Automated data processing of a large-scale airborne time-domain electromagnetic survey by a deep learning expert system

The new generation of airborne electromagnetic (AEM) surveys yield large data sets of thousands of line kilometres. Parts of these data are often contaminated by noise from various sources, e.g. fences, power lines, which corrupts the data to a degree that it can no longer be used. The problem intensifies in urban areas where the risk of data corruption is highest due to dense infrastructure. The inversion of corrupted data risks interpreting spurious subsurface features and flawed geological interpretations. Therefore, in many cases, the corrupted data is identified and culled prior to inversion. This process of culling corrupted data is generally a manual task requiring specialists to examine the data in detail, which is an extremely complex and time-consuming process. Recently, we proposed a deep learning expert system to automate the complex AEM data processing workflows. The proposed method uses a deep convolutional auto-encoder to identify corrupted data, and was trained such that it generalises to diverse geological conditions and various survey areas. In this study, we investigate the generalisation capabilities of our deep learning method on a large AEM survey area in Northland, New Zealand. Our approach takes ~ 600s to process 3984 line kilometres of data and displays strong spatial correlation for the data identified as corrupted. The inversion results show very few potential anomalies in the model space which are being inspected by a manual operator. In general, the proposed approach is generalisable and displays high-quality data processing within short amounts of time, which requires minimal further quality inspection.