Retrieval Of Surface Soil Moisture Covered With Winter Wheat Based On RADARSAT-2 SAR Data

Soil moisture content(SMC)is the key parameter of land surface ecological water cycle,which controls surface runoff and surface water evaporation.In agricultural production,soil moisture is not only an important source for crop growth and development,but also a key factor affecting soil fertility.Therefore,farmland soil moisture can be used as an effective index to reflect farmland health status.The acquisition of large-scale farmland soil moisture can effectively guide the production of crops.Remote sensing technology provides a new technical support for the accurate monitoring of large-scale soil moisture.Microwave remote sensing has strong penetration,the ability of all time and all-weather monitoring,and the high sensitivity of microwave signal to soil moisture,which makes microwave remote sensing become the main means of remote sensing to monitor soil moisture.At present,the scattering models based on bare soil have been widely used in soil moisture inversion,but in the vegetation covered area,the inversion of soil moisture is more difficult,mainly for the following two reasons:Firstly,the use of microwave scattering model is limited by the lack of surface roughness,vegetation coverage and other parameters that are difficult to measure;Secondly,the existence of vegetation will attenuate the surface microwave signal and produce complex volume scattering,which reduces the sensitivity of microwave signal to soil moisture.In order to solve the above problems that may exist in the inversion of soil moisture on the surface covered by winter wheat,this study took the winter wheat planting area in Ontario,Canada as the study area,took the fully polarized RADARSAT-2 SAR data as the main remote sensing data source,and combined the ground survey data and optical remote sensing data to construct a variety of soil moisture estimation models on the surface covered by winter wheat based on RADARSAT-2 SAR data.The completed work mainly included the following parts:(1)Two commonly used semi-empirical models(Ratio method and water cloud model(WCM))were used to characterize the scattering features and to separate the scattering contribution of vegetation.Then,a SMC estimation method based on semi-empirical models and optimal roughness parameter was constructed based on the calibrated integrated equation model(CIEM).CIEM was used to simulate the backscattering coefficient of bare soil,and the initial reference incidence angle was set to30°.LAI,RVI and NDVI were used to parameterize semi-empirical models,respectively.Based on the scattering coefficient estimated by the semi-empirical model and optimal roughness parameter,soil moisture can be retrieved by the look-up table constructed by CIEM.The results show that optimal roughness and appropriate reference incidence angle can effectively improve the estimation accuracy of SMC.The highest estimation accuracy can be obtained when NDVI was used to parameterize the ratio method and WCM.The coefficient of determination(R~2)of the estimation results were 0.68 and 0.66 respectively,and the root mean square error(RMSE)were 4.15 vol.%and 4.27 vol.%respectively.(2)Based on Freeman-Durden decomposition method,the vegetation scattering contribution was separated from the original SAR signal to obtain the surface scattering component,and a SMC estimation method based on polarimetric SAR decomposition was constructed.Firstly,H-A-alpha decomposition and Freeman-Durden decomposition were used to analyze the scattering mechanism in the study area,ignoring the dihedral scattering component with a small proportion.Then,different volume scattering matrices were used to separate the volume scattering contribution to obtain the backscattering coefficient corresponding to the surface scattering component.Finally,the SMC can be obtained from the CIEM and Dubois model combined with the optimal roughness parameters by further establishing a look-up table.The results show that the estimation accuracy based on Dubois model was generally higher than that of CIEM,with R~2 and RMSE were 0.63 and 5.16 vol.%respectively,and R~2 and RMSE based on CIEM were0.53 and 5.62 vol.%respectively.(3)A variety of polarization feature parameters were extracted from the original SAR image by using various polarization decomposition methods to expand the SAR data feature space.Combining with support vector machine(SVM),gradient boosting regression tree(GBRT),random forest(RF)and three feature selection methods,which are based on correlation coefficient,based on SVM recursive feature elimination(SVM-RFE)and the importance of RF factor,a SMC estimation model based on machine learning was constructed.The results show that the feature selection method based on the importance of RF factor and SVM-RFE was effective,and the models can achieve the best estimation accuracy by using a small number of features.The three models achieved the best SMC estimation performance with the feature selection of RF factor importance,and the estimation accuracy of RF model was the highest,the R~2 of the estimation result was 0.79 and the RMSE was 4.03 vol.%.