Automatic Identification of Basal Units in Ice Sheets Based on ResNet and Weight Control

In recent decades, radio-echo sounding (RES) has been extensively applied to study the flow and form of polar ice sheets. In certain regions within the ice sheet, the RES reveals a structure referred to as a “basal unit,” which differs from the overlying ice in terms of its characteristics, structure, and origin, and plays a significant role in the ice’s rheology and flow dynamics. However, methods for detecting basal units in RES data are semiquantitative and can lead to inconsistent identification. To address this issue, we propose an automatic “basal unit identification method” based on residual network (ResNet) and weight control. The method improves upon previous works in three aspects: 1) it simultaneously uses the signal and image features of RES and reduces inaccuracies associated with image analysis; 2) this method assigns weight to signal features that are affected by backscatter consistent with high particle concentrations in basal units, reduces the interference of concentration on signal characteristics, and improves the ability to identify basal unit; and 3) it provides weights related to signal feature recognition and calculates a composite recognition result that automatically identifies basal units. To validate the method’s effectiveness, we applied it to airborne RES data collected in recent years from the Gamburtsev Subglacial Mountains (GSMs) and Princess Elizabeth Land (PEL) regions in East Antarctica. A comparative analysis of the new method and previous methods indicates more accurate basal unit identification due to stronger resistance to interference from backscatter consistent with high particle concentrations.