Multi-source soil moisture data fusion based on high-resolution land surface simulation and machine learning

Soil moisture (SM) plays an important role in hydrological processes and land-atmospheric interactions, and serves as an important boundary condition for the weather forecasting and climate modeling. Influenced by global environment change, SM changes significantly at local scales which rises the great need of high-resolution SM products to provide locally relevant information. However, the three SM estimation approaches, namely in-situ observation, remote sensing retrieval and land surface modeling, all have their disadvantages. Although recent works produce a combined SM products by merging the in-situ observations and several land surface simulation products, the long-term high-resolution SM product integrating multivariate data including remote sensing products is still lacking. In this study, high-resolution land surface modeling, high-resolution remote sensing products and SM observations from more than 2000 stations will be combined to generate spatially continuous and temporally complete soil moisture data in China by using the random forest algorithm. We first performed land surface simulations by using the Conjunctive Surface-Subsurface Process version 2 (CSSPv2) model forced by three meteorological forcings including the China Meteorological Administration Land Data Assimilation System version 2.0 (CLDASv2.0), ERA5 and GLDASv2.1. The validations over 2090 in situ stations during 2012–2017 showed that CLDASv2.0/CSSPv2 soil moisture simulation performed better than ERA5 and GLDASv2.1 reanalysis products, with an increased correlation of 26%–68% and reduced errors of 14%–24% at the daily time scale. The improvements mostly originate from the use of an advanced LSM because CLDASv2.0/CSSPv2 only increased the correlation by 5%–35% and decreased the errors by up to 9% when compared with ERA5/CSSPv2 and GLDASv2.1/CSSPv2. Due to the high accuracy of CLDASv2.0/CSSPv2 product, it will be used as a background to fuse the in-situ observations and satellite remote sensing soil moisture. The 70% of the observation site data, remote sensing products and CLDASv2.0/CSSPv2 product will be used to train the random forest model and generate a high resolution soil moisture product from 2008 to 2017, and another 30% of the site data will be used to evaluate the accuracy of the results. Such a SM product can describe the spatial and temporal distribution characteristics of soil moisture heterogeneity more accurately, and thus provide sufficient data support for scientific research and social development.