Application of Self-Organizing Maps to characterize subglacial bedrock properties based on gravity, magnetic and radar data - an example for the Wilkes and Aurora Subglacial Basin region, East Antarctica

Abstract

Subglacial bedrock properties are key to understanding and predicting the dynamics and future evolution of the Antarctic Ice Sheet. However, the ice sheet bed is largely inaccessible for direct sampling, and characterization of subglacial properties has so far relied on expert interpretation of airborne geophysical data. To reduce subjective choices in the joint analysis of data and related biases, we present a Self-Organizing Map (SOM), an unsupervised machine learning technique. The concept of SOMs is briefly introduced and we discuss data selection and their associated attributes. First, we analyzed the correlation between attributes to provide a validation of an appropriate choice. Next, we trained the SOM on attributes derived from gravity, magnetics and ice-penetrating radar data for the Wilkes and Aurora Subglacial Basin region in East Antarctica. In contrast to earlier studies, our approach uses original line data as much as possible. These have a much higher resolution than the smooth gridded products used in previous studies. Our results show marked agreement with past studies on predicting regional bed characteristics such as the presence of crystalline basement and sedimentary basins. Additionally, our results indicate the ability to resolve finer details, demonstrating the potential in applying SOM to subglacial geologic mapping.