Rapid characterisation of landslide heterogeneity using unsupervised classification of electrical resistivity and seismic refraction surveys

Abstract

The characterisation of the subsurface of a landslide is a critical step in developing ground models that inform planned mitigation measures, remediation works or future early-warning of instability. When a landslide failure may be imminent, the time pressures on producing such models may be great. Geoelectrical and seismic geophysical surveys are able to rapidly acquire volumetric data across large areas of the subsurface at the slope-scale. However, analysis of the individual model derived from each survey is typically undertaken in isolation, and a robust, accurate interpretation is highly dependent on the experience and skills of the operator. We demonstrate a machine learning process for constructing a rapid reconnaissance ground model, by integrating several sources of geophysical data in to a single ground model in a rapid and objective manner. Firstly, we use topographic data acquired by a UAV survey to co-locate three geophysical surveys of the Hollin Hill Landslide Observatory in the UK. The data are inverted using a joint 2D mesh, resulting in a set of co-located models of resistivity, P-wave velocity and S-wave velocity. Secondly, we analyse the relationships and trends present between the variables for each point in the mesh (resistivity, P-wave velocity, S-wave velocity, depth) to identify correlations. Thirdly, we use a Gaussian Mixture Model (GMM), a form of unsupervised machine learning, to classify the geophysical data into cluster groups with similar ranges and trends in measurements. The resulting model created from probabilistically assigning each subsurface point to a cluster group characterises the heterogeneity of landslide materials based on their geophysical properties, identifying the major subsurface discontinuities at the site. Finally, we compare the results of the cluster groups to intrusive borehole data, which show good agreement with the spatial variations in lithology. We demonstrate the applicability of integrated geophysical surveys coupled with simple unsupervised machine learning for producing rapid reconnaissance ground models in time-critical situations with minimal prior knowledge about the subsurface.