Snow melting phases detection using apparent thermal inertia, radar data and numerical modelling

The timing of snowmelt has a crucial relevance for water management in mountain areas. Several approaches have been developed to detect the state of snow during the melting season using satellite observations, ground data, and numerical modelling, but these have rarely been compared directly. In this study, we used multitemporal (2016–2021) data from two meteorological stations located in the Aosta Valley (Western European Alps, Italy), namely, Torgnon (2050 m asl–sub-Alpine site) and Cime Bianche (3100 m asl–Alpine site), to force the SNOWPACK model and analyse the seasonal evolution of snow. Furthermore, we combined optical and thermal data continuously measured at the two sites to calculate the Apparent thermal inertia of snow (APs). This variable was found to resemble snow melting phases, i.e., warming, ripening, and output. In addition, Sentinel-1 Synthetic Aperture Radar (SAR) backscattering time series were analysed for both sites and compared with COSMO SkyMed backscattering time series. We extracted the start, end, and duration of the snow phases from all three datasets (APs, Sentinel-1, and SNOWPACK simulations) and compared them to assess their consistency. The comparison between Sentinel-1 and COSMO SkyMed demonstrates that both sensors independently capture the main stages of the snowmelt process showing consistent seasonal backscatter dynamics. While aggregated results showed a consistent representation of snowmelt among the three approaches (0.86 < R2 < 0.91; 4 <12), a phase- and site-specific analysis revealed that the Warming phase yielded lower statistics with respect to the Ripening and Output phase. and the comparison of warming-phase detection confirms an earlier onset of melting at the sub-Alpine site (DOY 60–160) than Alpine site (DOY 100–200), while also revealing method-dependent differences in phase identification. The estimated timing uncertainties correspond to 2.9%–8.9% of the snow season. Thus, these phase onset estimates are currently better suited to model evaluation and data assimilation than standalone peak-runoff prediction, and this uncertainty should be explicitly propagated in operational forecasting.