Glacier surface melt monitoring using Sentinel-1 SAR backscattering coefficient and polarimetric decomposition features at Greenland ice sheet

The synthetic aperture radar (SAR) imagery is sensitive to glacier melting because the dielectric constants vary when the ice surface melts. Previous studies mainly utilized the backscatter coefficients of SAR data. This study aims to explore the potential of using decomposed features of dual-polarized SAR imagery for detecting glacier melting. At five selected study sites in the Greenland ice sheet (GrIS) where automatic weather stations (AWS) are distributed, backscatter coefficient and polarimetric decomposition parameters of Sentinel-1 IW imagery are analyzed. AWS air temperature and MODIS land surface temperature are regarded as the ground truth. Glacier surface is firstly categorized into different radar glacier zones (RGZ). Then, an ensemble decision tree model is trained using the different feature values of both SAR features and RGZ categories to determine the glacier melt status. The overall accuracy of glacier melt detection only uses backscatter features reaches 76%, and 66% for the polarimetric decomposition features. Although the latter shows lower accuracy than the former, it is more sensitive in the areas with poor discrimination solely based on backscatter coefficient, especially at bare ice zones. Combining both backscatter coefficient and polarimetric decomposition features gives an accuracy of 80%. Finally, the accumulated melting days are analyzed at different RGZs. Accumulated melting days decrease with increasing altitude. Surface melting in the bare ice zone lasts up to 120 days, 60–120 days for the wet snow zone, and less than 10 days in the percolation zone. Strong melting events occurred in 2019 and 2021, corresponding to the most recent extreme weather events. Compared to the National Snow and Ice Data Center (NSIDC) polar daily land freeze/thaw status product, our results revealed consistent spatial patterns of melting, showing discrepancy limited within ± 10 days, possibly due to the detecting penetration depth difference between Sentinel-1 and AMSR2.