Detection of wet snow in the Arctic tundra from time-series fully-polarimetric RADARSAT-2 images

Abstract

Snow melt onset detection and mapping of snow extent during the spring melt period using spaceborne Synthetic Aperture Radar (SAR) is important for various applications including understanding and predicting snowmelt runoff, flood hazard risk, and freshwater resource supply. In this study, we investigated the time-series evolution of RADARSAT-2 C-band HH, VV, VH and HV radar backscattering coefficients and scattering contributions during the spring melt seasons of 2017 and 2018 over the Niaqunguk (Apex) River, Nunavut, Canada. Meteorological observations and optical Planet satellite images were analyzed to support interpretations of SAR data in relation to snowpack warming and melt processes at the high spatial resolution of 6 m. Our results suggest that the cross-polarization (HV and VH) backscatter from snow significantly dropped by ~10.7 dB when the air temperatures became positive and snowpack moisture content began to increase (i.e., melt onset). Co-polarization (HH and VV) backscatter from snow also decreased, but by a smaller magnitude of ~7.4 dB. Freeman–Durden decomposition analysis suggested that the volume scattering contribution dropped drastically, and surface scattering dominated at the melt onset. The timing of melt onset can be readily inferred from the temporal changes in backscatter signal, which was particularly pronounced in the HV time-series. The backscatter decrease (e.g., 2.5 dB for HH and 3.4 dB for HV) associated with melt onset was also pronounced for large areas, where both snow-covered and snow-free surfaces were present, and water was masked out. Cross-polarization signal demonstrated better separability between snow free and snow-covered surfaces than co-polarization signal throughout the melt season. This suggests that the HV backscatter signal is more suitable for snow melt onset and snow/ground detection compared to conventionally used HH and VV, given the instrument noise floor is sufficiently low to capture the HV signal variations. Our classification results demonstrated slightly higher overall accuracy for HV (83.9%) than for HH (82.7%) signal. However, snow and bare-ground classification accuracies were 49 limited to the late season period when class separabilities were lower compared to the earlier melt season, preventing robust assessment of classification accuracy using HV backscatter due to limited ground truth data.