Challenges with deriving sea ice roughness from SAR imagery using deep learning
Anton Korosov
This study presents a novel approach for retrieving sea ice roughness using Synthetic Aperture Radar (SAR) imagery, enhanced by convolutional neural networks (CNNs). By training the CNN with altimetry-derived roughness as the target, the method accurately captures surface features crucial for sea ice monitoring. Our model leverages the spatial resolution of SAR and the physical measurements from altimetry to improve roughness estimation across diverse ice conditions. Results demonstrate abiility to distinguishing between smooth and rough ice. This technique is expected to provide inputs for sea ice model with surface drag parametrisation and improve sea ice dynamics simulations.