A dual state-parameter updating scheme using the particle filter and high-spatial-resolution remotely sensed snow depths to improve snow simulation

Abstract

Snow varies widely in space and time over high mountain regions. Accurately representing the spatial distribution of snow water equivalent (SWE) is critically important for improving our understanding of snow accumulation and melt processes. Despite its importance, in situ observations are lacking in poorly gauged regions such as the headwater region of the Yangtze River (HRYR). Traditional remotely sensed retrievals are highly uncertain due to the effect of cloudiness (e.g., optical remote sensing-derived snow cover area) and the coarse spatial resolution (e.g., passive microwave remote sensing-derived snow depth). Hydrological modeling is a powerful way to understand the snow processes, but uncertainty still exists due to model deficiency and errors of forcing data. Assimilating high-spatial-resolution remotely sensed snow data into a snowmelt model may be potentially valuable for more accurate snow predictions. In this study, a high-spatial-resolution remotely sensed snow depth data set (500 m, derived by integrating snow cover area, land surface temperature, and passive microwave brightness temperature products) was assimilated into a snowmelt model within a particle filter (PF) assimilation framework, which updates both model state variables and parameters simultaneously. After assimilation, the time series of basin-averaged SWE estimation showed an appreciable improvement with respect to the simulation with the traditional calibration (TC) method, in terms of the Nash-Sutcliffe efficiency (NSE) coefficient increased by ~ 10–20% and root‐mean‐square error (RMSE) decreased by ~ 7–18%. The PF approach also greatly improved the spatial distribution of SWE estimation with RMSE decreased by ~ 15–30%. The SWE estimation from the PF was also comparable or even better than that from another assimilation scheme, namely, the direct insertion. Comparison with in situ snowfall data indicated that the simulated snowfall from the PF outperformed the TC, with RMSE decreased by ~ 15–32% and correlation coefficient increased by ~ 58–83%. Furthermore, the evolution of parameters suggested the applicability of the PF method with spatially variable parameters. With spatiotemporally variable parameters in the PF, the snow model could perfectly simulate the actual snow distribution particularly over high elevation regions where the average temperature was lower than 0 °C, while fixed parameters in the TC cannot simulate the variable snow distribution. The proposed data assimilation framework has large potential of improving the accuracy of snow prediction across poorly gauged high mountain areas.