| Soil moisture is a key parameter in the study of agriculture,hydrology and climate sys-tems,and plays an important role in the material exchange and energy cycle between surface and atmosphere.Especially in arid and semi-arid areas,soil water information has an im-portant impact on crop growth,water-saving irrigation,water and soil loss process,etc.How to timely obtain accurate temporal and spatial continuous soil water information has been the focus of scholars at home and abroad.At present,soil moisture information acquisition methods mainly include ground observation,model simulation and remote sensing monitor-ing.These three methods have their own advantages and limitations.Therefore,how to in-tegrate different soil moisture monitoring methods to complement each other’s advantages and effectively improve the prediction accuracy of soil moisture is a topic worth studying.Therefore,in this paper,20,000 mu of salinized farmland in Wuyuan County,Hetao irriga-tion District of Inner Mongolia is taken as the research area,and the artificial intelligence model of surface soil water inversion is constructed by using the soil backscattering coeffi-cient after removing the influence of vegetation cover by water cloud model.The HYDRUS-1D model was used to simulate soil moisture at different depths,and the simulation results were calibrated and verified.Surface remote sensing soil water inversion values and meas-ured soil water observation values were respectively taken as observation operators.The Ensemble Kalman Filter（EnKF）was applied to introduce the soil water observation values into the HYDRUS-1D model to construct soil water content data assimilation schemes of different depths,and the sensitivity analysis of the assimilation process was carried out.The optimal data assimilation scheme is determined.The main conclusions of the study are as follows:（1）Two inversion models of surface soil water content,BP neural network and RBF neural network,were constructed using the soil backscattering coefficient obtained after re-moving the influence of vegetation cover by water cloud model.The inversion results showed that:The R2 between predicted soil moisture content and measured soil moisture content of RBF neural network model is 0.7978,which is 0.1029 higher than that of BP neural network model.The RMSE was 0.020cm3·cm-3,which decreased by 0.006cm3·cm-3compared with BP neural network model.This indicates that the RBF neural network model is more suitable for soil water inversion research in the study area,and its inversion predicted value can be used as observation operator for assimilation simulation of surface soil water remote sensing data.（2）Based on the meteorological data and measured soil moisture data in the study area,the HYDRUS-1D model was used to simulate the soil moisture content and clarify the changes of soil moisture in the vertical direction.The RMSE of model simulation values at0～20cm,20～40cm,40～60cm and 60～80cm depth were 0.020～0.056cm3·cm-3,0.017～0.043cm3·cm-3,0.018～0.050cm3·cm-3,0.020～0.046cm3·cm-3,respectively.The results showed that the HYDRUS-1D model had a good simulation effect on soil water content.（3）A soil water data assimilation scheme based on EnKF was proposed.After sensitiv-ity analysis,it was determined that when the set number N was 100,the soil water assimila-tion effect was the best.The observation operator of data assimilation was surface remote sensing soil water inversion value and measured soil water observation value.The assimila-tion algorithm was EnKF algorithm,and the simulation operator of assimilation was the Hydraus-1D model.（4）Compared with the results of HYDRUS-1D model alone,the relative errors,root mean square errors and mean absolute errors between the analyzed and observed values of water assimilation in each soil layer at fixed observation stations were reduced to 0.025～0.063cm3·cm-3,0.010～0.017cm3·cm-3,0.008～0.01cm3·cm-3,respectively.It shows that the data assimilation method can effectively improve the simulation accuracy of soil water.The assimilation effect of 0～20cm soil was the best,followed by 20～40cm soil,and the effect of 40～80cm soil was poor.The simulation accuracy of the analysis value is better than that of the assimilation prediction value,and the simulation accuracy of the assimilation prediction value is better than that of the HYDRUS-1D prediction value.At the regional scale,the determination coefficient R2 between the assimilated soil moisture content and the measured soil moisture content at the 17 sampling points in different months ranged from0.6313 to 0.8747.The results of regional assimilation reflected the spatial distribution of soil moisture content in the study area,which was lower in the north and higher in the south,and the soil moisture content increased gradually with the increase of soil depth.Soil moisture content showed an increasing trend from April to September.The results show that the data assimilation scheme in this paper can effectively improve the simulation accuracy of soil water content,and has a certain reference value for monitoring the temporal and spatial dis-tribution of soil water information. |
