Item

Abstract

Sea ice thickness (SIT) retrieval using Global Navi-
gation Satellite System-Reflectometry (GNSS-R) is a challenging
problem in sea ice remote sensing, especially for SITs over 1 m,
which is still in the blank stage. In this article, a seamless
stacking-based retrieval method for SIT is proposed, which
reduces the root mean square error (RMSE) for thicknesses
below 1 m while ensuring the accuracy of SIT retrieval for
thicknesses above 1 m. Principal component analysis (PCA) was
used to extract delayed Doppler map (DDM) features, while the
scattering coefficient and incidence angle were calculated using
TechDemoSat-1 (TDS-1) data. Sea ice salinity and temperature
were derived from Soil Moisture and Ocean Salinity (SMOS)
data and used as inputs to the model along with other features.
The performance of four machine learning (ML) algorithms—
decision tree (DT), K-nearest neighbors (KNNs), support vector
regression (SVR), and random forest (RF)—was compared, and
the stacking model was constructed using these four algorithms
as the base learner to improve performance. Validation using
SMOS data for thicknesses up to 1 m showed that the stacking
algorithm significantly improved retrieval accuracy, reducing
the RMSE from 7 to 0.4 cm and improving the correlation
coefficient (r) from 0.94 to 0.99. For thicknesses greater than 1 m,
validation using Cryosat-2 data also showed strong performance.
In addition, the effect of sea ice parameters on retrieval accuracy
and sources of error was analyzed