Synthetic tests of passive microwave brightness temperature assimilation over snow covered land using machine learning algorithms

Abstract:

A novel data assimilation framework is evaluated that assimilates passive microwave (PMW) brightness temperature (Tb) observations into an advanced land surface model for the purpose of improving snow depth and snow water equivalent (SWE) estimates across regional- and continental-scales. The multifrequency, multipolarization framework employs machine learning algorithms to predict PMW Tb as a function of land surface model state information and subsequently merges the predicted PMW Tb with observed PMW Tb from the Advanced Microwave Scanning Radiometer (AMSR-E). The merging procedure is predicated on conditional probabilities computed within a Bayesian statistical framework using either an Ensemble Kalman Filter (EnKF) or an Ensemble Kalman Smoother (EnKS). The data assimilation routine produces a conditioned (updated) estimate of modeled SWE that is more accurate and contains less uncertainty than the model without assimilation.

A synthetic case study is presented for select locations in North America that compares model results with and without assimilation against synthetic observations of snow depth and SWE. It is shown that the data assimilation framework improves modeled estimates of snow depth and SWE during both the accumulation and ablation phases of the snow season. Further, it is demonstrated that the EnKS outperforms the EnKF implementation due to its ability to better modulate high frequency noise into the conditioned estimates. The overarching findings from this study demonstrate the feasibility of machine learning algorithms for use as an observation model operator within a data assimilation framework in order to improve model estimates of snow depth and SWE across regional- and continental-scales.