A robust data-worth analysis framework for soil moisture flow by hybridizing sequential data assimilation and machine learning

Abstract

As the collection of soil moisture data is often costly, it is essential to implement data-worth analysis in advance to obtain a cost-effective data collection scheme. In previous data-worth analysis, the model structural error is often neglected. In this paper, we propose a robust data-worth analysis framework based on a hybrid data assimilation method. By constructing Gaussian process (GP) error model, this study attempts to alleviate biased data-worth assessments caused by unknown model structural errors, and to excavate complementary values of multisource data without resorting to multiple governing equations. The results demonstrated that this proposed framework effectively identified and compensated for complex model structural errors. By training prior data, more accurate potential observations were obtained and data-worth estimation accuracy was improved. The scenario diversity played a crucial role in establishing an effective GP training system. The integration of soil temperature into GP training unraveled new information and improved the data-worth estimation. Instead of traditional evapotranspiration calculations, the direct inclusion of easy-to-obtain meteorological data into GP training yielded better data-worth assessment.