Towards Generalizable Deep Learning Algorithms for Echogram Layer Tracking

The accelerated melting of ice sheets in Greenland and Antarctica, driven by climate warming, is significantly contributing to global sea level rise. To better understand this phenomenon, airborne radars have been deployed to create echogram images that map snow accumulation patterns in these regions. Utilizing advanced radar systems developed by the Center for Remote Sensing and Integrated Systems (CReSIS), around 1.5 petabytes of climate data have been collected. However, extracting ice-related information, such as accumulation rates, remains limited due to the largely manual and time-consuming process of tracking internal layers in radar echograms. This highlights the need for automated solutions.

Machine learning and deep learning algorithms are well-suited for this task, given their near-human performance on optical images. The overlap between classical radar signal processing and machine learning techniques suggests that combining concepts from both fields could lead to optimized solutions.

In this work, we developed custom deep learning algorithms for automatic layer tracking (both supervised and self-supervised) to address the challenge of limited annotated data and achieve accurate tracking of radiostratigraphic layers in echograms. We introduce an iterative multi-class classification algorithm, termed “Row Block,” which sequentially tracks internal layers from the top to the bottom of an echogram based on the surface location. This approach was used in an active learning framework to expand the labeled dataset. We also developed deep learning segmentation algorithms by framing the echogram layer tracking problem as a binary segmentation task, followed by post-processing to generate vector-layer annotations using a connected-component 1-D layer-contour extractor.

Additionally, we aimed to provide the deep learning and scientific communities with a large, fully annotated dataset. This was achieved by synchronizing radar data with outputs from a regional climate model, creating what are currently the two largest machine-learning-ready Snow Radar datasets available, with 10,000 and 50,000 echograms, respectively.