Speaker
Description
Seasonal snow in mountain regions is an essential source of water. The monitoring of snow properties, such as snow depth (SD) or snow water equivalent (SWE), can benefit a range of applications, including water resource management, flood prediction and renewable energy production. A promising remote sensing method for the monitoring of mountain snow uses Synthetic Aperture Radar (SAR) observations from the ESA and Copernicus Sentinel-1 satellite mission, which operates at C-band (5.4 GHz) frequency in dual-polarization (VV and VH). This method relies on the observed increase in radar signal scattering, more precisely the increase in VH relative to VV backscatter, with the accumulation of dry snow. SD can be retrieved using a change detection algorithm, translating the observed temporal changes in radar backscatter into changes in SD. The resulting Sentinel-1 SD retrievals have demonstrated accuracy at 500-m resolution over mountain regions, including the European Alps. However, the optimization of the model parameters and their transferability to different mountain regions at the global scale have to date been insufficiently addressed.
This work provides a detailed investigation of the change detection model formulation, parameter calibration, and parameter transferability over three large mountain regions: the European Alps, Scandinavia, and western US-Canada. The parameter optimization is performed using Monte Carlo experiments, distinguishing between two scenarios: a single (global) optimization, and optimization per mountain region. Furthermore, the impact of vegetation cover (no, low, and high vegetation) on the parameter optimization is assessed. Finally, the optimization of the change detection algorithm is repeated for a regional study case over the Swiss-Italian Alps, using X-band SAR observations from COSMO-SkyMed Second Generation, to investigate the transferability of the approach to a different observation frequency, and to compare retrieval performances at X- and C-band.
The study shows promising retrieval results across the European Alps (R = 0.80, MAE = 0.34 m), Scandinavia (R = 0.68, MAE = 0.24 m) and western US-Canada (R = 0.81, MAE = 0.28 m), even when using a single change detection model parameterization for SD retrieval from Sentinel-1 backscatter observations (above metrics are excluding SD = 0). However, a regional parameter optimization allows to reduce small regional biases as obtained with the single model, i.e. a slight overestimation over Scandinavia, and underestimation over western US-Canada. Similarly, a parameter optimization distinguishing between no, low, and high vegetation classes can slightly improve the overall retrieval performance across vegetation conditions. The application of the change detection algorithm at X-band results in an almost identical retrieval performance compared to C-band. Slightly better results are obtained at X-band for shallow snow (SD < 0.75 m) and with the use of low satellite incidence angles (< 30°), while slightly better results are obtained at C-band for deeper snow.
The results demonstrate that the Sentinel-1 mission can provide useful snow depth information across diverse mountain ranges. Future research aims to further develop and validate the Sentinel-1 SD retrievals across the global mountains (including the Southern Hemisphere) covering the entire Sentinel-1 era, and to assimilate the retrievals into a snow model in order to provide daily, area-wide coverage for both SD and SWE estimates.