Remote Sensing Retrieval Of Soil Moisture Under Different Vegetation Conditions In A Typical Steppe

Soil moisture（SM）plays an important role in global water cycle and vegetation growth.The spatiotemporal dynamics of SM have a great influence on ecosystem functions and processes especially for plant growth.Inner Mongolia grassland located in arid and semi-arid areas of the Mongolian Plateau is an important component in temperate grassland of northern China.Soil Water content is the main limiting factor for plant growth.Timely and accurate acquisition of soil moisture is helpful in assessment of the plant growth status and grassland productivity.This study conducted soil moisture retrieval research using data from field surface measurement and Sentinel remote sensing in the typical steppe of Xilingol,Inner Mongolia.Based on the grazing control experimental platform,the soil moisture retrieval method is studied by using the measured surface data and Sentinel remote sensing data.By analyzing the influence of vegetation parameters and soil roughness parameters on radar backscattering coefficient under different grazing gradients,we constructed water cloud model（WCM）,bare soil model（AIEM,Oh）and machine learning soil moisture retrieval model.The major results are as follows:（1）The vegetation NDVI index showed a strong positive correlation with the measured aboveground biomass,which was susceptible to the influence of grassland utilization intensity and seasonal dynamics.It thus could be used as a proxy index for biomass.NDVI,NDWI₁ and NDWI₂ could be considered as proxy indicators of vegetation canopy water content in that they were significantly correlated with the measured vegetation canopy water content（VWC）.The VV polarized backscattering coefficient had a very significant positive correlation with the measured soil water content,which can effectively invert the soil water content of typical grassland.（1）The vegetation NDVI index showed a strong positive correlation with the measured aboveground biomass,which was susceptible to the influence of grassland utilization intensity and seasonal dynamics.It thus often be used as a proxy index for biomass in grasslands.NDVI,NDWI₁ and NDWI₂ could be considered as proxy indicators of vegetation canopy water content because that they were significantly correlated with the measured vegetation canopy water content（VWC）.The VV polarized backscattering coefficient had a very significant positive correlation with the measured soil water content,which can effectively invert the soil water content in the typical steppe.（2）The parameterization WCM model from NDVI and NDWI had universal applicability to the retrieval of soil moisture in typical steppe.The constructed parameterization schemes by NDVI and NDWI all passed the accuracy evaluation.However,the training and testing performance of WCM model is affected by the difference in plant community among grazing gradients and growing seasons as well as the vegetation index selected under different vegetation conditions.The results showed that the applicability of the model was segmented.NDVI was more suitable for NG,MG,HG under different grazing gradients and EGS in different growth stages,NDWI₂ was more suitable under LG,and NDWI₁ was more suitable for MGS-I and MGS-II.The accuracy of WCM model is highest when the aboveground biomass was more than 100 g/m² and the height was higher than 17 cm.（3）The dual-polarization empirical model of bare soil moisture retrieval was constructed by Oh model and AIEM model based on the effective combination of roughness parameters constructed in this paper,which improved the accuracy of soil moisture retrieval.The R² of VV and VH polarization increased from 0.1031 and0.0547 to 0.7585 and 0.6809,respectively（p<0.001）.Bare soil model is suitable for severely degraded grassland due to its lower vegetation height and biomass.（4）Four machine learning models,including random forest（RF）,extreme gradient boosting（XGB）,convolutional neural network（CNN）and genetic algorithm optimized BP neural network（GA-BP）,were combined with Boruta and RFE to construct feature parameter data sets.A total of eight soil moisture retrieval models were constructed to estimate soil moisture in the typical steppe.The XGB retrieval model driven by RFE was constructed and validated through measured data,which outperformed other models in terms of performance.The retrieval accuracy R²,RMSE and MAE were 0.7050,0.0321 and 0.0249,respectively,indicating that the model was robustness and could be used to invert soil moisture in a typical steppe.The results of this study also proved that the constructed XGB model can be applied to soil moisture retrieval at regional scale.The results of this study suggest that soil moisture inversion is mainly affected by surface vegetation parameters in the typical steppe.The double-layer scattering vegetation model is applicable for grassland with high vegetation coverage,while the single-layer bare soil scattering model is applicable for grassland with sparse vegetation.The vegetation influence must be considered when retrieving soil moisture in grasslands.Different parameters and models should be selected according to plant growth stage,utilization intensity and vegetation status.The results of this paper enrich the theory and method on soil moisture retrieval,which provide data and decision support for grassland scientific management,drought monitoring,drought disaster assessment and grassland insurance.