Mapping 3D Structure of Loose Quaternary Deposits Combining Deep Learning and Multiple-point Statistics: An example in Chencun, Northern Pearl River Delta
- 1School of Earth Sciences and Engineering, Sun Yat-Sen University, Guangzhou, 510275, China
- 2Guangdong Provincial Key Lab of Geodynamics and Geohazards, Guangzhou, 510275, China
- 3Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Zhuhai, 519080, China
- 4Guangdong Geology Survey Institute, Guangzhou, 510080, China
- 1School of Earth Sciences and Engineering, Sun Yat-Sen University, Guangzhou, 510275, China
- 2Guangdong Provincial Key Lab of Geodynamics and Geohazards, Guangzhou, 510275, China
- 3Southern Marine Science and Engineering Guangdong Laboratory (Zhuhai), Zhuhai, 519080, China
- 4Guangdong Geology Survey Institute, Guangzhou, 510080, China
Abstract. Reconstruction and cognition of structures of the Quaternary deposits, like thickness variation and displacement, is necessary for understanding neotectonics and the evolution of palaeo-valleys and deltas. Multiple-point statistics (MPS) is a useful method to reconstruct three-dimensional geological models in many fields. However, non-stationary spatial patterns and semantics in geological blocks are difficult to extract and reconstruct with the MPS-based methods, especially for those probability-based MPS methods. To reconstruct 3D characteristics of loose Quaternary deposits and the semantic relationship between them, an algorithm coupled MPS and deep artificial neural network (DANN) is proposed. The DANN is constructed and used to extract and simulate the global characteristics of geological structures. Process of sequential simulation and stratigraphic sequence calibration are implemented to build an initial model. To obtain a reasonable final realization, an iterative MPS simulation process with a multi-scale strategy is implemented. With several cross-sections and trench profiles used as modeling dataset, two concrete examples of constructing the Quaternary sediments in Chencun, South China are given. The displacements of sedimentary formation belonging to the Pleistocene reveal the strata rupture caused by the fault activities. The modeling results illustrated that the DANN used in the method can extract and simulate global structures of Quaternary deposits, and MPS simulation with the Expectation-Maximization-like iteration process can optimize local characteristics in results effectively.
- Preprint
(3321 KB) -
Supplement
(0.42 KB) - BibTeX
- EndNote
Weisheng Hou et al.
Status: open (until 07 Sep 2022)
-
RC1: 'Comment on gmd-2022-83', Anonymous Referee #1, 21 Jul 2022
reply
This paper gives a new solution for the model construction with DANN and MPS. The DANN extract global characteristics of geological structures. and the MPS simulation with Expectation-Maximization-like iteration process for local characteristics of the geobodies.The idea is interesting and may attract the focus of the modeller.Because in real reservoir,to guarantee the long range conenctivity and characteristics of small architectural elements is often a difficult task for modellers. The method will be adopted for modeling facies in oilfield.In my opinion the paper is excellent and should be accepted if some comments are addressed:
1.The cross section for global structure is parallel,why? is it sufficient for the reproduction of the characteristics .In another direction,the structure may be different, how to reproduce in 3D space with DANN? please clarify
2.The small structue is reproduced with the constrain of trenches to update the initial global structure. is it a 3D template for comparison.How to guarantee the consistency between the global stuctures and the local trenches. May be an illustration is needed.
3. In line 261-263. the method of constructing 3D models from 2D training image isstudiedï¼So maybe a revised depiction is suitable.
-
AC1: 'Reply on RC1', Henggaung Liu, 08 Aug 2022
reply
Thanks for the comments.
- The cross section for global structure is parallel,why? is it sufficient for the reproduction of the characteristics .In another direction,the structure may be different, how to reproduce in 3D space with DANN? please clarify
Response:
In many cases, 3D model is constructed with parallel cross-sections. Therefore, we provide a real case with parallel cross-sections in this study. However, as you pointed out, the parallel cross-sections cannot sufficient information to reproduce structures in the other direction. Therefore, as shown in the example in section 4.2, seven non-parallel trenches are used as data source for modeling. In the proposed method, whether the cross-sections used for modeling are parallel or not is not a key issue.
In this study, the modeling area is partitioned into regular grids. Then, the core of modeling is to specify the geological attribute on each grid. To simplify the problem, geological objects are formed with closed surfaces, which the same attribute appear in it. So, the problem is transferred into obtaining the elevation of subsurface as the interpolation process.
Here, the objective function of the DANN is constructed with the elevation of subsurface. Note that each object has two networks for top surface Mmax(i) and bottom surface Mmin(i). Values of the top surface and bottom surface are normalized and used as training dataset to learn the spatial distribution of geological surfaces. Here, coordinates (x, y) of the simulation grid are the input data, and the corresponding elevations hmax(x,y,Atti) and hmin(x,y,Atti) of each geological attribute are labeled for training Mmax(i) and Mmin(i).
The DANN is trained with the depths of contact lines from the cross-sections. Then, the elevations of the top and bottom surface of the Atti are predicted when the coordinates of grids to be simulated are input into the Mmax(i)and Mmin(i) . After completing the traversal, the top and bottom surface of each attribute can be obtained.
- The small structue is reproduced with the constrain of trenches to update the initial global structure. is it a 3D template for comparison.How to guarantee the consistency between the global stuctures and the local trenches. May be an illustration is needed.
Response:
The two models are independently built by the presented method, although they are in the same area (shown in the followed figure). The model with trenches are not constrained by the global model. Two trenches of BTI2 and BTI3 crosses the TI3 for the major model. In these two cross positions, the geological contacts are consistent in the trenches and parallel cross-sections. Because the TIs and BTIs are hard data for the modeling, the geometries of the two model in the cross positions are kept consistency.
- In line 261-263. the method of constructing 3D models from 2D training image isstudied,So maybe a revised depiction is suitable.
Response:
As pointed out in lines 261-263, the 2D TIs cannot provide information in third dimension. In this study, to obtain 3D patterns for modeling, we provided a compromise way by expanding 2D images to 3D data rather than directly extract 2D patterns for constructing 3D structures. The detail process of expansion is provided in lines 265-286. To express the process clearly, here, a simple process is given as follow:
- The first step is defining an expansion area Buffsec with or bigger than a template size in the simulation grid (SG), in which the 2D TI is in the middle or on the boundary. The position of Buffsecdepends on the location of 2D TI.
- A grid node neighboring to the known data or to the node with appointed data is randomly selected and marked as the current access grid nodeuc in the lateral layer. The attribute that appears with the maximum number in the moving window of 3×3 grid nodes of which the center node is uc is assigned to the attribute value of uc.
- Repeating step (3) until to values of all grids are assigned inBuffsec layer by layer, the 3D TD with a template size is obtained.
Note that the presented method is implemented with multiple scales. Therefore, in each scale, the 3D TD should be constructed before simulation.
-
AC1: 'Reply on RC1', Henggaung Liu, 08 Aug 2022
reply
-
RC2: 'Comment on gmd-2022-83', Anonymous Referee #2, 08 Aug 2022
reply
The paper combines two methodologies, MPS and DL, to create realistic geological models in a very specific case study.
A number of claims are made regarding literature that are not correct, plus the literature review is significantly lacking in major recent contributions in this area
- “the MPS method difficult to reconstruct global spatial features with anisotropic and non-stationary characteristics”. Many methods of non-stationary MPS exist, please consult the chapter “Non-stationary MPS” in the book of Mariethoz & Caers, 2015.
- “However, three-dimensional structure cannot be directly extracted from 2D cross-sections in the MPS-based simulation method”. Many methods exist that use 2D cross sections to create: e.g. Comunian, A., Renard, P. and Straubhaar, J., 2012. 3D multiple-point statistics simulation using 2D training images. Computers & Geosciences, 40, pp.49-65. This problem (of stereology) is quite common in 3D imaging (e.g. X-ray; MRI) and many methods of DL exists to do this. CNNs are very popular for this.
There is also no mention of the recent contribution of GAN methods starting with Laloy
Laloy, E., Hérault, R., Jacques, D. and Linde, N., 2018. Trainingâimage based geostatistical inversion using a spatial generative adversarial neural network. Water Resources Research, 54(1), pp.381-406.
Song, S., Mukerji, T., Hou, J., Zhang, D. and Lyu, X., 2022. GANSimâ3D for conditional geomodelling: theory and field application. Water Resources Research, p.e2021WR031865.
In addition, one would wonder if the case study specified would not be better solved with surface-based or implicit domain geological modeling. This literature is also missing:
Frank, T., Tertois, A.L. and Mallet, J.L., 2007. 3D-reconstruction of complex geological interfaces from irregularly distributed and noisy point data. Computers & Geosciences, 33(7), pp.932-943.
Consider an example with much larger complexity than presented in this manuscript:
Yang, L., Achtziger-ZupanÄiÄ, P. and Caers, J., 2021. 3D modeling of large-scale geological structures by linear combinations of implicit functions: Application to a large banded iron formation. Natural Resources Research, 30(5), pp.3139-3163.
Also consider rule-based geological modeling methods: Pyrcz, M.J., Sech, R.P., Covault, J.A., Willis, B.J., Sylvester, Z., Sun, T. and Garner, D., 2015. Stratigraphic rule-based reservoir modeling. Bulletin of Canadian Petroleum Geology, 63(4), pp.287-303.
The question for me is: what is the new contribution of this paper? What has been achieved that is new and hence exportable to other cases, applications? My take on this question is that the contribution remains narrow
- The methodology seems tailored to their specific case study. As a result, it contains many, many ad-hoc choices and tuning parameters that I would not know how they extend to other cases.
- The methodology is directly applied to the case study, there is no other verification, for example, we do not learn about how it would apply to some simpler synthetic models.
- Because of the many ad-hoc tuning, the methodology is very complex for what may be easier solved with other geological modeling approaches. There is small likelihood that others would use this method for that reason. The paper does not contain any comparisons, except for broad methological comparison which (see above) are not always accurate.
The manuscript is a significant amount of work, and a lot of thinking went into modeling this specific case study. But then, my question for the editor is if this is sufficient for publication in GMD where the aim would be to share modeling approaches across many applications.
Weisheng Hou et al.
Weisheng Hou et al.
Viewed
HTML | XML | Total | Supplement | BibTeX | EndNote | |
---|---|---|---|---|---|---|
173 | 39 | 8 | 220 | 11 | 3 | 1 |
- HTML: 173
- PDF: 39
- XML: 8
- Total: 220
- Supplement: 11
- BibTeX: 3
- EndNote: 1
Viewed (geographical distribution)
Country | # | Views | % |
---|
Total: | 0 |
HTML: | 0 |
PDF: | 0 |
XML: | 0 |
- 1