4.4 Article

Automatic Semivariogram Modeling by Convolutional Neural Network

期刊

MATHEMATICAL GEOSCIENCES
卷 54, 期 1, 页码 177-205

出版社

SPRINGER HEIDELBERG
DOI: 10.1007/s11004-021-09962-w

关键词

Automatic semivariogram modeling; Machine learning; Deep learning; Convolutional neural network

向作者/读者索取更多资源

Modeling semivariograms to characterize spatial continuity is subjective and noisy due to experimental variations. The proposed ASMC method uses deep learning to automate semivariogram modeling, improving objectivity and utilization of spatial data for a wide range of spatial modeling projects. The CNN-based approach successfully learns spatial characteristics and predicts semivariogram parameters with high accuracy, demonstrating the effectiveness of the machine learning workflow.
Modeling the semivariogram to characterize spatial continuity requires expert geostatistical knowledge and domain expertise about the spatial phenomenon of interest. Moreover, although practitioners may have experience in semivariograms, their interpretations may vary due to experimental semivariogram noise and ambiguity. In general, modeling semivariograms remains highly subjective. This paper presents a data-driven, deep learning-based automated semivariogram modeling method known as automatic semivariogram modeling with convolution-based deep learning (ASMC) that improves the utilization of available spatial information to reduce the subjectivity of semivariogram modeling. Training models are generated by sequential Gaussian simulation (SGS) and labeled with their associated semivariogram parameters (i.e., maximum correlation length, aspect ratio of major and minor direction ranges, and azimuth of major direction). ASMC consists of two convolutional neural networks (CNNs). The first CNN model maps the sparse spatial samples to the exhaustive SGS-derived spatial models, and the second CNN maps the SGS spatial model to the semivariogram parameters. Both CNNs are trained with realistic spatial training data, and their validity is also checked with validation data withheld from training. Two-dimensional synthetic, but realistic, case studies demonstrate that the first CNN successfully learns the spatial characteristics among spatial data and generates realistic subsurface model estimates. The second CNN learns the spatial context of the estimated subsurface model and successfully predicts the semivariogram parameters with greater than 96% accuracy. The proposed machine, deep learning-based workflow improves the utilization and objectivity of spatial data in semivariogram-based spatial continuity modeling. With the optimal design of the experiment for training and tuning of model hyperparameters, this method may be generalized for application for a wide range of spatial modeling projects.

作者

我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。

评论

主要评分

4.4
评分不足

次要评分

新颖性
-
重要性
-
科学严谨性
-
评价这篇论文

推荐

暂无数据
暂无数据