Impact of rescaling approaches on the simple fusion of soil moisture products

Authors: ASHFAR, M.H. - Collaborator; YILMAZ, M.T. - Collaborator; Crow, Wade

Submitted to: Water Resources Research
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 9/1/2019
Publication Date: 9/30/2019
Citation: Ashfar, M., Yilmaz, M., Crow, W.T. 2019. Impact of rescaling approaches on the simple fusion of soil moisture products. Water Resources Research. 55:7804-7825. https://doi.org/10.1029/2019WR025111. 2019.
DOI: https://doi.org/10.1029/2019WR025111. 2019

Interpretive Summary: Due to recent advances in remote sensing and modelling, soil moisture products are now available from a range of different satellite-based sensors and modelling systems. These products are valuable for important agricultural applications including: drought monitoring, irrigation scheduling and optimizing fertilizer usage. However, to maximize the accuracy, sampling frequency and historical availability of these data sets, strategies are needed to optimally merge concurrent soil moisture products acquired from different sources. This manuscript compares mathematical strategies for merging multi-source soil moisture products and makes specific recommendations regarding the effectiveness of each strategy. Applying the approaches recommended by this paper will eventually improve the quality of soil moisture data products available to inform decision support for agricultural water management applications.

Technical Abstract: Soil moisture plays a key role in weather forecasting, hydrologic modeling, climate change studies and water resource management. There are multiple ways to estimate this essential variable (i.e., remote sensing, modeling, station-based observations) and clear benefits associated with merging independent estimates of it. However, the time series of these products generally contain systematic differences that must be removed through rescaling before the application of data merging approaches (e.g., data assimilation or data fusion). Here, performance differences between various rescaling approaches in the framework of data fusion are explored. Pairwise combinations of four different soil moisture products: Advanced Scatterometer (ASCAT), Advanced Microwave Scanning Radiometer for EOS (AMSR-E), Antecedent Precipitation Index (API), Global Land Data Assimilation System (GLDAS) NOAH, are fused via an equal weighting data fusion scheme and the application of multiple re-scaling approaches. The rescaling approaches applied here utilize different: 1) methods (i.e., variance matching-, regression-, CDF matching-, MARS-, and SVR-based), 2) stationarity assumptions (i.e., constant or time-varying rescaling coefficients), and 3) techniques (i.e., using periodic or non-periodic high and low frequency components) to remove systematic differences between soil moisture products. Given that statistical measures of reference datasets (e.g., their standard deviation, correlation coefficient, etc.) are utilized in these rescaling approaches, the accuracy of the selected reference dataset also impacts the data fusion results via sampling errors in these statistics. Accordingly, experiments are performed using six different soil moisture product pairs, five different rescaling methods, three different rescaling techniques, two different rescaling stationarity assumptions, and four different reference datasets over four United States Department of Agriculture (USDA) Agricultural Research Service (ARS) watershed sites between 2007 and 2011 (for a grand total of 6*5*3*2*4*4=2880 experiments). Results reveal that application of a smooth-deviance decomposition rescaling technique (i.e., the rescaling techniques using non-periodic low frequency component) produces, on average, a correlation improvement of 0.03 [-] against the validation data compared to the other widely implemented rescaling approaches.