Estimating snow depth on Arctic sea ice using satellite microwave radiometry and a neural network
Snow lying on top of sea ice plays an important role in the radiation budget because of its high albedo and the Arctic freshwater budget, and it influences the Arctic climate: it is a fundamental climate variable. Importantly, accurate snow depth products are required to convert satellite altimeter measurements of ice freeboard to sea ice thickness (SIT). Due to the harsh environment and challenging accessibility, in situ measurements of snow depth are sparse. The quasi-synoptic frequent repeat coverage provided by satellite measurements offers the best approach to regularly monitor snow depth on sea ice. A number of algorithms are based on satellite microwave radiometry measurements and simple empirical relationships. Reducing their uncertainty remains a major challenge. A High Priority Candidate Mission called the Copernicus Imaging Microwave Radiometer (CIMR) is now being studied at the European Space Agency. CIMR proposes a conically scanning radiometer having a swath >1900 km and including channels at 1.4, 6.9, 10.65, 18.7 and 36.5 GHz on the same platform. It will fly in a high-inclination dawn–dusk orbit coordinated with the MetOp-SG(B). As part of the preparation for the CIMR mission, we explore a new approach to retrieve snow depth on sea ice from multi-frequency satellite microwave radiometer measurements using a neural network approach. Neural networks have proven to reach high accuracies in other domains and excel in handling complex, non-linear relationships. We propose one neural network that only relies on AMSR2 channel brightness temperature data input and another one using both AMSR2 and SMOS data as input. We evaluate our results from the neural network approach using airborne snow depth measurements from Operation IceBridge (OIB) campaigns and compare them to products from three other established snow depth algorithms. We show that both our neural networks outperform the other algorithms in terms of accuracy, when compared to the OIB data and we demonstrate that plausible results are obtained even outside the algorithm training period and area. We then convert CryoSat freeboard measurements to SIT using different snow products including the snow depth from our networks. We confirm that a more accurate snow depth product derived using our neural networks leads to more accurate estimates of SIT, when compared to the SIT measured by a laser altimeter at the OIB campaign. Our network with additional SMOS input yields even higher accuracies, but has the disadvantage of a larger “hole at the pole”. Our neural network approaches are applicable over the whole Arctic, capturing first-year ice and multi-year ice conditions throughout winter. Once the networks are designed and trained, they are fast and easy to use. The combined AMSR2 + SMOS neural network is particularly important as a precursor demonstration for the Copernicus CIMR candidate mission highlighting the benefit of CIMR.