Assimilating high-resolution satellite snow cover data in a permafrost model

Abstract. In high-latitude and mountain areas, the seasonal snow cover features considerable variability over spatial scales of tens of meters which strongly impacts the ground thermal regime. In permafrost areas, spatial patterns of snow accumulation and melt strongly influence the landscape-scale response to atmospheric warming, creating a need to develop simulation tools that can represent such small-scale variability. In this study, we introduce a data assimilation scheme designed to integrate satellite-derived fractional snow covered area (FSCA) into the permafrost model CryoGrid. The idea is to reconstruct the winter snowpack from the satellite-derived timing of the meltout, which at the same time constrains the insulating effect of the snow cover and thereby improves simulations of the ground thermal regime. For this purpose, we employ an iterative ensemble-based data assimilation approach combing a Particle Batch Smoother with Adaptive Multiple Importance Sampling, which proves efficient to manage the high computational demands of the CryoGrid model. The outcomes of the study are assessed using spatially distributed measurements of ground surface temperature (GST) and snow water equivalent collected during four years in an area of approximately 0.5 km² on Svalbard. The data assimilation scheme is evaluated using idealized, synthetic FSCA observations compiled from the GST measurements, as well as Sentinel-2 derived FSCA at a spatial resolution of 10 m. For both data sets, the assimilation workflow markedly enhances the representation of the annual GST cycle in the model when simulating extremes in the snow distribution (wind-exposed ridges and snowdrifts) at the scale of individual pixels. However, due to frequent cloudiness and spatial mismatches between the point-scale GST and the 10 m resolution Sentinel-2 data, the enhancements are less consistent using Sentinel-2 derived FSCA, even resulting in poorer performance in certain years and locations. The data assimilation scheme is further adapted to simulate GST distributions over the entire study area using snowmelt statistics compiled from the 10 m Sentinel-2 FSCA pixels. The results show that the algorithm can not only reproduce the observed spatial variability of mean annual GST of up to 4.5 °C, but also the temporal pattern of the spatial variability, with GST varying significantly more in space during snow-covered winter and the snowmelt period compared to the snow-free months. The study showcases that a close integration of observations and models can significantly enhance our ability to characterize the state of the terrestrial cryosphere. In particular, we demonstrate that the data assimilation can uncover the "hidden" variable GST when using a process-rich land surface model like CryoGrid.