Abstract: Abstract: Accurate and real-time information on soil salinity is required to understand the evolution of soil salinization, to develop appropriate management strategies, and to implement practices to improve the soil productivity and ecological restoration. Therefore, describing the accurate process of soil salt transport is of great significance for the precise management of salt-affected soils. Using the proximal soil sensor (electromagnetic induction, type EM38) and ensemble Kalman filter (EnKF) method, this study investigated the feasibility of soil salinity estimation by assimilating 1-D hydrological model (HYDRUS-1D) and apparent electrical conductivity data measured by EM38. Soil sampling and periodical EM38 survey at 11 dates was performed in the experimental site, located in a marine-terrestrial interlaced area in north Jiangsu Province. Soil physical and chemical properties, groundwater attributes and meteorological data were also collected as driving data of assimilation system during November 2015 and October 2016. The inversion model relating apparent electrical conductivity to soil salinity was adopted as observation operator, and EnKF method was applied to HYDRUS-1D model to simulate soil salinity on the profile. This study also examined the sensitivity of simulation accuracy to ensemble number, error level and number of soil salinity observation data during assimilation procedure. The main conclusions included: 1) EnKF assimilation method improved the simulation accuracy of soil salinity on 0-1 m profile. In comparison with simulated value after EnKF assimilation and HYDRUS-simulated value, the root mean square error of EnKF assimilation value decreased and the NSE of EnKF assimilation value increased. This indicated that EnKF assimilation value was more accurate than the simulated value after EnKF assimilation, whereas the simulated value after EnKF assimilation was better than HYDRUS-simulated value; 2) The simulated value after EnKF assimilation was closer to the measured value than HYDRUS-simulated value, and the soil salinity adjustment of EnKF assimilation value was greater than those of simulated value after EnKF assimilation and HYDRUS-simulated value during the assimilation procedure. In general, the adjustment amount of soil salinity was small when the temporal dynamics of soil salinity was flat for EnKF assimilation, whereas drastic soil salinity dynamics resulted in the increase of adjustment amount, for instance the soil salt leaching in the rainy season, indicating the improvement of simulation accuracy. The difference between simulated value after EnKF assimilation and measured value varied from (0.137 to 0.227 g/kg, with an average of 0.097 g/kg, whereas the difference between HYDRUS-simulated value and measured value ranged between 0.082 and 0.437 g/kg, with an average of 0.289 g/kg. 3) Soil salinity assimilation was not sensitive to ensemble size, whereas the error level and number of observation data were sensitive to soil salinity assimilation. The EnKF simulation result of soil salinity for ensemble size 75 was similar to that for ensemble size 100, and no further improvement was observed when ensemble size increased to 75. Also, no further improvement occurred when the error level of observation data exceeded 10% during the EnKF assimilation. Generally, high error level and low involved number of observation data in the EnKF assimilation resulted in large deviation, and vice versa. The improvement of surface soil observation data to the simulation of deep soil salinity attenuated with the increase of soil depth, and the assimilation of deep soil salinity was more sensitive to the involved number of observation data than that of surface soil salinity. It was concluded that, using the ensemble Kalman filter method, the coupled application of HYDRUS model and apparent electrical conductivity data improved the simulation performance of soil salinity. This study provided an effective way for the prediction of large scale ecological processes using multi-source data and mechanism model. More efforts should be diverted to the integration and assimilation of multi-source data at larger scales, such as the proximally sensed data and remote sensing data, and the optimization of model parameter and observation data acquisition frequency.