Automatic extraction of outcrop cavity based on a multiscale regional convolution neural network
Introduction
One of the important parameters in carbonate reservoir characterization and evaluation is determination of pore space characteristics (Yang et al., 2016; Schön, 2015). In reservoir evaluation, it is of great significance to quantitatively identify the distribution of fractures, pores in carbonate rocks and their internal geometric structure (Sun et al., 2012). Traditionally, the pore space characteristics of reservoirs are obtained through drilling and seismic data (Ping et al., 2013; Thiemeyer et al., 2014; Al Hinai et al., 2014). The conventional method of pore space extraction is thin-section observation (Verwer et al., 2011; Borazjani et al., 2016; Lai et al., 2018) or core CT scan (Tiwari et al., 2013, Xiaobo et al., 2020). This method has some limitations and technical bottlenecks in microscopic pore characterization. It can only be interpreted based on fine characterization of samples and inversion of spectral or laser intensity data, with relatively poor reliability. Recent techniques, such as the use of the current powerful processing mineralogy tool scanning electron microscopy (QEMSCAN), play an important role in pore space and mineral segmentation (Armitage et al., 2010; Amao et al., 2016). Compared with drilling and seismic data, outcrop data have more intuitiveness, integrity, accuracy and testability. Outcrop research is traditionally based on field investigations and manual drawings, which consume considerable manpower and time, and accuracy is not guaranteed. On the other hand, the latest digital outcrop technology based on laser scanning can realize fine, quantitative and simulated outcrop research. Therefore, automatic image pore recognition based on laser scanning digital outcrops can improve efficiency and facilitate information extraction and parameter characterization of quantitative pores. After the cave structure is identified, the geological characteristics of outcrops (i.e., structure, sedimentation and diagenesis) are combined to clarify the genetic types and development rules of pores and caves to provide a basis for reservoir prediction in the process of regional geological exploration. Therefore, how to use digital image recognition technology to complete the automatic recognition of cavities in outcrop areas is a subject of practical significance.
Based on the image segmentation technology proposed by Haralick and Shapiro in 1985 (Haralick and Shapiro, 1985), this technology has been applied in geoscience fields (Meyer and Beucher, 1990). Since then, image processing technology has begun to be applied to the automatic identification of fractures and cavities in logging imaging data (Li et al., 2005). Many scholars have widely used methods such as threshold segmentation (Ke and Xu, 2006; Cao, 2015; Xie, 2015; Tuan et al., 2020), edge extraction (Tian et al., 1999; Li, 2010) and machine learning (Shen and Gao, 2007; Qu et al., 2009; Shi, 2008; Chen et al., 2015; Wang et al., 2017; Gomes et al., 2016; Viana et al., 2016) to realize automatic identification of fractures and caves in logging imaging data. However, outcrops will be greatly affected by natural conditions such as illumination differences, rock fragmentation and weathering. Traditional automatic image recognition methods are inefficient, so artificial intelligence methods are mainly preferred in the identification of outcrop fractures and cavities (Su et al., 2005). In recent years, deep learning algorithms have emerged and achieved great success in the fields of computer vision (Simonyan and Zisserman, 2014), speech recognition (Hannun et al., 2014), semantic recognition (Parikh et al., 2016) and reinforcement learning (Mnih et al., 2015), which have promoted the rapid development of artificial intelligence. Deep learning also provides an effective tool for building a new, data-driven earth system science model (Reichstein et al., 2019), and has been successfully applied in image-based earth science-related problems (Koeshidayatullah et al., 2020; Dias et al., 2020; Saporetti et al., 2018). Among them, target recognition based on convolutional neural networks (CNN) has made a breakthrough in natural image applications. At present, target detection based on CNN includes two broad categories of methods (Kwok et al., 2018; Oliveira et al., 2019). The first category is based on regional recommendations. Its core idea is to determine the possible locations of the target in the image to be detected in advance through candidate regions. This approach mainly includes spatial pyramid pooling networks (SPP-NET) (He et al., 2015), fast region-based convolutional neural networks (Fast R-CNN) (Girshick, 2015), faster region-based convolutional neural networks (Faster R-CNN) (Ren et al., 2015), region-based fully convolutional networks (R-FCN) (Dai et al., 2016), mask region-based convolutional neural networks (Mask R-CNN) (He et al., 2017) and other algorithms. The second category is based on regression, such that a regression operation is carried out for the positioning and categorization of the target in a convolutional neural network. It mainly includes algorithms such as you only look once (YOLO) (Redmon et al., 2016) and single shot multibox detector (SSD) (Liu et al., 2016). On the whole, the first type of methods is more accurate, but the second type of methods is faster. We proposed an improved method named Mask R-CNN, which is a new method of cavity detection based on a multiscale regional convolutional neural network, and compared it with several existing common cavity recognition methods through experiments. Then, according to the cavity extraction results of this method, three cavity feature parameters, including the number of cavities, the surface porosity and the average cavity area, are calculated. The accuracy of cavity feature parameter extraction is evaluated by referring to the manual extraction results. Finally, by taking the digital outcrop profile of the Dengying Formation (2ndMember) in Xianfeng, Ebian, as an example, the method proposed in this paper is used to automatically identify the cavities in the outcrop profile, calculate the cavity parameters in layers, and quantitatively analyse their distribution characteristics.
Section snippets
Image segmentation algorithm
Traditional image segmentation methods, such as Otsu segmentation based on threshold detection (Otsu, 1979), watershed segmentation based on mathematical morphology and topology theory (Vincent and Soille, 1991), back propagation neural network trained by error back propagation algorithm (Rumelhart et al., 1986) and support vector machine method based on region segmentation (Cortes and Vapnik, 1995), use low-level semantics such as pixel colour, texture and shape to segment images with complex
Data preparation
The data are collected with high-definition photos of the Dengying Formation (2nd Member) in Xianfeng, Ebian, which are taken by a Pentax 645D high-resolution digital camera. The resolution of the image is 2 mm and the size is 800 × 800 mm. Then, we used the open-source image annotation tool VGG Image Annotator (Dutta and Zisserman, 2019) to create a sample dataset. We collect a total of 189 cavities as training samples. In addition, two regions with the same size of 800 × 800 mm are selected
Application of outcrop cavities in Dengying Formation (2nd member) in Xianfeng of Ebian
In this section, the multiscale regional convolutional neural network method is applied to the quantitative characterization of the extraction and distribution of outcrop cavities in the Dengying Formation (2nd Member) in Xianfeng, Ebian.
Results and discussion
In this section, we will use the multiscale regional convolutional neural network model designed in the previous section to conduct cavity extraction experiments on outcrop profile images. Furthermore, we will compare with traditional image recognition methods to verify the advantages of the model.
We presented the detection cavity results at different scales in two areas (Fig. 7). The resolution of the image is 2 mm and the size is 800 × 800 mm. Combined with the different number of detection
Conclusion
Data from the study of outcrops are the most intuitive, reliable and detailed resources in for sedimentary reservoir geology research. Based on high-precision profile image data from digital outcrops, a new method of cavity detection with a multiscale regional convolutional neural network was proposed, and used to automatically extract cavities and characterize the parameters of the Dengying Formation (2nd Member) in Xianfeng, Ebian. Through the research of this paper, the following
Computer code availability
Name of code: Automatic-Extraction-of-Outcrop-Cavity.
Contact details: School of Geosciences, Yangtze University, Wuhan, 430,100, China;
Email: [email protected].
Year first available: 2021.
Program language: Python.
Details on how to access the source code: https://github.com/irene-wsq/Automatic-Extraction-of-Outcrop-Cavity.
Authorship statement
Qing Wang: guided the algorithm design, designed the experiment and helped to organize the manuscript.
Siqi Wu: wrote the manuscript and performed the experimental analysis.
Qihong Zeng and Youyan Zhang: provided the data used in the experiment.
Yanlin Shao: obtained the input data and contributed in the code.
Fan Deng: obtained the input data and provided ideas for the experimental part.
Yuangang Liu: adjusted the experimental parameters and checked the manuscript.
Wei Wei: supervised the work and
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
The authors would like to thank the Remote Sensing Institute of Research Institute of Petroleum Exploration and Development (RIPED) in China for providing the awesome field outcrop cavity data set. We would like to appreciate the financial support from the National Natural Science Foundation of China (Grant No. 41701537) and science and technology research project of Education Department of Hubei Province (Grant No. B2021040). We would also like to thank the developers in the Tensorflow and
References (79)
- et al.
Diagenetic and sedimentary controls on porosity in Lower Carboniferous fine-grained lithologies, Krechba field, Algeria: a petrological study of a caprock to a carbon capture site[J]
Mar. Petrol. Geol.
(2010) - et al.
Two intelligent pattern recognition models for automatic identification of textural and pore space characteristics of the carbonate reservoir rocks using thin section images[J]
J. Nat. Gas Sci. Eng.
(2016) - et al.
Automatic detection of fractures and breakouts patterns in acoustic borehole image logs using fast-region convolutional neural networks[J]
J. Petrol. Sci. Eng.
(2020) - et al.
An algorithm for automatic detection and orientation estimation of planar structures in LiDAR-scanned outcrops
Comput. Geosci.
(2016) - et al.
Image segmentation techniques[J]
Comput. Vis. Graph Image Process
(1985) - et al.
Segmentation of digital rock images using deep convolutional autoencoder networks[J]
Comput. Geosci.
(2019) - et al.
Fully automated carbonate petrography using deep convolutional neural networks[J]
Mar. Petrol. Geol.
(2020) - et al.
Investigation of pore structure and petrophysical property in tight sandstones[J]
Mar. Petrol. Geol.
(2018) - et al.
Morphological segmentation
J. Vis. Commun. Image Represent.
(1990) - et al.
Machine learning approaches for petrographic classification of carbonate-siliciclastic rocks using well logs and textural information[J]
J. Appl. Geophys.
(2018)
Application of deep learning for semantic segmentation of sandstone thin sections[J]
Comput. Geosci.
Superiorities of Support Vector Machine in Fracture Prediction and Gassiness Evaluation
Characterization of oil shale pore structure before and after pyrolysis by using X-ray micro CT[J]
Fuel
A region-growing approach for automatic outcrop fracture extraction from a three-dimensional point cloud
Comput. Geosci.
Deep learning of rock images for intelligent lithology identification[J]
Comput. Geosci.
Comparisons of pore size distribution: a case from the Western Australian gas shale formations[J]
Journal of Unconventional Oil and Gas Resources
Automated mineralogical methodology to study carbonate grain microstructure: an example from oncoids[J]
Environ. Earth Sci.
A quantified study of segmentation techniques on synthetic geological XRM and FIB-SEM images[J]
Comput. Geosci.
The Acoustic Parameters Experimental Research of Fractured Rocks and the Log Evaluation of Fractured Reservoirs
Semantic Image Segmentation with Deep Convolutional Nets and Fully Connected crfs[J]
Logging identification method of fillings in fractures and caverns in carbonate reservoir based on BP neural network
Geophys. Prospect. Pet.
Encoder-decoder with atrous separable convolution for semantic image segmentation[C]
Stitcher: Feedback-Driven Data Provider for Object detection[J]
Support-vector networks
Mach. Learn.
Improving Accuracy of Automatic Fracture Detection in Borehole Images with Deep Learning and GPUs[C]//2017 30th SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI)
R-fcn: object detection via region-based fully convolutional networks
The VIA annotation software for images, audio and video
Fast RCNN
Rich Feature Hierarchies for Accurate Object Detection and Semantic segmentation[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition
Deep Speech: Scaling up End-To-End Speech recognition[J]
Spatial pyramid pooling in deep convolutional networks for visual recognition
IEEE Trans. Pattern Anal. Mach. Intell.
Deep residual learning for image recognition
Mask R-CNN
Adaptation in Natural and Artificial Systems: an Introductory Analysis with Applications to Biology, Control, and Artificial intelligence[M]
A method study on quantitatively evaluating sidewall imaging well logging data
Nat. Gas. Ind.
Deep learning approach for rock outcrops identification
Automatic Recognition and Quantitative Evaluation Study on Electric Image Logging
Automatic Recognition of Geological Aperture in Carbonates
The Identification of Slope Crack Based on Convolutional Neural Network[C]//International Conference on Artificial Intelligence and Security
Cited by (11)
Mineral identification based on natural feature-oriented image processing and multi-label image classification
2024, Expert Systems with ApplicationsScale-space effect and scale hybridization in image intelligent recognition of geological discontinuities on rock slopes
2024, Journal of Rock Mechanics and Geotechnical EngineeringDifferentiating outcropping strata from oblique photogrammetric data using an octree-based convolutional neural network with spatial parameters
2022, Computers and GeosciencesCitation Excerpt :Outcropping strata are often the preferred target of geological research because of their surface exposure. In traditional geological work, research on outcropping strata has mainly adopted the methods of artificial field geological surveys (Wu et al., 2022). Advancements in technical equipment, such as lidars, have provided high-precision data support for geological research work (Qiao et al., 2015; Liang et al., 2022).
Densely multiscale framework for segmentation of high resolution remote sensing imagery
2022, Computers and GeosciencesCitation Excerpt :The major limitation of using the logistic regression classifier is its assumption of linearity between the dependent and independent variables, making it difficult to capture complex relationships. Similarly, the region proposal networks (RPN) are also considered regardless of their structural complexity (Wu et al., 2022). Several approaches have shown the effectiveness of integrating dense convolution into networks to improve the discrimination ability.
Research framework for distributive fluvial system
2024, Lithologic ReservoirsEvaluation of Porous Media Using Digital Core Analysis by Pore Network Modeling Method: A Comprehensive Review
2023, Journal of Chemical and Petroleum Engineering