Airborne Radar Sounders (RSs) are active sensors that acquire subsurface data for Earth observation. RS data (radargrams) provide information on buried geology by imaging subsurface dielectric discontinuities. Recently, several automatic RS target identification techniques have been proposed, being convolutional neural network (CNN)-based methods the most promising. However, they require numerous labeled data that are hard to retrieve in the subsurface environment targeted by RS. Further, they are not designed to effectively deal with problems showing unbalanced classes like RS segmentation. We introduce newer cryosphere subsurface targets in the inland and coastal areas that can have a very low probability. To deal with the higher complexity and variability than previous works, we propose a transfer learning framework for RS data to mitigate the need for a large amount of labeled data and handle extremely unbalanced target classes. Herewith, we propose two transfer learning-based mechanisms for radargram segmentation. The first uses a lightweight architecture whose pre-training is supervised with a large labeled dataset from other domains. The second mechanism uses a deep architecture pre-trained in the RS domain, considering the pretest task of radargram reconstruction. The architectures are modified to deal with the characteristics of RS data and the radargram segmentation task. Finally, both methods are fine-tuned with a few labeled radargrams to learn radargram features useful for segmentation. We reveal experimental results on radargrams acquired in Antarctica by MCoRDS-1 and MCoRDS-3. The results demonstrate the effectiveness of transfer learning for radargram segmentation.