Dimensionality Reduction Modeling And Inversion Of Soil Moisture In Ejina Oasis By Remote Sensing

Soil moisture affects water,energy and biogeochemical cycles and is a key parameter in meteorological,hydrological,ecological and agricultural systems.Water resources are scarce in arid and semi-arid regions,hence accurate monitoring of soil moisture can provide basic data and decision support for preventing drought events and predicting crop yields.At present,remote sensing has become an important technical means for dynamic monitoring of soil moisture in a more extensive area.Polarimetric Synthetic Aperture Radar(Pol SAR)image contains information of the geometric characteristics,backscattering characteristics and polarization characteristics of the scattering target,showing prominent advantages in target detection and recognition,texture feature and surface parameter extraction.Pol SAR data has demonstrated great potential in soil moisture monitoring in vegetated areas with the rapid development and wide application of polarization target decomposition technology.However,related researches in arid and semi-arid areas are limited.Red-edge band is sensitive to vegetation growth,and the band information is of great significance to soil moisture monitoring in vegetated areas.Some works have been reported in recent years,and further study is reqiured.As we known that soil moisture is influenced by many factors and the relationship among them is complex.How to establish a more effective soil moisture inversion model in vegetated areas based on remote sensing has become the main focus of scholars at home and abroad.Owing to the characteristics of fewer tuning parameters,multifactor integration,and highly complex nonlinear mapping,machine learning models,such as Random Forest(RF)and Support Vector Machine(SVM),provide a new clue for soil moisture inversion using both Pol SAR and optical data.However it may cause information redundancy while allowing multi-factor input.Therefore,it is necessary to further explore a factor dimensionality reduction method to improve soil moisture inversion accuracy.Based on Radarsat-2 C-band Pol SAR data and Sentinel-2A optical data,this paper aims to explore a dimensionality reduction modeling method of soil moisture by remote sensing in Ejina Oasis of Alxa League in western Inner Mongolia.Firstly,the radar backscattering coefficients were extracted from Pol SAR data,and various radar characteristic parameters such as polarization scattering features are extracted by H-A-α decomposition,FreemanDurden decomposition,An&Yang decomposition and van Zyl decomposition.Optical characteristic parameters such as spectral characteristics,red-edge vegetation indices and non-red edge vegetation indices were obtained based on Sentinel-2A optical data.Secondly,according to Mean Decrease Accuracy(MDA),the importance of radar and optical characteristic parameters was scored respectively and a variety of parameter combination schemes were developed.Then multicollinearity test was performed on all the schemes according to Variance Inflation Factor(VIF).Parameter dimensionality reduction using Principal Component Analysis(PCA)was implemented for combinations that do not meet the conditions,and RF and SVM models were established to evaluate the effect and applicability of each parameter combination scheme and PCA dimensionality reduction in soil moisture retrieval in arid desert oasis area.Finally,soil moisture distribution map was obtained based on the dimensionality reduction model,and the soil moisture spatial distribution characteristics as well as the distribution patterns of different land use types were analyzed.The main conclusions are as follows:1 、 In MDA evaluation,each radar characteristic parameter presents different importance levels in grassland,arableland,woodland and overall samples,indicating the necessity of discussing on different vegetation types.Among the four backscattering coefficients,the cross-polarization backscattering coefficients are of high importance compared with the co-polarization coefficients.In H-A-α decomposition,A and RVI are of high importance.Among the three decomposition methods with three components,the importance of dihedral scattering and volume scattering to soil moisture retrieval is relatively significant for grassland samples,while for the arableland samples,the contribution of surface scattering is greater,and a certain regularity is showed in the three-component polarization decomposition.2、In the soil moisture inversion models based on radar characteristic parameters,compared with single-polarization and dual-polarization,the full-polarization scheme has higher inversion accuracy.Compared with the backscattering information,the polarization target decomposition method has higher accuracy,and the Freeman-Durden,An&Yang and van Zyl decomposition based on model have more potential in soil moisture retrieval than the H-A-α decomposition based on eigenvalues and eigenvectors.By comparing the inversion accuracy of the M3 and M11 schemes screened by MDA with the three 3-component scheme and the 11-component scheme of H-A-α polarization target decomposition with the same number of input features,the effectiveness of MDA method is demonstrated.Appropriately increasing the number of parameters will improve the accuracy of moisture inversion to a certain extent,but too many parameters will cause the multicollinearity problem.In the soil moisture retrieval models based on optical characteristic parameters,red-edge vegetation indices usually have higher accuracy than non-red edge vegetation indices and spectral features.3 、 PCA dimensionality reduction was carried out for the schemes with serious multicollinearity in combination schemes after MDA importance scoring,namely M11 and MALL schemes based on radar characteristic parameter combination,and ALL scheme based on optical characteristic parameter combination,respectively,and all were used as input to RF and SVM model.The results indicate that the model accuracy is effectively improved after dimensionality reduction.Moreover,it is beneficial to model weight reduction and computing efficiency improvement.4、By comparing the soil moisture inversion accuracy between RF and SVM models,it is found that the accuracy in both training set and validation set of RF models is generally higher than those of SVM models,indicating that RF model had higher accuracy in soil moisture inversion in the study area.Compared with the soil moisture retrieval models based on optical characteristic parameters,models based on radar characteristic parameters are more suitable for soil moisture retrieval in the study area.5、The soil moisture content in most parts of the study area was low,and grassland covered area has the highest water content,followed by woodland,arable land and bare land.Soil moisture distribution is related to vegetation cover,vegetation structure,water supply and evapotranspiration.The grassland covered area has relatively good soil moisture condition due to water nourishment.The woodland has certain soil and water conservation ability due to the developed root system.The arable land can reduce evapotranspiration to a certain extent because its developed veins are conducive to surface cooling.The model inversion results can better reflect the soil moisture distribution pattern in the study area.