Depth to bedrock accuracy: comparing ANN with interpretation techniques used in early time case studies

Introduction

Infrastructure cost overruns and delays are persistent challenges for engineers and project owners. Assessing geological risk is a significant part of planning; however, this risk is hard to control given the high cost of detailed ground investigation programs using traditional approaches (i.e. geotechnical drillings). Airborne geoscanning is a technology that is increasingly being used to mitigate geological uncertainty.

We have translated complicated geophysical models to parameters valuable for engineers using artificial neural networks. In this study, we illustrate the applicability of airborne geoscanning surveys to derive bedrock topography (i.e.depth of cover).

Conclusion

The most significant added value of helicopter geoscanning to an infrastructure project is provided when the survey is carried out in an early project phase. With only a very small number of control boreholes the presented algorithm creates a representative bedrock topography model. The more boreholes are being acquired, the more accurate the model becomes. Thus, a frequent update of the derived model is crucial to extract the full value of the survey investment.

The ANN provides an objective and consistent analysis of the models outperforming human subjectivity. The human expert can thus instead invest valuable time in quality controlling the ANN results. At a later project stage, when high spatial resolution and high accuracy is needed along the final infrastructure location, the geophysical method’s resolution sets limitations. The accuracy of the two bedrock modelling methods (borehole triangulation and AEM resistivity guided ANN) converge at around 20 boreholes per square kilometer (~200 m borehole spacing), which is just slightly more than the footprint of a single AEM sounding (~150 m, Figure 4 in the article).

However, there is a high variability in performance of neural networks. In the Nes study, the average error of a bedrock model ranged from 28% to 62% when only 5 boreholes were used to train a neural network. Performance is thus highly dependent on the input training data. Early indications are that the neural network performs best where a wide range of geological conditions, resistivity models, locations, and depths are represented in a sample. We have successfully applied cluster analysis to optimize a borehole investigation plan based on early AEM models. Further developments of ANN based geotechnical interpretation linking the 3D resistivity models to geotechnical soundings and samples have shown convincing results.

Acknowledgement

This study was carried out within the Norwegian Research Council’s FORNY2020 project “Airborne Geo-Intelligence”. Our project partners Bane NOR (Norwegian Railway Authorities) and Statens Vegvesen (National Public Road Authorities) have kindly granted permission to re-analyze and publish the presented data and have actively contributed in the assessment of new results. Our gratitude further extends to our colleagues M. Romøen, K. Kåsin and others within NGI and throughout the Norwegian geotechnical industry that have either inspired or contributed to aspects of the presented work.

References

The full paper can be requested through the download link above or found directly at www.app.box.com.

Pfaffhuber, A. A., Lysdahl, A. O., Christensen, C. W., Vöge, M., Kjennbakken, H., & Mykland, J. (2019). Extraction of depth of bedrock from airborn electromagnetic data using artificial neural networks. In 32nd Annual Symposium on the Application of Geophysics to Engineering and Environmental (SAGEEP 2019). Portland, Oregon USA.