Almost automatic geological mapping from AEM surveys

Interpretation of electromagnetic survey data for underlying geology is a common task that is complicated by the volume of data collected in airborne electromagnetic (AEM) surveys. We test a supervised learning approach for AEM data collected in the La Grange groundwater area, Western Australia. We use machine learning to identify the most likely geological setting at each station and use this to derive the probable extent of the seawater interface. We employ standard techniques such as cross validation to benchmark machine learning algorithms such as nearest-neighbour, naive Bayes and support vector networks. The good agreement between a qualitative interpretation and the best-performing machine learning algorithm, here a random forest algorithm, for the seawater interface extent suggests that automatic classification has the potential to speed up the interpretation of large airborne electromagnetic surveys. Our results also suggest that careful use of machine learning algorithms trained on high quality interpretations can lead to more objective geological interpretations particularly when airborne electromagnetic data are collected in order to map regional geology. A modest effort spent interpreting small but representative survey portions can be leveraged to geological mapping of the survey as a whole.