Snow Radar Echogram Layer Tracker: Deep Neural Networks for radar data from NASA Operation IceBridge

This paper documents the performance of two deep learning models developed to automatically track internal layers in Snow Radar echograms. A novel iterative RowBlock approach is developed to circumvent the small training-data problem peculiar to radar data by recasting pixel-wise dense prediction problem as a multi-class classification task with millions of training data. The proposed models, Skip_MLP and LSTM_PE, achieved tracking accuracies of 81.2% and 87.9%, respectively, on echograms from the dry snow zone in Greenland. Moreover, 96.7% and 97.3% of the errors are less than or equal to two pixels for both models respectively. The tracked layers were used to estimate annual accumulation over two decades and compared with Regional Atmosphere Model (MAR) estimates to yield a coefficient of determination of 0.943, thus validating this approach.