Soil Moisture Retrieval In Farmland Surface Based On Sentinel Multi-source Remote Sensing Data

As an important component of the earth ecosystem,soil moisture is of great significance in the fields of crop growth monitoring,crop yield estimation,variable irrigation and other related applications.With the rapid development of the technology and theory of microwave remote sensing,soil moisture inversion with remote sensing data has been widely used at home and abroad.The multi-source remote sensing data for quantitatively retrieving soil moisture in this study is acquired from Sentinel-1 radar and Sentinel-2 optical satellites,belonging to ESA’s Sentinel series,so there are great advantages in space,time and data registration in monitoring soil moisture.To deal with the problem that soil moisture retrieval is greatly affected by surface vegetation covers,this study firstly applies Oh model to retrieve soil moisture after removing the influence of vegetation by water cloud model.Secondly,taking the great advantages of machine learning algorithms into account,this study takes various vegetation indexes as inputs for support vector regression（SVR）and generalized regression neural network（GRNN）models,without separating vegetation scattering from soil scattering.The vegetation indexes are calculated out with Sentinel-2 optical remote sensing data,including normalized difference vegetation index（NDVI）,modified soil adjusted vegetation index（MSAVI）and difference vegetation index（DVI）.In order to excavate the feature information of satellite data in soil moisture retrieval,the intensity information and phase information of active microwave data are fused,and then the polarization decomposition feature is analyzed on the soil moisture inversion based on the CNN model.The main conclusions are drawn as follows:（1）In order to solve the problem of vegetation cover in farmland surface soil moisture retrieval,water cloud model was used to remove the effects of vegetation.The experimental results show that the soil moisture inversion accuracy with Oh model after removing vegetation influence by water cloud model has been increased,as we expected.After removing the vegetation,the R² of the test set increased by 0.0297and the RMSE decreased by 0.0022 cm³/cm³.It can be seen that the effect of soil moisture retrieval is improved after removing the vegetation.（2）In view of the limitations of traditional soil moisture retrieval models,this paper constructs soil moisture inversion models based on SVR and GRNN.The inversion accuracies of SVR model with MSAVI and NDVI are higher than that of Oh model,and the retrieval effect of GRNN model is better than the Oh model when the input contains H₀ and VV parameters.The optimal input combination of SVR model composed of five characteristic parameters,including dual-polarization radar backward scattering coefficients（i.e.,VV and VH）,altitude,local incident angle,and MSAVI achieves the best inversion accuracy with correlation coefficient of 0.8497 and root mean square error of 0.0214cm³/cm³ respectively.Compared with the GRNN optimal combination,the R² of the SVR optimal combination is increased by 0.0682,and the RMSE is reduced by 0.0031 cm³/cm³.Compared with the results of the Oh model after removing the vegetation,the R² of the optimal combination of SVR increased by 0.2229,and the RMSE decreased by 0.0081 cm³/cm³.（3）Since retrieved farmland soil moisture could be influenced by the dual-polarization radar backward scattering coefficient,altitude,local incident angle and vegetation index,this study defines the equivalent number to evaluate the quantitative influence of each characteristic parameter based on the soil moisture inversion results.The results verify the important effects of radar backscatter coefficient,altitude,local incidence angle,and vegetation index on the retrieval of farmland surface soil moisture.Among the three vegetation indexes,MSAVI has the strongest correlation with soil moisture content,followed by NDVI and DVI is the weakest.（4）To verify the effect of SAR phase information on soil moisture inversion,a model based on CNN was constructed.By comparing and analyzing the inversion results of the CNN model,it was found that when the training set ratio test set is 1:1,3:1 and 5:1 respectively,compared with the non-polarized decomposition feature CNN,the R² of CNN model with the combined polarization decomposition feature test set is increased by 0.0291,0.0307 and 0.0337,RMSE decreased by 0.0126,0.0064 and0.007cm³/cm³.The results showed that the CNN model had better performance with the increase of the training sample size and fusing the intensity and phase information of SAR data could improve the accuracy of soil moisture inversion.