The Effect of Three Different Data Fusion Approaches on the Quality of Soil Moisture Retrievals from Multiple Passive Microwave Sensors

Long-term climate records of soil moisture are of increased importance to climate researchers. In this study, we aim to evaluate the quality of three different fusion approaches that combine soil moisture retrieval from multiple satellite sensors. The arrival of L-band missions has led to an increased focus on the integration of L-band-based soil moisture retrievals in climate records, emphasizing the need to improve our understanding based on its added value within a multi-sensor framework. The three evaluated approaches were developed on 10-year passive microwave data (2003–2013) from two different satellite sensors, i.e., SMOS (2010–2013) and AMSR-E (2003–2011), and are based on a neural network (NN), regressions (REG), and the Land Parameter Retrieval Model (LPRM). The ability of the different approaches to best match AMSR-E and SMOS in their overlapping period was tested using an inter-comparison exercise between the SMOS and AMSR-E datasets, while the skill of the individual soil moisture products, based on anomalies, was evaluated using two verification techniques; first, a data assimilation technique that links precipitation information to the quality of soil moisture (expressed as the Rvalue), and secondly the triple collocation analysis (TCA). ASCAT soil moisture was included in the skill evaluation, representing the active microwave-based counterpart of soil moisture retrievals. Besides a semi-global analysis, explicit focus was placed on two regions that have strong land–atmosphere coupling, the Sahel (SA) and the central Great Plains (CGP) of North America. The NN approach gives the highest correlation coefficient between SMOS and AMSR-E, closely followed by LPRM and REG, while the absolute error is approximately the same for all three approaches. The Rvalue and TCA show the strength of using different satellite sources and the impact of different merging approaches on the skill to correctly capture soil moisture anomalies. The highest performance is found for AMSR-E over sparse vegetation, for SMOS over moderate vegetation, and for ASCAT over dense vegetation cover. While the two SMOS datasets (L3 and LPRM) show a similar performance, the three AMSR-E datasets do not. The good performance for AMSR-E over spare vegetation is mainly perceived for AMSR-E LPRM, benefiting from the physically based model, while AMSR-E NN shows improved skill in densely vegetated areas, making optimal use of the SMOS L3 training dataset. AMSR-E REG has a reasonable performance over sparsely vegetated areas; however, it quickly loses skill with increasing vegetation density. The findings over the SA and CGP mainly reflect results that are found in earlier sections. This confirms that historical soil moisture datasets based on a combination of these sources are a valuable source of information for climate research.