Mapping of deep internal reflection horizons, method modifications and applications.

The task of mapping of deep internal reflection horizons (IRH) of ice sheets has been a crucial step for a variety of glaciological studies, for instance relating ice core age-depth relationships, tuning ice sheet models, and extend dated layers beyond ice core sites. However, mapping a sufficient number of IRHs is a time-consuming and error-proned task. Thus, there have been ongoing endeavors for automatized pipelines to perform this.

In this work, a complete pipeline for automatic mapping of deep IRH, which determine ice layer boundaries, is presented. This pipeline is tested on radargrams from Dronning Maud Land Antarctica and shows good performance in mapping a number of deep IRHs. The model shows great promise to be used on snow radargrams and obtaining recent accumulation rates as well.

We have applied convolutional neural networks (CNN) to achieve this. The training data is composed of a small set of complete hand-labeled radargrams as well as radargrams that are labeled using conventional feature extraction methods. This task requires dense pixel-level predictions, and ground-truth collection is time-consuming and prone to errors, therefore a group of modifications have been implemented on the model. The role of post-processing is discussed, since the output of the model is a raw image and much work is done on the model output. The potential of such a deep mapped stratigraphy is discussed and various applications are pointed out.