Bias correction in SMAP soil moisture assimilation using a neural network approach

Statistical techniques permit the retrieval of soil moisture estimates in a model climatology while retaining the spatial and temporal signatures of the satellite observations. As a consequence, they can be used to implement an alternative bias correction to the local cumulative distribution function matching typically used in soil moisture data assimilation (DA) systems. In this study, a statistical neural network (NN) retrieval algorithm is calibrated using SMAP brightness temperature observations and modeled soil moisture (which is also used to calibrate the SMAP Level 4 DA system). Daily values of surface soil moisture are estimated using the NN and then assimilated into the NASA Catchment model. We assess the skill of the NN retrieval and the assimilation estimates through a comprehensive comparison to in situ measurements from the SMAP core and sparse network sites. The NN method compares well against the official RTM based approach and is able to extract information from the SMAP observations that is complementary to the model. Additionally, we compare the NN method to more traditional bias correction approaches and analyze the potential of using spatially variable error estimates to improve the relative impact of observations in the assimilation.