New journal paper accepted: “Machine Learning for Arctic Sea Ice Physical Properties Estimation Using Dual-Polarimetric SAR Data”

We are happy to announce that a new paper has been accepted. The paper is summarized below.

“Machine Learning for Arctic Sea Ice Physical Properties Estimation Using Dual-Polarimetric SAR Data” by Katalin Blix, Martine M. Espeseth and Torbjørn Eltoft, IEEE Transactions on Geoscience and Remote Sensing, 22.09.2020

Early access: https://ieeexplore.ieee.org/document/9203812

Monitoring the rapidly changing Arctic sea ice has great scientific, economical and political importance. The most efficient way to monitor Arctic sea ice is by using remote sensing technology. In this regard, there are many different kinds of spaceborne sensors providing useful information. Data collected by the synthetic aperture radar (SAR) sensor onboard the Canadian Radarsat-2 (RS2) satellite allows for full polarimetric (quad-pol) observations, which enables us to generate parameters that can say something about the sea ice characteristics, such as surface roughness and salinity. However, the swath-width of quad-pol RS2 scenes is not large enough for operational sea ice monitoring. On the other hand, the Copernicus mission’s Sentinel 1 (S1) satellite not only allows large scale monitoring, but the data is also freely available. Although the SAR on S1 provides large coverage, the dual-polarimetric configuration of the sensor does not allow to retrieve the same sea ice information as RS2.

In this work, our goal was to overcome this trade-off between large coverage and high polarimetric information by using Machine Learning (ML) approaches. The objective was to establish a learnt functional relationship between data obtained by RS2 and S1, so that the advantageous properties of both sensors can be exploited.

In practice, we have shown how ML methods can be used on freely available, large swath S1 data to estimate parameters that describe sea ice characteristics which actually rely on quad-pol observations. Hence, our results indicate how advances in ML can potentially help to achieve improved, cost-effective, large-scale sea ice monitoring.

The image below shows the up-scaled indicator of salinity generated from a dual-pol S1 image using an ML method. (For more information, see https://ieeexplore.ieee.org/document/9203812).

Print Friendly, PDF & Email