Radar Remote Sensing Research On Soil Moisture In Arid Areas Using Machine Learning And Polarization Feature Optimization

Soil moisture is an important parameter of the earth’s water cycle process,which affects the exchange of energy and water in processes such as ecology,climate,and agriculture.Microwave remote sensing has penetrating characteristics,will not be disturbed by weather factors,can monitor soil moisture in real time,and is highly sensitive to soil moisture.Among them,synthetic aperture radar（SAR）technology has developed rapidly in recent years,from single polarization to multi-polarization and full polarization,and fully polarized synthetic aperture radar（Pol SAR）includes with more polarization information,it has been widely used in research fields such as ground object classification,target recognition,and inversion of ground parameters.However,the polarization scattering characteristics of different ground objects are different,and under different test conditions,the effect of polarization characteristics on soil moisture retrieval is also different.Therefore,it is of great significance to study the potential relationship between polarization characteristic parameters and soil moisture.However,existing studies mostly use the traditional method of simple linear regression to retrieve soil moisture through a limited number of polarization parameters,so it is impossible to comprehensively and systematically measure the role of each polarization characteristic parameter in soil moisture retrieval.Compared with traditional regression methods,machine learning is not limited by the number and types of input parameters,and can learn nonlinear and complex mapping relationships.This paper will use machine learning model combined with polarized SAR characteristic parameters to carry out soil moisture retrieval.This research has important significance and reference value for the ecological environment protection,sustainable economic development and drought monitoring research in Juyanze area.In this paper,Juyanze,a typical arid area,is used as the research area.Based on the fully polarized Radarsat-2 data and combined with the measured data of soil moisture in the research area,a machine learning soil moisture inversion model in arid areas with different combinations of characteristic parameters is constructed to explore the machine learning model.Applicability in arid regions.First,use standard intensity and phase processing to extract the backscattering coefficients,and use target decomposition methods to extract the characteristic parameters contained in the SAR image;then,analyze the correlation and importance of the characteristic parameters,and analyze the correlation and importance of the characteristic parameters.Based on the comprehensive analysis results of normalization and addition,all feature parameters are sorted from high to low,and a number of Random Forest（RF）and Support Vector Machine（SVM）combinations with different parameter combinations are constructed and BP Artificial Neural Network（BP-ANN）models.Ten-fold cross-validation is used to evaluate the performance of the model.The three machine learning models are compared horizontally to select the best parameters for arid areas.Combine machine learning models;finally,select the model with the best performance to invert the soil moisture in Juyanze area,use H-α-Wishart and Freeman-Wishart classification methods to classify land use in the study area,and combine soil moisture inversion the results analyze the spatial distribution of soil moisture in Juyanze.The main research results are as follows:（1）The comprehensive analysis result of the normalized addition method shows that the normalized result value of the entropy H is the highest,which is consistent with the higher correlation between the entropy H and soil moisture and the higher importance score;the normalized addition result has the lowest value Is the single scattering component F_Odd;among the four backscattering coefficients,the normalized value of the cross-polarized componentσ_VHis the highest,the co-polarized componentσ_HH is the second,and the lowest is the cross-polarized componentσ_HV;In the mechanism,the normalized value of even-scattered component is better than single-scattered component and volume-scattered component;（2）In general,the training set R²,RMSE,and MAE of all models are better than the validation set,indicating that the performance of each model in the training set is overall better than the validation set.Regardless of the training set or the validation set,the applicability of the RF model for soil moisture retrieval in this study area is better than that of the SVM model and the BP-ANN model;the training set and validation set of the 41-parameter RF model in all parameter combinations RMSE is not the lowest.The R² and RMSE of the combined SVM model and BP-ANN model training set and validation set are not optimal,indicating that the increase of model input parameters does not mean the improvement of model performance,but also proves that the model The necessity of parameter optimization;from the perspective of R²,RMSE,and MAE of the comprehensive training set and validation set,the RF model with a 20-parameter combination performs best,and the model can explain 91%of soil moisture changes.（3）Based on the difference map of the inversion results,the inversion results are analyzed from the perspective of the scattering mechanism,and it is found that the H single-parameter model,the 5-parameter combination model and the 10-parameter combination model are overestimated near Swan Lake,and the remaining models are underestimated.From a point of view,the models with different combinations of parameters have differences in overestimation and underestimation in Swan Lake and its surrounding areas,and most of the remaining areas are overestimated to varying degrees.Secondly,by comparing the results of Freeman-Wishart classification with the results of H-α-Wishart classification,it is more intuitive to find that the Freeman-Wishart classification is more refined,and the classification results are better,and the boundaries of each category are clear.Combining Kappa coefficient,overall classification accuracy and classification result shows that Freeman-Wishart classification results are better.（4）Combining the inversion results and the land use classification results,it can be found that the soil moisture content in most areas of Juyanze is concentrated below 10%.The soil water content near Swan Lake is higher than the surrounding area;the surface of saline-alkali soil forms a hard salt crust layer,the existence of the salt crust layer can reflect solar radiation to a certain extent,and the soil water evaporation under the salt crust layer is small.Its soil water content is relatively high;sandy soil has large pores and poor soil water holding capacity,resulting in low soil water content;bare soil has relatively small pores and relatively large soil viscosity,and soil water content is slightly higher Sandy.Taken together,the spatial distribution of soil moisture in the study area is mainly related to the surface vegetation coverage,soil texture and soil heat conditions.