Back to search

FRINATEK-Fri prosj.st. mat.,naturv.,tek

Thickness of Arctic sea ice Reconstructed by Data assimilation and artificial Intelligence Seamlessly

Alternative title: Rekonstruksjon av tykkelsen av sjøien i Arktisk ved sømløs bruk av dataassimilering og kunstig intelligens

Awarded: NOK 11.5 mill.

Satellites have been measuring the extent of the whole Arctic sea ice since the late 70’s. But only since 2010 are they also able to measure its thickness and thus monitor the decline of its total volume. This information would have been invaluable for climate studies, if only these new satellites hadbeen sent in orbit ten years earlier. Short of owning a time machine, can we learn enough from the last well-observed decade (since 2010) to reconstruct an earlier decade? What we have is a numerical model that can reproduce the known physics and assimilate satellite measurements of the ocean and sea ice during the well-observed decade; this is a well-established tool but it cannot use the ice thicknesses measured before the satellites: research cruises, older satellites, and proxy measurements because they are too scattered. A more recent tool introduced by the TARDIS team now allows to combine data assimilation and machine learning to build data-driven relationships between ice thickness and other variables. The tool has been developed and validated against independent mooring data this year, proving more accurate than a crude bias correction. To further enhance the sea ice thickness reconstruction by classes of ice thickness, we have exploited a new satellite called IceSAT2 which provides several thousands of ice thickness values within a single model grid cell. This is our training data to learn the subgrid-scale probability distribution of the Arctic sea ice thickness, which contains the memory of the sea ice evolution, and is therefore a key information to reverse its history. TARDIS will apply this method in order to finally reconstruct the past ice thickness for climate studies. If TARDIS succeeds at time travelling and spatial interpolation over the Arctic sea ice, the team will consider analogous applications of the same methodology.

The project will explore two possible ways to extend Data Assimilation (DA) techniques using Machine Learning (ML). The first will reconstruct a Sea Ice Thickness (SIT) in a past period when satellite data was incomplete. The second ML approach will be trained to retrieve the unobserved subgrid-scale Ice Thickness Distribution (ITD). TARDIS will start from a DA reanalysis of the Arctic Ocean (without ML), it will then carry out a multivariate statistical exploratory analysis to select auxiliary variables interest to train the ML networks. The selection and testing of different ML algorithms will be done on the most recent 10 years, when satellite SIT observations are gapless. The empirical relationships between SIT and predictor variables will then be extrapolated to extend the SIT time series to the previous decade and retrieve the ITD after DA. Independent in situ data will be used to tune the ML hyperparameters. The oceanographic interpretation of the new dataset in terms of Arctic heat budget will also be carried out within TARDIS. The project will last for 4 years, and will employ researchers working at the Nansen Center, plus a 3-years post doc to be recruited and a remote sensing expert at NORCE. International partners in China will contribute to the interpretation of the results, while partners in France and UK will bring expertise in ML and DA techniques. The acronym TARDIS refers to a device able to travel in time and space in the British cult series Doctor Who.

Publications from Cristin

No publications found

No publications found

No publications found

Funding scheme:

FRINATEK-Fri prosj.st. mat.,naturv.,tek