Development of a high-resolution land data assimilation system with integrated machine learning

Land data assimilation systems, by assimilating land surface remote sensing observations, such as soil moisture (SM) products from SMAP, SMOS, and AMSR2, and combining the advantages of the land surface model, are able to produce spatiotemporally seamless data on the state of the land surface, including soil moisture and temperature in the surface layer and rooting zone, as well as energy fluxes at the surface. However, because of the coarse resolution of prevailing passive microwave soil moisture remote sensing products as well as the lesser accuracy of high-resolution soil moisture products, there is no high-resolution land data assimilation system available.

In this study, we developed a high-resolution land data assimilation system by using a machine learning algorithm in combination with a dual-cycle assimilation system. We first used random forests to generate high-resolution soil moisture products from passive microwave soil moisture, and then used the dual-cycle assimilation system to correct the bias of the soil moisture products and assimilated them into the land surface model, and finally produced high-resolution land surface state datasets.  The high-resolution assimilation system was validated on observations from three soil moisture observation networks on the Tibetan Plateau. The results show that the system is capable of producing reliable soil moisture products at different resolutions, such as 5 km, 10 km, etc., with ubRMSE less than 0.06m³/m³.