Snow Water Equivalent Reanalysis Over a Scarce Data Region Via Assimilation of Snow Covered Area from Landsat 5, 7 and 8

Abstract:

In this work we apply a Bayesian methodology to reconstruct historical SWE estimates over two test watersheds in the semi-arid Andes (33°S). The approach combines climatological reanalysis forcing data, in-situ precipitation, a Land Surface Model (LSM) and a snow depletion model to generate ensembles of prior SWE for each pixel over the watersheds. The precipitation forcing data is then updated by assimilating fractional snow covered area (fSCA) from the Landsat 5, 7 and 8 satellites using an Ensemble Kalman Smoother approach, generating new precipitation fields that are then used in a posterior LSM simulation. The resulting (posterior) SWE estimates are validated using in-situ snow depth measurements surveyed during 2009 to 2013. Results show significant improvements in error statistics, including an increase in correlation, a reduction in bias and a reduction in Root Mean Square error. After assimilation the posterior fSCA temporal evolution is consistent with the remotely sensed fSCA data, showing significant differences with the original prior estimates generated in the open-loop model simulation. The framework is robust to negative and positive biases in prior precipitation, resulting in similar posterior metrics regardless of the initial bias level. The resulting SWE fields are analyzed in terms of spatial distribution and relation with physiographic characteristics. A high correlation with slope suggests that gravitational distribution mechanisms play a significant role in snow distribution over these watersheds. Despite the fact that the LSM has no gravitational redistribution algorithm, the assimilation framework was able to capture this source of variability.