Modeling And Inversion Of Soil Moisture In Ejina Oasis Integrating Optics And Radar Remote Sensing

Soil moisture is a key factor affecting the global water cycle and atmospheric energy exchange,and plays a crucial role in ecohydrology,drought monitoring,crop yield estimation and other fields.In northwest China,water resource is scarce and vegetation coverage is low.Accurate acquisition of soil moisture can provide necessary decision-making support for local ecological environment protection and drought monitoring.Polarimetric Synthetic Aperture Radar（Pol SAR）in active microwave remote sensing not only has the characteristics of high spatial resolution,all-weather and all-time imaging,but also can provide favorable polarization characteristic parameters for soil moisture inversion.Vegetation Supply Water Index（VSWI）,Global Vegetation Moisture Index（GVMI）and other traditional spectral indices in optical remote sensing were widely used in soil moisture inversion researches.Meanwhile,thermal infrared band also has high sensitivity to soil moisture.Therefore,multi-source data fusion inversion method integrating optical and microwave remote sensing has attracted much attention in current researches.However,few rearches have combined traditional spectral index with thermal infrared band,and radar data used previously are mostly acquired by C-band SAR,while L-band SAR data is rarely investigated.Besides,traditional regression models show some limitations in the process of multi-source data fusion inversion,machine learning models have stronger learning and fitting capabilities compared to traditional ones,and can be used to evaluate the role of different parameters in soil moisture retrieval accurately and comprehensively.Accordingly,multi-source data fusion and machine learning models based on optical and microwave remote sensing have become the focus of soil moisture inversion researches.In this study,radar,optical,and radar-optical schemes were constructed to retrieve soil moisture in the Ejina oasis in western Inner Mongolia based on L-band ALOS-2 PALSAR-2and Landsat-8 data with the help of machine learning models and water cloud models to explore the effectiveness and applicability of multi-source data fusion and machine learning models in soil moisture retrieval in arid desert oasis regions.Firstly,radar backscattering coefficients were extracted based on ALOS-2 PALSAR-2 data,and vegetation influence was eliminated by water cloud model.The polarization characteristic parameters were extracted based on several polarization decomposition methods including Freeman-Durden,Yamaguchi and so on.The traditional spectral indices and land surface temperature were obtained based on Landsat-8image,and the sensitive thermal infrared band was introduced into traditional spectral indices to generate an improved index.Secondly,the importance of different parameters in soil moisture retrieval were evaluated using the Mean Decrease Accuracy（MDA）.Optimal radar,optical and radar-optical collaborative inversion schemes were designed according to importance ranking to construct three kinds of machine learning models including Least Absolute convergence and Selection Operator（LASSO）,Random Forest（RF）and Support Vector Machine（SVM）.Finally,the inversion accuracy of different schemes and models was compared and analyzed,and soil moisture content in the study area was retrieved by the model with better performance and higher applicability.Soil moisture distribution pattern in this region was further discussed by combining with different land use types.The main conclusions are as follows:（1）According to the processing results of the water cloud model,the attenuation effect of cross polarization is better than that of co-polarization.The attenuation ranges ofσ_HH andσ_VVof the co-polarization backscattering coefficients are 0.03-1.11dB and 0.03-0.98dB,respectively,and the average attenuation values are 0.27dB and 0.26dB,respectively.The attenuation ranges ofσ_VH andσ_HV of cross polarization backscattering coefficients are 0.06-1.59dB and 0.07-1.52dB,respectively,and the average attenuation values are 0.40dB and0.41dB,respectively.The above results indicate that it is very necessary to use water cloud model to remove the influence of vegetation in soil moisture inversion in low vegetation area to obtain a more real surface backscattering coefficient.（2）After adding thermal infrared bands to the traditional spectral indices for improvement,the correlation values of Enhanced Enhanced Vegetation Index（EEVI）and Enhanced Ratio Vegetation Index（ERVI）was significantly improved,with an increase of 0.038 and 0.092,respectively,showing that after improvement the indicating effect of EEVI and ERVI on soil moisture has increased,and the correlation of the remaining indices decreased after the improvement.（3）In radar characteristic parameters,the importance of surface/odd scattering and volume scattering components is relatively high,while the importance of dihedral angular scattering and helix scattering components is relatively low according to the importance scores whether before or after the removal of vegetation impacts.Among the backscattering coefficients,co-polarization backscattering coefficients are always better than those of the cross polarization.As for optical characteristic parameters,VSWI is the most important one,and the importance of the remaining indices changes to varying degrees before and after improvement.（4）Accuracy comparison of different schemes and models shows that the radar scheme is always superior to the optical scheme when using single data source.The R²and RMSE of the radar scheme ranges from 0.363 to 0.700 and 2.160%to 2.760%,respectively,with an average value of 0.513 and 2.515%,respectively.The R² and RMSE of the optical scheme ranges from0.060 to 0.568 and 2.323%to 3.156%,respectively,with an average value of 0.309 and 2.760%,respectively.In the radar-optical collaborative scheme,R² and RMSE of each model ranges from 0.278 to 0.772 and 1.984%to 2.589%,respectively,with an average value of 0.574 and2.314%,respectively,indicating a relatively high accuracy which means that multi-source data fusion is more advantageous in soil moisture retrevial in the arid desert oasis region than single data source.Among LASSO,SVM and RF models,RF model has the best performance.The collaborative scheme 4 has the highest accuracy,and the R² and RMSE are 0.772 and 1.984%,respectively.Compared with VIM_Ra18 and VIM_Oa6 of the single data source scheme,R²increased by 0.072 and 0.204,and RMSE decreased by 0.188%and 0.339%,respectively.（5）Combining soil moisture retrieval and land use classification results,it can be concluded that soil moisture content in the study area is generally low,and presents a trend of low in the west and high in the east.The eastern part of the study area is mainly occupied by forest grassland,cultivated land,and other land types.The presence of vegetation and crops can effectively prevent water evaporation,and has a certain water retention capacity,resulting in relatively high soil moisture content.The western part is mainly composed of bare land and sandy land usually with intense transpiration and scarce perennial rainfall,so the soil moisture content is relatively low without the nourishment of rivers and rain.The model inversion results are in good agreement with the actual situation in the study area.