Retrieval of ice concentration data from satellite sensors at high spatial resolution
The declining sea ice extent in the Arctic is closely monitored due to its role as an important climate change indicator. This monitoring is primarily carried out using sea ice concentration retrievals from passive microwave sensors. While passive microwave sensors have the advantage of being relatively insensitive to weather conditions and do not require sunlight to return meaningful data, the data from these sensors have a coarse spatial resolution (5 km - 55 km). This can lead to blurring of the ice edge and small scale features in the ice cover (like floes and filaments of ice). We have found retrievals from these sensors are typically biased in marginal ice zones, such as the Gulf of Saint Lawrence and the Labrador Coast, both situated along the eastern Coast of Canada. This makes it very difficult (if not impossible) to use current ice concentration retrievals from passive microwave data to monitor changes in the marginal ice zone, for example due to warming air temperatures or changes in ocean currents.
To overcome these shortcomings we are (i) investigating the reasons for these biases (ii) developing methods to retrieve ice/water and ice concentration from synthetic aperture radar (SAR) imagery, and (iii) developing methods to assimilate or fuse both SAR and passive microwave data. SAR images have relatively high spatial resolution (eg., 50m) with the main challenge being automated interpretation of the data. Increasing volumes are becoming available through the European Sentinel-1 mission and the Canadian Radarsat Constellation mission, and using these data in operational forecasting systems is becoming significant priority.
At the same time, the recent release of RADARSAT-1 data archives by the Canadian Space Agency means almost 25 years of SAR data are available for analysis. Although these data are not continuous in space and time, we can work toward building a climatology of sea ice from SAR for specific regions (for example those monitored regularly by the Canadian Ice Service should have good coverage). This will then be used with passive microwave data to enhance our understanding of sea ice change.
Left, Ice eddies in Lancaster Sound during freeze-up, Right, ice eddies near the tip of Labrador, mid-winter. (Synthetic aperture radar (SAR) images: Copyright MacDONALD DETTWILER and ASSOCIATES LTD (2010)-All Rights Reserved)
Identification of ice eddies in SAR Imagery
Ice eddies can be seen in the marginal ice zone, for example along the Canadian east coast and the outflows of the Canadian Arctic Archipelago. We know very little about these eddies. In contrast, ocean eddies in regions without sea ice have been studied since the widespread use of satellite altimeter data. They are known to transport heat anomalies and nutrients, and their spatial distribution can impact the stratification of the mixed layer (e.g., in the Labrador Sea). We are presently working on automatic identification of ice eddies in SAR imagery. I am interested in the role of these eddies play in transport, on the formation of pressured ice (when they enter a consolidated ice zone) and on the exchange of heat and moisture with the atmosphere.
Data driven approaches for satellite based retrievals and downscaling
We have been very sucessful in using data-driven approaches for the retrieval of ice/water and ice concentration data from SAR. In this realm, we have shown convolutional neural networks are able to extract meaningful features from these images and produce high-quality sea ice concentration states. A significant challenge in this realm is the training data. Image analysis charts are manual analyses where each analysis corresponds to a given SAR image. These charts are therefore conicident in time and space with the SAR image. However, in these charts each data point yields ice concentration and ice type (new ice, first-year ice etc) information over a larger, and irregularly shaped, spatial region. This makes it a challenging downscaling problem.
We are now looking into estimating uncertainty information in tandem with the retrieved value. This type of information is key to using the data in downstream applications (data assimilation, decision support). There are several unresolved issues in validating both uncertainty information and high-resolution retrievals. We are looking into this both for sea and lake ice, with plans to expand to land surface applications.