conductivity, is a key control on the thermal state of near surface permafrost. At the same time, accurately estimating the seasonal snow cycle at the kilometre scale is a considerable hydrometeorological challenge. Consequently, snow represents a major source of uncertainty in permafrost models. To constrain this snow induced uncertainty we propose a new ensemblebased snow data assimilation framework (ESDA) for fine scale snow state estimation that fuses a simple subgrid snow model and fine scale satellite-based surface albedo retrievals using the ensemble Kalman filter (EnKF; reviewed in Evensen [2009]). The potential of ESDA is demonstrated for the Bayelva catchment near Ny Åelsund (Svalbard, Norway) where independent ground-based observations of snow cover and the near surface ground thermal state were available to perform validation. On the modeling side of ESDA we adopt the subgrid snow distribution model (SSNOWD; see Liston [2004]) to estimate the snow water equivalent depth distribution, snow cover fraction and surface albedo at the grid scale (1 km). These model runs are forced by melt and net precipitation rates based on the energy and water balance derived from the meteorological fields provided by a (3 km resolution) Weather Research and Forecasting (WRF) model run. For observations our system makes combined use of two relatively new high level products: frequently available coarser scale (500 m) albedo retrievals from MODIS (MCD43A version 6) and intermittently available finer scale (30 m) albedos derived from Landsat8 surface reflectance retrievals. In the last step of the framework we apply the EnKF; a robust sequential data assimilation method that yields the optimal estimate of a system state based on the combined information from model results and observations, both of which are uncertain, provided a set of assumptions hold (see Evensen [2009]). The EnKF has been successfully implemented for a range of applications in numerous fields including oceanography, meteorology, hydrology, mining and reservoir geophysics, although to our knowledge this is the first time it is being applied directly to permafrost modeling. Simply stated an ensemble (a set) of model realizations, in this case capturing uncertainties in the meteorological forcing, are propagated forward in time and sequentially updated by the observations whenever these are available. The magnitude of the updates depends on the deviation of the model realizations from the observations as well as the respective uncertainties. Thereby, the result of the EnKF is expressed in terms of an ensemble of corrected model states, where the ensemble mean is interpreted as the most likely estimate and the ensemble spread is a measure of the uncertainty. Our results are promising; the evolution of the ensemble mean estimated snow cover using ESDA at Bayelva is shown to be much closer to the ground-truth, as observed by an independent automatic camera system, than that of the open-loop (no assimilation) estimate. Finally, we incorporate ESDA into the recently developed CryoGrid3 surface energy-balance driven permafrost model described in Westermann et al. [2016]. The results, with and without ESDA, are compared to in situ measurements from an array of randomly distributed ground surface temperature measurements within the modeled grid cell. A significant improvement in the skill of the model at capturing the near-surface ground thermal state is demonstrated, particularly in the ablation season. Thus, ESDA provides improved estimates of the state of permafrost at Bayelva. Due to the cheap computational cost, the framework is also applicable to much larger model domains. Moreover, given the robustness, owing to the global span of the satellite retrievals and the option of running SSNOWD with reanalysis data (e.g. ERA-Interim), it is possible to apply this framework to most permafrost regions on the planet.