A new approach for generating optimal GLDAS hydrological products and uncertainties

This study proposes a new approach that can be used to generate the optimal surface state information and associated uncertainties from the estimates provided by the six land surface models used by the Global Land Data Assimilation System(GLDAS). The Forstner and best quadratic unbiased variance component estimators are used simultaneously with the least-squares method to calculate optimal values and the associated uncertainties. To demonstrate the concept, the research focused on three GLDAS hydrological products, namely soil moisture (SM), snowwater equivalent (SWE), and canopywater (CAN) over the Canadian Prairies. When the Forstner estimator is applied, the estimated SMand SWE differ fromtheir correspondingmean values by 26 mm and 9 mm respectively. Almost similar resultwas found with the best quadratic estimator. The estimated maximum uncertainties of each component including SM, SWE and CAN vary from year to year (e.g. 35 mm in 2006, 12 mm in 2007 and 2009 and 0.1mmin 2001, respectively). The uncertainties of the total water storage (TWS) are almost similar to that of SM, which contributes more importantly to TWS in the area considered. The results obtained by the two proposed estimatorswere compared to the waterGAP hydrologicalmodels (WGHM), and to the Gravity Recovery and Climate Experiment (GRACE) terrestrial water storage anomalies. The optimalSWE anomalies generated fromGLDAS using the proposed approach showa maximumcorrelation of r= 0.97 with theWGHMSWE anomalies. The optimal TWS anomalies have a correlation of r = 0.91 with WGHM, and r = 0.71 with GRACE. However, the correlation jumps to r = 0.81 when GRACE TWS is corrected for groundwater signals (with a mean RMSE of 8.5 mm). The RMSE and mean absolute error between our proposed methods and WGHM and GRACE are better than those obtained with each individual LSM or their average value. No significant mean bias error is observed in each case. Finally, the analysis of the time-lag characteristics of the resonance period between the results and their coherence was done by using a cross wavelet transform and a wavelet coherence analysis. (C) 2020 Elsevier B.V. All rights reserved.