Improving Fully Distributed Snowpack Simulations by Mapping Perturbations of Meteorological Forcings Inferred From Particle Filter Assimilation of Snow Monitoring Data

Snow plays a crucial role in the water balance of mountainous regions by affecting the timing and magnitude of runoff and, thus, water availability and flood hazards. However, estimating snow water equivalent (SWE) in mountainous regions is challenging due to its substantial spatial variability, the lack of accurate distributed measurements, and the uncertainties of snow models. Model uncertainties are primarily bound to uncertainties in the meteorological forcings. This study proposes an assimilation scheme to identify and correct spatiotemporal error patterns in the meteorological forcing data. Using a particle filter, we assimilated in situ snow depth observations from 444 stations across Switzerland into an ensemble simulated by the multi-layer, physics-based snow model FSM2OSHD. The ensemble is created by applying traceable, fixed perturbations to the energy input and the amount and phase of precipitation. This allows us to identify and correct errors in the meteorological forcing data for each station site and each 3-day assimilation window. Leveraging spatial correlation in these errors, we distribute the corrections across the entire model domain using a weighted three-dimensional spatial interpolation method. The refined meteorological data then serve as forcing for improved model runs, allowing unobserved grid points to benefit from the point assimilation. A leave-one-station-out cross-validation shows marked improvements in root-mean-squared error and bias for estimates of snow depth and SWE over the entire elevation range and multiple winter seasons. The proposed scheme is a promising step in developing comprehensive data assimilation solutions for large-scale, fully distributed, near real-time snow modeling applications, taking into account operational constraints and practical considerations.