Transfer learning for the semantic segmentation of cryoshpere radargrams

Radar sounders (RSs) mounted on airborne platforms are active sensors widely employed to acquire subsurface data of the cryosphere for Earth observation. RS data, also called radargrams, provide information on the buried geology by identifying dielectric discontinuities in the subsurface. Recently, a strong effort can be observed in designing automatic techniques to identify the main targets of the cryosphere. However, most of the methods are based on target-specific handcrafted features. Newly convolutional neural networks (CNNs) automatically extract meaningful features from data. However, supervised training requires numerous labeled data that are hard to retrieve in the RS domain. In this work, we adopt a CNN pre-trained in domains other than RS for automatically segmenting cryosphere radargrams. To adapt to the radargram characteristics, we introduce convolutional layers at the beginning of the pre-trained network. We modify the top layers of the network to a U- fashion autoencoder to extract relevant features for the target task. The new layers are fine-tuned with few labeled radargrams to identify and segment five targets: free space, continental ice layering, floating ice, bedrock, and EFZ and thermal noise. The pre-trained weights are not updated during fine-tuning. We applied the proposed approach to radargrams from Antarctica acquired by MCoRDS3, obtaining high overall accuracy. These results demonstrate the effectiveness of the method in segmenting radargrams and discriminating continental and coastal ice structures.