Land data assimilation (DA) is an effective method to provide high-quality spatially and temporally continuous soil moisture datasets that are crucial in weather, climate, hydrological, and agricultural research. However, most existing land DA applications have used remote sensing observations, and are based on one-dimensional (1D) analysis, which cannot be directly employed to reasonably assimilate the recently expanded in-situ soil moisture observations in China. In this paper, a two-dimensional (2D) localized ensemble-based optimum interpolation (EnOI) scheme for assimilating in-situ soil moisture observations from over 2200 stations into land surface models (LSMs) is introduced. This scheme uses historical LSM simulations as ensemble samples to provide soil moisture background error covariance, allowing the in-situ observation information to be propagated to surrounding pixels. It is also computationally efficient because no additional ensemble simulations are needed. A set of ensemble sampling and localization length scale sensitivity experiments are performed. The EnOI performs best for in-situ soil moisture fusion over China with an ensemble sampling of hourly soil moisture from the previous 7 days and a localization length scale of 100 km. Following the evaluation, simulations for in-situ soil moisture fusion are also performed from May 2016 to September 2016. The EnOI analysis is notably better than that without in-situ observation fusion, as the wet bias of 0.02 m³ m⁻³ is removed, the root-mean-square error (RMSE) is reduced by about 37%, and the correlation coefficient is increased by about 25%. Independent evaluation shows that the EnOI analysis performs considerably better than that without fusion in terms of bias, and marginally better in terms of RMSE and correlation.