Retrieval of daily sea ice thickness from AMSR2 passive microwave data using ensemble convolutional neural networks

• Published in: GIScience & Remote Sensing
• Main Authors: Junhwa Chi, Hyun-Cheol Kim
• Format: Article in Journal/Newspaper
• Language: English
• Published: Taylor & Francis Group 2021
• Subjects: amsr2; arctic sea ice; convolutional neural network; deep learning; passive microwave; sea ice thickness; Mathematical geography. Cartography; GA1-1776; Environmental sciences; GE1-350; Arctic; Global warming; Sea ice
• Online Access: https://doi.org/10.1080/15481603.2021.1943213; https://doaj.org/article/d9a6a10dad474ddc9b2c538a3a5a48cd

Recently, measurement of sea ice thickness (SIT) has received increasing attention due to the importance of thinning ice in the context of global warming. Although altimeter sensors onboard satellite missions enable continuous SIT measurements over larger areas compared to in situ observations, these sensors are inadequate for mapping daily Arctic SIT because of their small footprints. We exploited passive microwave data from AMSR2 (Advanced Microwave Scanning Radiometer 2) by incorporating a state-of-the-art deep learning (DL) approach to address this limitation. Passive microwave data offer better temporal resolutions than those from a single altimeter sensors, but are rarely used for SIT estimations due to their limited physical relationship with SIT. In this study, we proposed an ensemble DL model with different modalities to produce daily pan-Arctic SIT retrievals. The proposed model determined the hidden and unknown relationships between the brightness temperatures of AMSR2 channels and SITs measured by CryoSat-2 (CS2) from the extended input features defined by our feature augmentation strategy. Although AMSR2-based SITs agreed well with CS2-derived gridded SIT values, they had similar uncertainties and errors in the CS2 SIT measurements, particularly for thin ice. However, based on quantitative validations using long-term unseen data and IceBridge data, the proposed retrieval model consistently generated SITs from AMSR2 at 25 km spatial resolution, regardless of time and space.