3D classification of clay types using Machine Learning

Introduction

Sensitive glaciomarine clays, often referred to as ‘quick clay’, commonly occur in many countries at high, northerly latitudes, causing frequent and occasionally devastating landslides. The salt content of quick clay is strongly correlated to both its shear strength and electrical resistivity. Hence, it can be mapped using electromagnetic methods more efficiently than traditional intrusive methods, the latter of which can often be slow and costly. However, the resistivity signature of quick clay is non-unique, leading to ambiguous, imprecise interpretations of geophysical models. In this study, we present an improved method for predicting the probability of quick clay using airborne electromagnetics.

Using machine learning algorithms, we combine geophysical models with geotechnical data to address the issue of their non-unique resistivity signature. Beyond resistivity values, the machine learning algorithms use spatial derivatives of resistivity and spatial attributes. We evaluate the performance of this method using data collected from a road construction project in central Norway. Results show that this method is able to make plausible and accurate predictions of quick clay occurrence using as few as 10 boreholes across an area of 14.8 km², and that it outperforms a simple interpretation based on resistivity intervals alone.

In addition to a ‘best guess’ categorical classification, these algorithms output probability estimates, and we demonstrate that they are a reliable indication of uncertainty. The accuracy of these predictions also tends to increase as more geotechnical data are included as training data, helping compensate for the limited resolution of the airborne electromagnetics data. Given that the petrophysics of the clays at this test site are consistent with observations in other regions, we expect this method has the potential to make quick clay hazard mapping more efficient by offering valuable early-phase insights, leading to large time and cost savings for both infrastructure planning and regional hazard mapping.

‍

‍

Figure 1. Bird’s-eye view of 3D models showing voxels of predicted quick clay (red) and ambiguous zones (white) at Åsen resulting from different classifiers using different sets of boreholes.

Note that the ambiguous zones have 70% transparent and appear as light grey or light red depending on whether they occlude bedrock (dark grey) or brittle zones behind them. Voxels where non-brittle clay is predicted (Pbrittle < 0.4) are omitted.

‍

‍

‍

‍

Conclusion

We have demonstrated that our method, which combined airborne electromagnetics, geotechnical data and machine learning, is a reliable tool for performing large-scale, early-phase mapping of quick clay. Despite the resolution limitations of our resistivity models, we have demonstrated that brittle clays have a resistivity signature that is distinct enough from other types of sediment to be modelled with machine learning algorithms. The key is to incorporate spatial attributes to overcome the non-unique resistivity signature of leached glaciomarine clay and other sediments.

Though our predictions of quick clay probability do have uncertainty, the probability predictions associated with our models are an accurate indicator of the uncertainty in our predictions. With as few as 10 geotechnical soundings and an airborne electromagnetic (AEM) data set, we can delineate areas with a low or high probability of quick clay. With an increasing number of geotechnical soundings used as training, we can continually refine and update these predictions. Our workflow cannot replace detailed, high-resolution sampling of quick clay needed in some projects, but having an overview of quick clay zones early on may provide major cost savings for large infrastructure and construction project as well as large-scale hazard mapping. Given that the petrophysical properties of sediments at our test site in central Norway are similar to those elsewhere in areas prone to quick clay occurrence, we expect that our methods should be applicable elsewhere.

Acknowledgment

Thank you to Nye Veier for permission to publish these results, and specifically to Kari Charlotte Sellgren for her helpful comments and inputs on this publication. We also thank colleagues at NGI, particularly Katharina Kahrs, for assistance in preparing geotechnical data sets for this study.

Reference

The full paper can be requested through the download link above or found directly at www.earthdoc.org:

Christensen, C. W., Harrison, E. J., Pfaffhuber, A. A., & Lund, A. K. (2021). A machine learning–based approach to regional‐scale mapping of sensitive glaciomarine clay combining airborne electromagnetics and geotechnical data. Near Surface Geophysics, 19(5), 523-539.