Improving the estimation of snow depth in the Noah-MP model by combining particle filter and Bayesian model averaging

Abstract

Noah with multiparameterization (Noah-MP) is a new-generation land surface model that is enhanced by Noah and has multiparameterization as a major feature. Although various schemes have improved the applicability of this model, the ability to reduce the uncertainties of the simulation results remains a challenge. To improve the accuracy of snow-depth prediction based on the multiparameterization schemes within the Noah-MP model, we proposed a novel snow data assimilation framework configured using Bayesian model averaging (BMA) and genetic particle filter (GPF). Within this framework, we employed a kernel density smoother to create a probability density function of the model ensembles that are created using the GPF. We selected representative parameterization scheme combinations within the Noah-MP model to conduct the simulation experiments. We examined the feasibility of BMA applied in snow-depth prediction. On this basis, we investigated the performance of the BMA-GPF at eight seasonally snow-covered study sites with different snow climates. The results showed that BMA can be used to obtain a deterministic snow-depth prediction with a determination coefficient of 93.51% and a containing ration of prediction interval of 94.68%. These results were significantly higher than the four scheme combinations within the training period; however, the bias of deterministic prediction was well below expectations. The containing ration of uncertainty prediction interval was just 70% and was even lower than the four scheme combinations within the test period. The BMA-GPF snow data assimilation framework helped to obtain snow-depth prediction with a containing ration that ranged from 95% to 98% at eight sites and could be used in different snow climates. Further comparative results indicated that the BMA-GPF snow data assimilation framework no longer depended on the correction coefficients of the model prediction, and the assumption that the posterior probability of the state variable obeys the Gaussian distribution was not required. Finally, we concluded that the BMA-GPF is a suitable candidate approach to snow-depth multimodel ensemble simulation.