A triple collocation-based 2D soil moisture merging methodology considering spatial and temporal non-stationary errors

Authors: ZHOU, J. - Hohai University; Crow, Wade; WU, Z. - Hohai University; DONG, J - US Department Of Agriculture (USDA); HE, H. - Hohai University; FENG, H. - Hohai University

Submitted to: Remote Sensing of Environment
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 5/12/2021
Publication Date: 5/31/2021
Citation: Zhou, J., Crow, W.T., Wu, Z., Dong, J., He, H., Feng, H. 2021. A triple collocation-based 2D soil moisture merging methodology considering spatial and temporal non-stationary errors. Remote Sensing of Environment. 263:112509. https://doi.org/10.1016/j.rse.2021.112509.
DOI: https://doi.org/10.1016/j.rse.2021.112509

Interpretive Summary: Within the past decade, a large number of model-based and remotely sensed surface soil moisture products have been developed and distributed. These products are of potential value for a range of agricultural applications including drought monitoring, yield forecasting, and irrigation scheduling. However, in order to maximize the accuracy and length of available soil moisture data sets, strategies must be developed to optimally merge concurrent soil moisture products acquired from multiple sources. This manuscript presents a novel mathematical strategy for obtaining reliable error statistics for independent soil moisture products acquired from a variety of sources. We then apply these statistics to merge multiple soil moisture products into a single, optimized estimate that is superior to any of the underlying products it is based on. Application of this strategy will eventually improve the quality of soil moisture data products available to inform agricultural water management.

Technical Abstract: Random error in remotely sensed (RS) and modeled soil moisture (SM) products is typically as-summed to be statistically stationary for the purpose of SM merging and land surface data assimilation (DA) applications. In fact, such error is commonly non-stationary which may undermine applications based on a stationary assumption. Here, we introduce an ergodic time-space (2D) SM merging approach that considers a class of inferred error that is non-stationary and undetectable in time (or space) but stationary and detectable in space (or time). Such 2D merging is realized in a least-squares framework where spatial and temporal error variances for each product are estimated via triple collocation. As a test case, a 2D-merged SM product is obtained by combing two independent RS SM products and one mod-eled SM product. The method and the resulting merged product are evaluated using both synthetic experiments and independent in-situ observations acquired from a dense SM ground network in the Huang-Huai-Hai River Basin of China. Results show that inferred nonstationary error can be effectively filtered by the 2D merging method to produce a merged SM product that is superior, with regards to both SM spatial patterns and time series precision, to that produced by classical 1D merging methods. By providing a more effective way to detect and remove non-stationary errors during the process of merging multi-source SM products, this approach is meaningful for improving current SM merging applications.