地铁沿线地质结构三维随机重建方法研究

doi:10.13745/j.esf.sf.2025.4.75

摘要/Abstract

摘要：

地铁是缓解大城市交通拥挤、增强城市综合承载能力和发展韧性的有效交通工具之一。高精度的三维地质模型是厘定地下空间的地质构造和不良地质体分布的重要数据基础,也是保证地铁工程建设安全的关键因素之一。然而,地铁工程地质数据整体量不多但局部密度高的特点,制约了地质体分布模式的有效识别和重建。本研究以广州地铁十一号线某区段为对象,针对白垩系、第四系沉积层及次火山岩复杂地质条件,系统对比了随机森林(RF)、XGBoost以及融合深度学习与多点统计学(DL+MPS)3种建模方法的性能。结果表明:DL+MPS方法通过深度神经网络提取全局特征,且与MPS局部优化相结合,在准确率(99.16%)、F1分数(98.91%)和ROC曲线AUC值(0.93~0.99)等关键指标上表现最优,能准确重建断层破碎带与火成岩体的空间接触关系,避免出现地层异常延伸和地质语义错乱现象。相较之下,随机森林和XGBoost虽在模型拟合阶段表现出较高训练精度(准确率分别达到99.60%和98.64%),但其模拟结果存在地质体离散分布、不合理外推及地层穿插等问题,钻孔验证准确率(最低为69.93%)显著低于DL+MPS方法(73.33%~87.50%)。研究表明:深度学习模型凭借强大的非线性特征提取能力,能有效应对地铁工程数据空间分布不均的挑战,为复杂地质条件下三维建模提供了更优解决方案,对提升地下工程安全性和数字孪生系统应用具有重要实践价值。

关键词: 三维重构, 地铁线路, 多点统计学, 深度学习

Abstract:

Constructing the metro system is one of the effective solutions to relieve traffic congestion in big cities and enhance the comprehensive carrying capacity and development resilience. High-precision three-dimensional (3D) geological model is an important data infrastructure for determining the geological structure and distribution of unfavorable geological bodies underground. And it is one of the keys to ensuring the safety of metro engineering construction. However, the characteristic that the overall amount of geological data of metro engineering is not large but the local density highly restricts the effective identification and reconstruction of the distribution patterns of geological bodies. Taking the geological structure at a station of Guangzhou Metro Line 11 as the concrete example, this study systematically compares the performance of three modeling methods: random forest (RF), XGBoost and hybrid method of deep learning and multi-point statistics (DL+MPS), under the complex geological conditions of Cretaceous strata, Quaternary sedimentary strata and subvolcanic rocks. The results illustrate that combining the advantages of simulating global features by deep neural network and the local optimization of MPS, the DL+MPS method shows the best performance in key indicators such as accuracy (99.16%), F1 score (98.91%) and AUC value of ROC curve (0.93-0.99). The DL+MPS method can accurately reconstruct the spatial relationship between fault fracture zone and igneous rock mass, and avoid abnormal extension of strata and geological semantic disorder. In contrast, although RF and XGBoost show high training accuracy in the model fitting stage with the accuracy rates of 99.60% and 98.64% respectively, some problems such as discrete distribution of geological bodies, unreasonable extrapolation and strata interpenetration appear in their simulation results. The minimum borehole dispersion value of the results by RF and XGBoost methods reaches 69.93%, is which is significantly lower than that of DL + MPS method (73.33%-87.50%). The research shows that the deep learning model can effectively deal with the challenge of uneven spatial distribution of subway engineering data by virtue of its strong nonlinear feature extraction ability, which provides a better solution for 3D modeling under complex geological conditions, and has important practical value for improving the safety of underground engineering and the application of digital twin system.

Key words: three-dimensional geological models, metro line, multipoint statistics, deep learning

中图分类号:

P628

陈勇华, 侯卫生, 郭清锋, 杨松桦, 叶舒婉, 李鑫. 地铁沿线地质结构三维随机重建方法研究[J]. 地学前缘, 2025, 32(4): 222-234.

CHEN Yonghua, HOU Weisheng, GUO Qingfeng, YANG Songhua, YE Shuwan, LI Xin. Study on stochastic reconstruction methods for 3D geological structures along metro lines[J]. Earth Science Frontiers, 2025, 32(4): 222-234.

图/表 13

图1 研究区基岩地质图

Fig.1 Bedrock map of the study area

图2 RF的分类器示意图

Fig.2 Schematic chart of classifier of RF

图3 改进后的DL+MPS建模方法流程图

Fig.3 The schematic chart of improved DL+MPS method for constructing 3D geological model

表1 地质面AFCDNN超参数设置

Table 1 AFCDNN hyperparameters for constructing geological surfaces

参数设置		搜索空间
超参数	隐藏层数	[1, 10]
	隐藏层神经元数	50n (n∈[1, 10])
	学习率	[10^-5, 10^-1]
	批量训练样本数量	2n (n∈[5, 10])
	网络训练周期	100n (n∈[1, 10])
评估指标	均方误差损失	最小化指标
搜索周期	30

图4 研究区钻孔和剖面分布

Fig.4 Distribution of boreholes and cross-sections in the study area

图5 采用不同方法模拟的地质体空间展布形态 a—断层破碎带和火成岩(DL+MPS);b—白垩系(DL+MPS);c—第四系(DL+MPS);d—断层破碎带和火成岩(RF);e—白垩系(RF);f—第四系(RF);g—断层破碎带和火成岩(XGBoost);h—白垩系(XGBoost);i—第四系(XGBoost)。

Fig.5 The spatial distribution patterns of geological bodies simulated using different methods

图6 采用不同方法模拟的断层破碎带形态对比 a—DL+MPS方法;b—RF方法;c—XGBoost方法。图中黑色的椭圆标识模型之间的差异。

Fig.6 Comparison of fault fracture zone morphology simulated using different methods

图7 采用不同方法模拟的切片对比(EW=180) a—DL+MPS方法;b—RF方法;c—XGBoost方法。图中黑色的椭圆标识结果之间的差异。

Fig.7 Comparison of slices simulated using different methods (EW=180)

表2 不同方法模型的评价指标值

Table 2 The values of evaluation indicators for 3D geological model by different methods

方法	准确率/%	精确率/%	召回率/%	F1分数/%
DL+MPS	99.16	98.92	98.91	98.91
RF	99.60	98.68	98.14	98.40
XGBoost	98.64	95.67	95.31	95.49

图8 采用不同方法的混淆矩阵 a—DL+MPS方法;b—RF方法;c—XGBoost方法。

Fig.8 Confusion matrix of 3D models by different methods

图9 不同方法的ROC曲线 a—DL+MPS方法;b—RF方法;c—XGBoost方法。

Fig.9 The ROC curves of different methods

表3 不同方法模拟结果钻孔验证准确率

Table 3 The borehole dispersion values of different method

钻孔编号	不同方法模拟结果钻孔验证准确率/%
钻孔编号	DL+MPS	RF	XGBoost
B1	79.17	81.25	77.08
B2	83.33	80.21	80.24
B3	73.33	69.93	72.27
B4	74.14	71.42	72.21
B5	87.50	89.09	83.52
B6	81.67	79.73	80.77
平均值	79.86	78.61	77.68

表4 不同方法的效率对比

Table 4 Efficiency comparison of different methods

方法	训练时间(相对值)	调参复杂度	特征工程需求
DL+MPS	高(需多轮迭代)	极高(层数、激活函数和优化器等)	高(需标准化和处理缺失值)
RF	低	低(主要调树数量和深度)	低(容忍缺失值和噪声)
XGBoost	中	中(学习率、树深度和正则化项)	中(需处理类别型特征)

参考文献 39

[1]	CHEN Z L, CHEN J Y, LIU H, et al. Present status and development trends of underground space in Chinese cities: evaluation and analysis[J]. Tunnelling and Underground Space Technology, 2018, 71: 253-270.
[2]	LIN D, BROERE W, CUI J Q. Metro systems and urban development: impacts and implications[J]. Tunnelling and Underground Space Technology, 2022, 125: 104509.
[3]	XIE H P, ZHANG Y H, CHEN Y Y, et al. A case study of development and utilization of urban underground space in Shenzhen and the Guangdong-Hong Kong-Macao Greater Bay Area[J]. Tunnelling and Underground Space Technology, 2021, 107: 103651.
[4]	黄超, 邱建, 谭晓红. 中心城区 “向地下要空间” 的城市遗产保护策略: 以徐州市为例[J]. 西南交通大学学报(社会科学版), 2023, 24(6): 99-117.
[5]	王成善, 周成虎, 彭建兵, 等. 论新时代我国城市地下空间高质量开发和可持续利用[J]. 地学前缘, 2019, 26(3): 1-8. DOI
[6]	HOU W S, CHEN Y H, LIU H G, et al. Reconstructing three-dimensional geological structures by the multiple-point statistics method coupled with a deep neural network: a case study of a metro station in Guangzhou, China[J]. Tunnelling and Underground Space Technology, 2023, 136: 105089.
[7]	BIAN X, FAN Z Y, LIU J X, et al. Regional 3D geological modeling along metro lines based on stacking ensemble model[J]. Underground Space, 2024, 18: 65-82.
[8]	HOULDING W S. 3D geoscience modeling: computer techniques for geological characterization[M]. Berlin: SpringerVerlag, 1994.
[9]	JIA Q R, LI W W, CHE D F. A triangulated irregular network constrained ordinary kriging method for three-dimensional modeling of faulted geological surfaces[J]. IEEE Access, 2020, 8: 85179-85189.
[10]	JESSELL M, AILLÈRES L, DE KEMP E, et al. Next generation three-dimensional geologic modeling and inversion[J]. Economic Geology, 2014, 18: 261-272.
[11]	OUYANG J, ZHOU C, LIU Z, et al. Triangulated irregular network-based probabilistic 3D geological modelling using Markov Chain and Monte Carlo simulation[J]. Engineering Geology, 2023, 320: 107131.
[12]	CAUMON G, GRAY G, ANTOINE C, et al. Three-dimensional implicit stratigraphic model building from remote sensing data on tetrahedral meshes: theory and application to a regional model of La popa basin, NE Mexico[J]. IEEE Transactions on Geoscience and Remote Sensing, 2013, 51(3): 1613-1621.
[13]	邹艳红, 黄望, 阳宽达, 等. 基于杨赤中推估法空间插值的三维地质隐式建模[J]. 地质学刊, 2017, 41(3): 384-393.
[14]	SMIRNOFF A, BOISVER E, PARADIS S J. Support vector machine for 3D modelling from sparse geological information of various origins[J]. Computers & Geosciences, 2008, 34(2): 127-143.
[15]	SHI C, WANG Y. Development of subsurface geological cross-section from limited site-specific boreholes and prior geological knowledge using iterative convolution XGBoost[J]. Journal of Geotechnical and Geoenvironmental Engineering, 2021, 147(9): 04021082.
[16]	CLAYTON C R. Managing geotechnical risk: improving productivity in UK building and construction[M]. London: Thomas Telford Ltd, 2001.
[17]	MADSEN R B, HØYER A-S, ANDERSEN L T, et al. Geology-driven modeling: a new probabilistic approach for incorporating uncertain geological interpretations in 3D geological modeling[J]. Engineering Geology, 2022, 309: 106833.
[18]	JIA R, LV Y K, WANG G W, et al. A stacking methodology of machine learning for 3D geological modeling with geological-geophysical datasets, Laochang Sn camp, Gejiu (China)[J]. Computers & Geosciences, 2021, 151: 104754.
[19]	WU S C, ZHANG J M, WANG R. Machine learning method for CPTu based 3D stratification of New Zealand geotechnical database sites[J]. Advanced Engineering Informatics, 2021, 50: 101397.
[20]	WANG X R. Uncertainty quantification and reduction in the characterization of subsurface stratigraphy using limited geotechnical investigation data[J]. Underground Space, 2020, 5(2): 125-143.
[21]	WANG Y, SHI C, LI X. Machine learning of geological details from borehole logs for development of high-resolution subsurface geological cross-section and geotechnical analysis[J]. Georisk: Assessment and Management of Risk for Engineered Systems and Geohazards, 2022, 16(1): 2-20.
[22]	WEI Y, XING Z, JIAN C et al. Use of tree-based machine learning methods for stratigraphic classification in 3D geological modelling[J]. IOP Conference Series: Earth and Environmental Science, 2021, 861(7): 072039.
[23]	GUARDIANO F B, SRIVASTAVA R M. Multivariate geostatistics: beyond bivariate Moments[M]// Geostatistics Tróia’92.Dordrecht: Springer Netherlands, 1993: 133-144.
[24]	CHEN Q Y, MARIETHOZ G, LIU G et al. Locality-based 3-D multiple-point statistics reconstruction using 2-D geological cross sections[J]. Hydrology and Earth System Sciences, 2018, 22(12): 6547-6566.
[25]	GUO J T, WANG Z X, LI C L, et al. Multiple-point geostatistics-based three-dimensional automatic geological modeling and uncertainty analysis for borehole data[J]. Natural Resources Research, 2022, 31(5): 2347-2367.
[26]	MARIETHOZ G, CAERS J. Multiple-point geostatistics[M]. Hoboken: Wiley, 2014,
[27]	HINTON G E, OSINDERO S, TEH Y W. A fast learning algorithm for deep belief nets[J]. Neural Computation, 2006, 18(7): 1527-1554. DOI PMID
[28]	LECUN Y, BENGIO Y, HINTON G. Deep learning[J]. Nature, 2015, 521(7553): 436-444.
[29]	郭甲腾, 刘寅贺, 韩英夫, 等. 基于机器学习的钻孔数据隐式三维地质建模方法[J]. 东北大学学报(自然科学版), 2019, 40(9): 1337-1342. DOI
[30]	ZHAO Z, ZHOU X P. An integrated method for 3D reconstruction model of porous geomaterials through 2D CT images[J]. Computers & Geosciences, 2019, 123: 83-94.
[31]	GOODFELLOW I, POUGET-ABADIE J, MIRZA M, et al. Generative adversarial nets[J]. Advances in Neural Information Processing Systems, 2014: 2672-2680.
[32]	LI Y, JIAN P P, HAN G H. Cascaded progressive generative adversarial networks for reconstructing three-dimensional grayscale core images from a single two-dimensional image[J]. Frontiers in Physics, 2022, 10: 716708.
[33]	YANG Z X, CHEN Q Y, CUI Z S, et al. Automatic reconstruction method of 3D geological models based on deep convolutional generative adversarial networks[J]. Computational Geosciences, 2022, 26(5): 1135-1150.
[34]	FAN W Y, LIU G, CHEN Q Y, et al. Geological model automatic reconstruction based on conditioning Wasserstein generative adversarial network with gradient penalty[J]. Earth Science Informatics, 2023, 16(3): 2825-2843.
[35]	BREIMAN L. Bagging predictors[J]. Machine Learning, 1996, 24(2): 123-140.
[36]	BREIMAN L. Random forests[J]. Machine Learning, 2001, 45(1): 5-32.
[37]	BELGIU M, DRĂGUŢ L. Random forest in remote sensing: a review of applications and future directions[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2016, 114: 24-31.
[38]	CHEN T, GUESTRIN C. XGBoost: a scalable tree boosting system[C]// Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. San Francisco California USA. ACM, 2016: 785-794.
[39]	SHI C, WANG Y. Data-driven construction of three-dimensional subsurface geological models from limited Site-specific boreholes and prior geological knowledge for underground digital twin[J]. Tunnelling and Underground Space Technology, 2022, 126: 104493.