Statistical error estimation and optimal merging of MERRA and AMSR-E soil moisture and temperature datasets in preparation for SMAP

When modeling ecosystem carbon and water processes the question often arises as to which input meteorology dataset to use. Commonly used reanalysis and remotely sensed datasets can give significantly different model results. A computationally intensive solution is to produce an ensemble of model runs using a different combination of input datasets for each run. However, independent information from each dataset can be exploited a priori to provide a new dataset that is more accurate than any single data source alone. We develop and apply a statistical Kalman filter method for combining remotely sensed daily air temperature and surface soil moisture retrievals from AMSR-E with similar land parameters from the GMAO MERRA (Modern Era Retrospective-Analysis for Research and Applications) global reanalysis, using a simple time series model as the forecast equation. The filter is first calibrated using either triple-collocation or maximum likelihood to estimate error covariances and then run forward to estimate a new, optimal data time-series. The method is tested using in situ soil moisture and air temperature from locations spanning a variety of ecosystems. Filtering the anomaly components (daily variability) results in accurate error estimates and improved time series (1-21% R2 and 3-13% RMSE improvement (quartiles) for 41 soil moisture locations) relative to in situ observations. However, filtering the climatology components (seasonal variability) is challenging because auto-correlated errors and periodic biases hamper identification of error covariances. These results are being applied for optimal dataset combination as input for an ecosystem carbon flux model in preparation for the NASA Soil Moisture Active and Passive (SMAP) mission.