Testing simple local regression equations to derive long‐term global soil moisture datasets from passive microwave observations
Résumé
The European Space Agency's Water Cycle Multimission Observation Strategy (WACMOS) was set up in 2010‐2012 and supported by the Support To Science Element (STSE) program. Within WACMOS first long term soil moisture (SM) data records from passive and active microwave data have been developed. In January 2012 the ESA Climate Change Initiative (CCI) SM program started and refined the SM products obtained during the preliminary WACMOS project. In June 2012, the first long term SM datasets have been made public available through the ESA CCI portal (Su et al., 2010) and in July 2014 the second version was available also on the same portal. This product was computed by merging surface SM (SSM) data sets retrieved from several microwave sensors in an attempt to produce the most complete and consistent long‐term time series of SM (1978‐2013) (http://www.esa‐cci.org/)(Liu et al., 2012). Two dedicated SM spaceborne missions, the Soil Moisture and Ocean Salinity (SMOS) and Soil Moisture Active/Passive (SMAP) satellites, were launched just recently, and therefore not yet considered for the CCI product. SMOS is an innovative in space technology that has been providing multi‐angular microwave brightness temperatures TB observations at L‐band since 2010 (Kerr et al., 2012). The ESA established a passive microwave SSM fusion study to investigate the inclusion of SMOS SSM in a long‐term SM datasets. For this purpose, three approaches have been selected: the Land Parameter Retrieval Model (LPRM) algorithm (Owe et al., 2001; Owe et al., 2008; Van der Schalie et al., 2015), neural networks (Rodriguez‐Fernandez et al., 2014; Rodríguez‐Fernández et al., 2015), and local statistical regressions (Wigneron et al., 2004). The present study addresses only the local regressions, as the first two methods are addressed by other groups. Objectives: The main objectives of this study are to: (i) test a physically based local regression method to retrieve SSM from the AMSR‐E TB observations at the global scale, (ii) extend the SMOS SSM into the past (2003‐2009), (iii) compare the quality of this product to the other SM datasets produced from the two other approaches, with reference to modelled SSM datasets and in situ measurements at both global and local scales. Methods: Wigneron et al. (2004) and Saleh et al. (2006) have developed and evaluated physically‐based regression equations between the SSM and microwave reflectivity (i.e. one minus emissivity) derived analytically from the radiative transfer model (τ‐ω model) (Mo et al., 1982; Wigneron et al., 1995). More specifically, this study investigated the use of these local regression methods to retrieve SSM based on a combination of passive microwave remote sensing observations from the Advanced Microwave Scanning Radiometer (AMSR‐E; 2003 ‐ Sept. 2011) and the SMOS sensors. Regression coefficients were calibrated “locally” over each pixel using AMSR‐E horizontal and vertical TB observations and SMOS level 3 SM (SMOSL3; as a training data). This calibration process was carried out over the June 2010 ‐ Sept. 2011 period, over which both SMOS and AMSRE observations coincide. Results and conclusion: Based on these calibrated coefficients, global SSM maps were computed from the AMSR‐E TB observations during the 2003‐2011 period (referred here to as AMSR‐reg). The regression quality was assessed by evaluating the AMSR‐reg SSM maps against the SMOSL3 SSM products over the period of calibration, in terms of correlations (R) and Root Mean Square Error (RMSE). A good agreement, R (mostly > 0.75) and RMSE (mostly < 0.04 m3/m3), was obtained between the AMSR‐reg and SMOSL3 SSM products particularly over Australia, central USA, central Asia, and the Sahel. AMSR‐reg retrievals were compared to the original AMSR‐E SSM retrievals derived from the Land Parameter Retrieval Model (AMSR‐LPRM), against two kinds of reference SSM: (i) global MERRA‐Land SSM simulations (2003 ‐ 2009) and (ii) in situ measurements (2008‐2009). The results demonstrated that both AMSR‐reg and AMSR‐LPRM (better) successfully captured the temporal dynamics of the used SM references in terms of correlation values. The AMSR‐reg was more consistent with MERRA‐land than AMSR‐LPRM in terms of unbiased RMSE (unbRMSE) particularly over high latitude regions with a global average of unbRMSE of 0.059 and 0.098 m3/m3, respectively. In conclusion, the statistical regression appears to be a promising approach for merging the SSM data sets retrieved from both AMSR‐E and SMOS in the framework of the ESA CCI project.