Compensating for penetration in radar altimetry cryosphere elevation estimates using deep learning

Abstract

Radar altimetry is commonly used for monitoring changes within the cryosphere and itscontribution to sea-level rise. Recent advances in Swath altimetry processing, using theinterferometric mode of CryoSat-2, have enabled fine 500m spatial resolution surface elevationmodels from each satellite pass. However, there is variability in radar elevation estimates, oftendue to the penetration of radar waves into snow and firn, yielding differences compared tolocal airborne LiDAR altimeters which are less impacted by penetration. While neuralnetworks are increasingly being used in a wide variety of domains to enhance predictions anddetect changes, they have not previously been applied to radar altimetry to correct forelevation biases within the cryosphere. In this study, we present a novel approach to adjustingfor elevation bias by creating a neural network that is trained to predict the elevation providedby local airborne LiDAR from CryoSat-2 radar elevation data where both data are availableand then apply that model where only radar data are present. We investigate the challengeswith building such models and review the variety of configurations and considerations. Finally,we present two proof of concepts that show good spatial and temporal transfer ability andcompensate for 70-90% of the mean penetration, while reducing the root mean squared errorby 10-17%.