Machine learning method for CPTu based 3D stratification of New Zealand geotechnical database sites

https://doi.org/10.1016/j.aei.2021.101397Get rights and content

Highlights

  • Random forest machine learning model for soil classification based on CPTu data.

  • Automatic classification for silty-sand based on fines content.

  • Modified WTMM layer boundary identification method considering the influence of transition zone.

  • GRNN based method for interpolation of stratification at discrete locations into 3D site.

  • High accuracy for 1D soil stratification and 3D site stratification in New Zealand.

Abstract

Three-dimensional (3D) geotechnical site stratification is of vital importance in geotechnical practice. In this study, a set of methods for 3D site stratification based on CPTu measurements of New Zealand Geotechnical Database (NZGD) sites is proposed. One-dimensional (1D) soil stratification at discrete CPTu points is first conducted and then interpolated in 3D to achieve 3D site stratification. 1D soil stratification is achieved through a proposed soil classification model combined with a proposed soil layer boundary identification method, which achieves a correct soil profile length identification rate of 93%. The soil classification machine learning model classifies the soil within NZGD into three types, i.e. Gravel, Sand, and Silt, and is able to reflect the fines content for silty sand. The model innovatively incorporates local variation information of CPTu curves in the input for a random forest algorithm to significantly improve identification accuracy to over 90%. Accurately locating soil layer boundaries is achieved through proposing a modified WTMM boundary identification method. 3D site stratification is then realized through 3D interpolation of 1D stratification at discrete CPTu points using a generalized regression neural network (GRNN) method. The 3D site stratification method is validated for two independent geotechnical sites within NZGD, exhibiting the effectiveness of the proposed set of methods.

Introduction

Three-dimensional (3D) soil stratification of geotechnical sites is the basis for design, construction, and maintenance of geotechnical engineering projects [30], [48]. 3D soil stratification is an important part of information models such as BIM (Building Information Model), especially for tunnelling and underground construction [20]. The piezocone penetration test (CPTu) has been a popular tool for in-situ soil investigation, with the benefits of being cost-effective, rapid, continuous, and reliable [26], [45], [46], [52], [22]. The efficient reconstruction of 3D soil stratum model of geotechnical sites based on CPTu without the additional need for soil sampling is a nontrivial task, which mainly consists of three processes: (1) soil type classification at discrete CPTu locations, (2) soil layer boundary identification at discrete CPTu locations, and (3) interpolation of soil types and layers from discrete points over the entire site.

Various existing methods for soil type classification via CPTu have been developed. Many use charts to link CPTu measurements (e.g., cone resistance (qc), sleeve friction (fs), and penetration pore pressure (u2)) with soil types [14], [44], [45], [24], [38], [15], [42], [49], [50], [31], such as the SBTn Qtn-Fr chart [45] and Qu2/σv0 chart [49]. A common drawback of these classification charts is that they are based on “soil behavior type”, which is inconsistent with physical characteristics-based soil classification systems commonly used in design, such as the Unified Soil Classification System (USCS) ([3], D2487-11). Several probabilistic soil classification approaches have been developed (e.g., [61], [51], [27], [9]), most of which are still based on the SBT/SBTn chart framework. The mapping between soil types and CPTu measurements can be very complex, which may not be sufficiently reflected using a two-dimensional chart in some cases. Machine learning methods can provide solutions for such complex nonlinear mapping. Generalized regression neural network (GRNN) has been used in soil classification based on CPTu [28]. Reale et al. [47] explored the link between cone penetration test (CPT) measurements and fines content (FC). A key issue that needs to be further resolved in adopting machine learning methods for soil type classification is to select and extract effective input features from CPTu curves to establish a strong mapping to soil.

Different soil properties of adjacent layers would result in transition zones for CPTu measurements near soil layer boundary. Various approaches have been developed for locating the soil layer boundaries using CPTu measurements, such as T ratio method [56], clustering method [18], [35], [8], modified Bartlett method [39], [40], wavelet transform modulus maxima (WTMM) method [13], and Bayesian methods [10], [53], [54], [11]. Zhao and Wang [62] proposed a method for identifying soil layer boundaries based on sparse data, which integrates Bayesian supervised learning with modified k-means clustering. Among these methods, WTMM method has demonstrated both accuracy and efficiency in identifying boundaries of soil layers. However, the WTMM method is known to suffer from having relatively low accuracy for boundary location when there is a long transition zone in the CPTu data [13]. In this study, WTMM method is modified and combined with the machine learning soil classification model to achieve 1D soil stratification at CPTu soundings.

Linear interpolation of discrete 1D soil stratification results is the simplest approach to further form a 2D site cross-section or 3D site model [37]. However, actual soil layers are often not linearly distributed. Ordinary Kriging method has been one of the most frequently used methods for interpolating spatial data [4], [55], [59], [41], [23], and can be applied for 2D or 3D site stratificationLi et al. [32], [12]. Stratification is described by the variogram and covariance functions in the ordinary Kriging method, which can be difficult to estimate based on inadequate CPTu/borehole data in practice [7]. Markov chain, a kind of statistical model, is popular in the problem of 2D or 3D soil stratification [16], [29], [33], [34]. The direction-dependent coupled Markov chain (CMC) method was proposed for improvement by presetting the simulation directions based on rational rules [34], but is relatively complicated for use in practice.

This study aims to automate the process of 3D soil stratification across a geotechnical site, via CPTu data. A new machine learning CPTu soil classification model compatible with the USCS is established, new features in CPTu data are extracted to enhance the performance of the model, and soil types according to physical characteristics are predicted. The WTMM method is modified and combined with the proposed soil classification model to achieve 1D soil stratification at CPTu points, achieving more accurate soil layer boundary detection. A GRNN interpolation method is proposed for 2D and 3D soil stratigraphy, the local mapping feature of the method allows for high accuracy estimation of soil categories. Compared with the direction-dependent CMC method, the GRNN method can achieve similar accuracy while being simpler and easier to use for engineering application. The entire set of 3D soil stratification methods proposed can benefit the effective and accurate setup of numerical analysis models of geotechnical sites based on CPTu measurements.

Section snippets

The study area and available data

CPTu and borehole data from NZGD are used in this study for the development and validation of procedures for 3D site stratification, respectively. CPTu and borehole soundings less than 2 m apart are considered as a pair and are assumed to have the same soil profile. In each pair, the borehole logs can be used as validation for CPTu based soil stratification method. In this study, 137 pairs of CPTu and borehole soundings are identified within NZGD and used for establishment and validation of the

Overall framework

To achieve 3D site stratification of geotechnical sites using CPTu data, two main procedures are needed: (1) use of information from CPTu soundings to obtain 1D soil stratification; (2) interpolation of 1D soil stratification at CPTu points to obtain 3D site stratification. 1D soil stratification comprises two sub-steps: (1) classification of soil types within layers; (2) identification of boundaries between different soil layers. The overall framework is shown in Fig. 2. In this study, a new

Validation of soil classification model

The proposed soil classification model is validated using two independent validation sets of CPTu and borehole pairs, from Christchurch and the North Island, respectively. SM data is not considered during this performance assessment, as there are not enough fines content measurements for the CPTu and borehole pairs. It is assumed that the 37% fines content threshold for Sand-like and Silt-like SM obtained based on available data is reliable.

The performance of the classification model is

Choice of features for soil classification model

Variation information of CPTu curves are extracted as input features to enhance the performance of soil classification model for the first time in this study, this is characterized by standard deviation (s(Qtn), s(Fr), s(Bq)) or local deviation (s*(Qtn), s*(Fr), s*(Bq)). Observation of CPTu results in different types of soil (gravel, sand, silt, etc.) show clear differences in the variation information, as shown in Fig. 3, which was hypothesized to be helpful for soil type identification. Fig. 3

Conclusions

A set of methods for 3D site stratification based on CPTu measurements of NZGD sites is proposed in this study. 1D soil stratification at discrete CPTu points is first conducted and then 3D interpolated to ultimately achieve 3D site stratification. 1D soil stratification is achieved through the proposed soil classification model and soil layer boundary identification method. 3D interpolation is achieved using GRNN.

A physical characteristics-based soil classification machine learning model is

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.

Acknowledgements

The authors would like to acknowledge the National Natural Science Foundation of China (No. 52022046 and No. 52038005) for funding the work presented in this paper.

References (62)

  • K.M. Yost et al.

    Assessment of the efficacies of correction procedures for multiple thin layer effects on Cone Penetration Tests

    Soil Dyn. Earthquake Eng.

    (2021)
  • T. Zhao et al.

    Interpolation and stratification of multilayer soil property profile from sparse measurements using machine learning methods

    Eng. Geol.

    (2020)
  • M.M. Ahmadi et al.

    Thin layer effects on the CPT qc measurement

    Can. Geotech. J.

    (2005)
  • Y. Amit et al.

    Shape quantization and recognition with randomized trees

    Neural Comput.

    (1997)
  • ASTM

    D2487–11. Standard Practice for Classification of Soils for Engineering Purposes (Unified Soil Classification System)

    (2011)
  • G.B. Baecher et al.

    Reliability and Statistics in Geotechnical Engineering

    (2003)
  • L. Breiman

    Random forests

    Mach. Learn.

    (2001)
  • R.W. Boulanger et al.

    Inverse filtering procedure to correct cone penetration data for thin-layer and transition effects

  • S.F. Carle

    Use of a transition probability/Markov approach to improve geostatistical of facies architecture

  • G.J. Cai et al.

    Soil classification using CPTU data based upon cluster analysis theory

    Chin. J. Geotech. Eng.neering

    (2009)
  • K.O. Cetin et al.

    CPT-based probabilistic soil characterization and classification

    J. Geotech. Geoenviron. Eng.

    (2009)
  • Z. Cao et al.

    Bayesian approach for probabilistic site characterization using cone penetration tests

    J. Geotech. Geo-environ. Eng.

    (2013)
  • Z.-J. Cao et al.

    Bayesian identification of soil stratigraphy based on soil behaviour type index

    Can. Geotech. J.

    (2019)
  • J. Ching et al.

    Cone penetration test (CPT)-based stratigraphic profiling using the wavelet transform modulus maxima method

    Can. Geotech. J.

    (2015)
  • B.J. Douglas, R.S. Olsen, Soil classification using electric cone penetrometer, in: Proceedings of the Symposium on...
  • A. Eslami et al.

    Pile capacity by direct CPT and CPTu methods applied to 102 case histories

    Can. Geotech. J.

    (1997)
  • A. Elfeki et al.

    A Markov chain model for subsurface characterization: theory and applications

    Math. Geol.

    (2001)
  • M.E. Harr

    Reliability-based Design in Civil Engineering

    (1987)
  • Y.A. Hegazy et al.

    Objective site characterization using clustering of piezocone data

    J. Geotech. Geo-environ. Eng.

    (2002)
  • T.K. Ho

    Random decision forests

    Proceedings of the 3rd International Conference on Document Analysis and Recognition, Montreal, QC

    (1995)
  • I.M. Idriss, R.W. Boulanger, Soil liquefaction during earthquakes. Earthquake Engineering Research Institute (EERI),...
  • Cited by (21)

    View all citing articles on Scopus
    View full text