Long time series of soil moisture retrieved from AMSR‐E and SMOS Observations

Introduction: Soil moisture (SM) represents less than 1/10000 of the total water of our planet but it plays an important role as it affects the water and energy exchanges at the land surface/atmosphere interface and it is the reservoir of water for agriculture and vegetation in general. SM has been endorsed by the Global Climate Observing System (GCOS) as an Essential Climate Variable. In order to use SM information for climate modeling, SM datasets spanning long time periods are needed. In the context of the European Space Agency (ESA) Climate Change Initiative (CCI) several strategies have been evaluated to merge SM datasets from different microwave sensors (Owe et al. 2008). These strategies consist typically in merging a posteriori several SM datasets computed with different algorithms applied to data from different sensors. In addition, they do not include data from the Soil Moisture and Ocean Salinity (SMOS) satellite (Kerr et al. 2001), which is the first mission specifically designed to retrieve SM from space. Therefore, in the context of an ESA funded project, several approaches have been studied to add SMOS data to long term SM datasets. In a first phase, three different approaches are tested to merge ESA SMOS and NASA/JAXA Advanced Scanning Microwave Radiometer (AMSR‐E). This paper is devoted to a method using statistical retrieval algorithms to compute a priori a long time dataset that is consistent along time by construction. Methods Rodriguez‐Fernandez et al. (2015) have shown the good performances of a SM retrieval from SMOS observations using neural networks trained with ECMWF numerical weather prediction models (Balsamo et al. 2009). In addition, the same approach but using SMOS L3 SM as reference has proven to be useful to develop a Near‐Real‐Time SM retrieval algorithm (Rodriguez‐Fernandez et al. this conference). This study uses both approaches to obtain larger SM time series adding AMSR‐E as input: i) Using a Land Surface model as reference, two neural networks algorithms have been defined and optimized using AMSR‐E or SMOS as input data in the periods 2003‐Oct 2011 and 2010‐2014, respectively. ii) An alternative approach which is independent of land surface models has also been studied. It consist in using SMOS L3 SM retrievals as reference to train a NN using as input AMSR‐E Tb's. Results The best input data to retrieve SM using ASMR‐E data as input is using the three lower frequency bands and two polarizations. Soil temperature and a vegetation index improve the ability of the NN to capture the SM variability of the ECMWF model simulations and SMOS L3 SM. The NN performances are higher when using Tb's as input instead of the polarization index ( PI =(Tb^V ‐ Tb^H)/(Tb^V + Tb^H) ) used by the LPRM algorithm (Owe et al. 2001) or other NNs algorithms (Santi et al. 2012). As shown in Rodriguez‐Fernandez et al. (2015) for the SMOS case, the performance of the NN using AMSR‐E as input data and trained with ECMWF SM simulations improves when using soil texture maps as input (clay and sand fractions). For instance when using the lowest frequency channel, R increases from 0.79 to 0.83. In contrast, when more Tb's measured at higher frequencies are added as input, the contribution of NDVI decreases. When using all the frequency bands, including the 89 GHz channel, the NDVI contribution is negligible and the correlation of NN SM and ECMWF SM is as high as 0.9 (only using AMSR‐E plus soil texture information as input). For comparison, the global score obtained using SMOS Tb's from 7 angle bins from 25deg to 65deg, soil texture and NDVI is R = 0.88 (NN SM with respect to ECMWF SM). In conclusion, both NNs using AMSR‐E or SMOS data as input exhibit similar performances to capture the SM variability in the ECMWF models. Discussion In order to check the consistency of the two NN SM datasets, the trained NNs have been applied to AMSR‐E and SMOS \Tb's in the period from June to September 2010, which has not been used for the training of the NNs. The two maps are very similar, probing that the produced SM is consistent using SMOS or AMSR‐E as input. In addition, the performances of the NN retrieved SM have been evaluated against more than 1100 measurements from in situ sensors in America, Europe, Africa and Australia in the 2003‐2013 period. Results show that, on average, remote sensing retrievals performances in Europe are below those of ERA‐Interim/Land and MERRA‐Land models. In contrast, the remote sensing retrievals performances with respect to in situ measurements in other locations are, on average, similar or better than those of the land surface models. Therefore, this is a promising method to compute long time series of SM that can be used for hydrological and climate application and it is the first step towards a longer dataset which will include additional sensors