationship. In addition, the magnetic susceptibility box plot generated by CTGAN showed that the range of values ​​differed greatly from the real data. In contrast, the concentrated distribution area of ​​the data scatter points generated by ICG-GAN was consistent with the real data, especially the discrete points in the high-magnesium and low-silicon region, and the regression curve trends of the two were consistent. In terms of marginal distribution, the concentrated range and discrete points of the magnetic susceptibility box plot of the ICG-GAN-generated data were similar to those of the original data.

[0093] Figure 3(d) shows the distribution relationship between iron and nickel content. The original data show that iron content covers a wide range of values, while nickel content exhibits some dispersion in the mid-to-high range. Overall, the data contain a small number of points with low iron (Fe) and high nickel (Ni) values, resulting in a more dispersed distribution. Compared with the original data, the data generated by CTGAN are shifted toward high nickel within the Fe-Ni plane, resulting in a regression relationship significantly higher than the original data. The box plot of Fe content shows that its value range is significantly higher than the original data and contains a large number of discrete points, indicating that CTGAN has introduced a significant amount of noise. However, the mean and variance of the normal distribution of nickel content are similar to those of the original data. In contrast, the data generated by ICG-GAN exhibit a more balanced distribution for this feature combination, with a similar range of Fe values ​​as the original data and consistent nickel dispersion across different Fe ranges. The box plot of Fe shows a distribution closer to the real data, and the regression line also tends to be consistent with the original data. However, in the nickel distribution plot, the variance of the dataset differs from that of the original data, primarily due to the data size.

[0094] Overall, Figure 3 The difference between CTGAN and ICG-GAN in their ability to fit multidimensional spatial data distributions is clearly demonstrated. The results show that ICG-GAN exhibits higher fidelity in reproducing the joint distribution of multidimensional features, marginal feature distributions, and regression relationships, approaching the statistical properties and intrinsic structure of the original data. Therefore, the proposed method is more suitable for augmenting geoscience data.

[0095] To verify the improved performance of the generated dataset, this example compared the performance of the original dataset, the dataset augmented using CTGAN, and the dataset augmented using the method of the present invention under different classifiers. This example randomly sampled 70% of the original dataset as the training set and 30% as the test set for prediction. The augmented training dataset was composed of the 70% of the original dataset combined with the generated dataset, while the test set remained the original 30%. Figure 4 The confusion matrices for four classifiers (Balanced Random Forest, Random Forest, Support Vector Machine, and XGBoost) are shown. Figure 4In (a)-(d), each corresponds to a classifier, where a is a balanced random forest, b is a random forest, c is a support vector machine, and d is XGBoost. Subscripts 1, 2, and 3 represent the confusion matrices of the original data, the data amplified based on CTGAN, and the data amplified based on the method of this study, respectively. Each column of the confusion matrix represents a predicted category, and the total number of columns is the number of data predicted to be in that category; each row represents a true category, and the total number of rows is the number of data instances in that category; the values ​​on the diagonal represent the number of true data predicted to be in that category. It can be seen from the confusion matrix that the model trained with the dataset amplified by ICG-GAN performs better than the other four classifiers. Subsequently, this embodiment further analyzes other classification indicators to more clearly and intuitively demonstrate the impact of generated data on the prediction model (see Figure 5 ).

[0096] Figure 5 The paper presents performance evaluation results of four classifiers—Balanced Random Forest (RF), Random Forest (RF), Support Vector Machine (SVM), and XGBoost—on different generated datasets (CTGAN, ICG-GAN, and original datasets). Experiments show that the classifier trained on the ICG-GAN dataset significantly improves performance across all evaluation metrics.

[0097] In quantitative analysis, the classification accuracy of Balanced RF on the ICG-GAN dataset reached 89.37%, which is 2.41% higher than that of CTGAN (86.96%) and 4.35% higher than that of the original dataset (85.02%). Figure 5 As shown in (a), the order of model performance improvement is ICG-GAN > CTGAN > original dataset. Compared with CTGAN, precision, recall, F1 score, and Matthews correlation coefficient (MCC) increased by 2.35%, 2.41%, 2.35%, and 5.30%, respectively. The RF classifier showed significant improvement on the ICG-GAN dataset, reaching a classification accuracy of 96.14%, an increase of 8.70% and 7.25% compared to CTGAN (87.44%) and the original dataset (88.89%), respectively. Figure 5As shown in (b), the data generated by CTGAN failed to improve model performance and instead resulted in a performance degradation. Analysis suggests that this phenomenon is primarily due to the fact that CTGAN uses discrete features as conditional inputs. These features are unevenly distributed in the original data, which can easily introduce generation bias during training, causing the distribution of generated samples to deviate from the real data. Furthermore, discrete features have significant limitations in expressing the complex feature correlations and physical properties found in geoscience data, making it difficult to fully capture the physical and chemical coupling between variables. These factors collectively affect the performance of CTGAN-generated samples in the lithology classification task. XGBoost achieved an accuracy of 94.69% on the ICG-GAN dataset, a 7.73% and 3.39% improvement over CTGAN (86.96%) and the original dataset (91.30%), respectively. Other evaluation metrics also showed significant advantages. In contrast, while SVM improved across various metrics on the ICG-GAN dataset, its overall performance remained inferior to both RF and XGBoost.

[0098] From an analysis of algorithmic characteristics, Random Forest (RF) and XGBoost performed exceptionally well on the ICG-GAN dataset, primarily due to their ability to handle complex feature interactions and nonlinear relationships. Balanced Random Forest (Balanced RF) effectively mitigated class imbalance through its sampling strategy. While Support Vector Machines (SVM) offer theoretical advantages in high-dimensional feature spaces, their performance is limited by kernel function selection and parameter optimization. Regardless of the classifier used, the ICG-GAN dataset significantly improved classifier performance, primarily due to its better preservation of the statistical characteristics and inter-class relationships of the original data during data generation. In summary, the experimental results fully demonstrate the superiority of ICG-GAN-generated datasets in improving classifier performance.

[0099] Figure 6 The training and testing results of four classifiers (balanced random forest, random forest, support vector machine, and XGBoost) for the original data, CTGAN-generated data, and ICG-GAN-generated data proposed in this study are shown, including a comparison of ROC curves and AUC values.

[0100] Analysis of the results shows that both the original data and the data generated using the proposed ICG-GAN (ICG-GAN) exhibit excellent classification performance in most cases. The AUC values ​​are generally high, with the curves closely aligned with the upper left corner, indicating a good balance between the true positive rate (TPR) and the false positive rate (FPR). Specifically, the AUC value for the original data using the Random Forest classifier was 0.9934, while that for the ICG-GAN-generated data increased to 0.9978, demonstrating significant improvement. This suggests that ICG-GAN not only approaches the real data distribution but also potentially provides a more stable decision boundary for the classifier. This trend is also evident with the support vector machine (SVM) and extreme gradient boosting (XGBoost) classifiers. Using the SVM classifier, the AUC for the ICG-GAN-generated data was 0.9879, surpassing the 0.9822 for the original data. Using the XGBoost classifier, the AUC for the ICG-GAN-generated data was 0.9980, further improving from the 0.9933 for the original data. Although the Balanced Random Forest classifier already achieved a high AUC (0.9842) on the original data, the ICG-GAN generated data still improved it to 0.9890, further highlighting the general performance gain of the ICG-GAN generated data.

[0101] Compared to the original data and ICG-GAN-generated data, the CTGAN-generated data performed poorly under all three classifiers in terms of Area Under Current Application (AUC). For example, in the Support Vector Machine (SVM) classifier, the AUC for the CTGAN-generated data was only 0.9687, significantly lower than both the original data and the ICG-GAN data. In other classifiers, such as Random Forest and XGBoost, while the AUC for the CTGAN-generated data remained relatively high (0.9927 for Random Forest and 0.9902 for XGBoost), it was consistently lower than that for the ICG-GAN and original data. Only in the Balanced Random Forest model did the AUC for the CTGAN-generated data slightly exceed that of the ICG-GAN, with an increase of only 0.14%.

[0102] Overall, ICG-GAN excels in data synthesis and augmentation, generating more realistic distribution features while also providing highly resolvable and stable training samples for mainstream classification algorithms. Across multiple model evaluation metrics, the classification performance of ICG-GAN-generated data matches or slightly exceeds that of the original data, significantly outperforming the commonly used generative model, CTGAN. This further validates the application value of ICG-GAN in data augmentation and improving classification performance.

[0103] The aforementioned implementation cases demonstrate that, compared to existing tabular data generation models (such as CTGAN), the proposed method, ICG-GAN, optimizes its network structure and conditional settings, making it more adaptable to the multi-source nature, feature correlations, and physical property constraints of geological data. Validation experiments using core analysis data demonstrate that the proposed method's statistical distribution is closer to real data and significantly improves downstream prediction tasks using various classifiers, such as balanced random forests, random forests, support vector machines (SVMs), and XGBoost. This improvement is reflected not only in single metrics such as accuracy, precision, recall, F1 score, and Matthews correlation coefficient (MCC), but also in a comprehensive, multi-dimensional evaluation system that combines qualitative and quantitative methods.

[0104] An embodiment of the present invention further provides a device for enhancing geospatial data based on deep learning, the device comprising:

[0105] A model building module is configured to establish a data enhancement model; wherein the data enhancement model includes a generator, a supervised classifier, and a discriminator; the generator outputs simulated continuous feature data in response to an input random noise vector and conditional information, wherein the conditional information is a vector obtained by preprocessing the continuous variables of the geoscientific table data, and the continuous variables of the geoscientific table data include magnetic susceptibility, apparent resistivity, and element content; the supervised classifier classifies the simulated continuous feature data to obtain corresponding category labels, and the simulated continuous feature data and their corresponding category labels are combined into sample pairs, and the discrete label completion process is implemented by denormalizing to the original scale to obtain generated samples; the discriminator is used to output a probability estimate of the data source based on the real data sample or the generated sample and the conditional information;

[0106] A model training module is configured to train the data enhancement model to obtain a trained data enhancement model;

[0107] The model evaluation module is configured to construct a multi-dimensional evaluation system to comprehensively evaluate the actual utility of the augmented data from two aspects: data similarity and downstream prediction task performance; wherein the augmented data is obtained based on the trained data augmentation model.

[0108] An embodiment of the present invention also provides a deep learning-based geological table data enhancement system, which includes: a memory for storing a computer program; and a processor for executing the computer program to implement the method described in the above embodiment.

[0109] An embodiment of the present invention further provides a non-transitory computer-readable storage medium storing instructions. When the instructions are executed by a processor, the method described in the above embodiment is executed.

[0110] The above embodiments are only used to illustrate the present invention, and are not intended to limit the present invention. Ordinary technicians in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also fall within the scope of the present invention. The scope of patent protection of the present invention should be defined by the claims.

Claims

1. A method for enhancing geospatial data based on deep learning, characterized in that: The method comprises: Establish a data enhancement model; wherein, the data enhancement model includes a generator, a supervised classifier and a discriminator; the generator outputs simulated continuous feature data in response to an input random noise vector and conditional information, wherein the conditional information is a vector obtained by preprocessing the continuous variables of the geoscientific table data, and the continuous variables of the geoscientific table data include magnetic susceptibility, apparent resistivity and element content; the supervised classifier classifies the simulated continuous feature data to obtain corresponding category labels, and the simulated continuous feature data and its corresponding category labels are combined into sample pairs, and the discrete label completion process is realized by denormalizing to the original scale to obtain generated samples; the discriminator is used to output a probability estimate of the data source based on the real data sample or the generated sample and the conditional information; Training the data enhancement model to obtain a trained data enhancement model; Constructing a multi-dimensional evaluation system to comprehensively evaluate the practical utility of augmented data from two aspects: data similarity and downstream prediction task performance; wherein the augmented data is obtained based on the trained data augmentation model; The supervised classifier includes a plurality of base classifiers. The supervised classifier classifies the simulated continuous feature data to obtain corresponding category labels in the following manner: Each base classifier outputs a probability distribution for all categories for the input simulated continuous feature data; Based on the output of each base classifier, determine the class label; Based on the output of each base classifier, the category label is determined by the following formula: ； Where, represents the category label, is the number of classifiers participating in the voting, Indicates the Classifiers for categories The predicted probability of X represents the input sample. k Indicates the number of categories.

2. The method according to claim 1, characterized in that When training the data augmentation model, a binary cross-entropy loss function is used to model the adversarial objective between the generator and the discriminator. The learning rate is decayed to 90% of the initial value after each set number of training rounds. In each mini-batch, the discriminator is updated first, followed by the generator, to maintain a dynamic balance in the training process. Set the weight decay parameter and introduce L2 regularization to suppress overfitting of the discriminator. Record the loss value after each set number of training rounds to monitor the training process.

3. The method according to claim 1, characterized in that The multi-dimensional evaluation system includes two dimensions: qualitative evaluation and quantitative evaluation.

4. The method according to claim 3, characterized in that The qualitative evaluation includes: visual analysis using single-feature distribution histograms and dual-feature group marginal plots; wherein the single-feature distribution histogram intuitively presents the similarities and differences in feature distribution of different data sets; the dual-feature group marginal plots include a central main plot, an upper box plot, and a normal distribution curve on the right. The central main plot displays the scatter distribution of two key features and their linear regression fitting curves, which are used to characterize the feature correlation and distribution patterns of different data sets; the upper box plot displays the distribution characteristics, dispersion degree, and potential outliers of the horizontal axis features to quickly identify the median, interquartile range, and distribution symmetry of the data; the right normal distribution curve describes the central tendency and dispersion degree of the vertical axis features.

5. The method according to claim 3, characterized in that The quantitative evaluation includes: comparing the performance of the original dataset and the augmented dataset on multiple classifiers using multiple evaluation indicators; the multiple evaluation indicators include confusion matrix, accuracy, precision, recall rate, F1 score, Matthews correlation coefficient and area under the ROC curve.

6. A device for enhancing geospatial data based on deep learning, characterized in that: The device comprises: A model building module is configured to establish a data enhancement model; wherein the data enhancement model includes a generator, a supervised classifier, and a discriminator; the generator outputs simulated continuous feature data in response to an input random noise vector and conditional information, wherein the conditional information is a vector obtained by preprocessing the continuous variables of the geoscientific table data, and the continuous variables of the geoscientific table data include magnetic susceptibility, apparent resistivity, and element content; the supervised classifier classifies the simulated continuous feature data to obtain corresponding category labels, and the simulated continuous feature data and their corresponding category labels are combined into sample pairs, and the discrete label completion process is implemented by denormalizing to the original scale to obtain generated samples; the discriminator is used to output a probability estimate of the data source based on the real data sample or the generated sample and the conditional information; A model training module is configured to train the data enhancement model to obtain a trained data enhancement model; a model evaluation module configured to construct a multi-dimensional evaluation system to comprehensively evaluate the practical utility of augmented data from two aspects: data similarity and downstream prediction task performance; wherein the augmented data is obtained based on the trained data augmentation model; The supervised classifier includes a plurality of base classifiers. The supervised classifier classifies the simulated continuous feature data to obtain corresponding category labels in the following manner: Each base classifier outputs a probability distribution for all categories for the input simulated continuous feature data; Based on the output of each base classifier, determine the class label; Based on the output of each base classifier, the category label is determined by the following formula: ； Where, represents the category label, is the number of classifiers participating in the voting, Indicates the Classifiers for categories The predicted probability of X represents the input sample. k Indicates the number of categories.

7. A deep learning-based geospatial data enhancement system, characterized by: The system comprises: memory for storing computer programs; A processor, configured to execute the computer program to implement the method according to any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium storing instructions, characterized in that: When the instructions are executed by a processor, the method according to any one of claims 1 to 5 is performed.

Images