Hybrid Deep Learning to Trace Snow Layers through Radargrams

Abstract

Global warming is drastically reducing the polar ice caps and leading to sea level rise. Continuous monitoring and measurement of this reduction is imperative to estimate the resulting socio-economic damage that may affect human lives. Polar ice sheets are typically monitored through airborne radar sensors which give a two- dimensional image of the subsurface profile. Multiple snow accumulation layers can be observed in these images, the changing thickness of which can contribute to glacier accumulation or ablation rates. A variety of semi-automated techniques have been proposed in the past which process the radargrams and identify these snow accumulation layers. This work proposes the use of deep learning to develop fully-automated algorithms for snow layer detection and discusses some of the issues faced while developing these algorithms, such as the low visibility of the layers, noise in the radargram and the lack of quality training labels. This work thus proposes ‘hybrid’ deep learning architectures where the architecture is not just data-driven, but also leverages additional physical information of the environment, or of the radar signal, to improve the generalizability, robustness, and scientific reliability of the architectures. Such hybrid approaches for neural network training not only provide techniques to augment a purely data-driven architecture with physical information, but also offer potential extensions to various radar systems; thereby enhancing snow layer detection and refining climate change projections.