Inversion Of Soil Moisture In Arid Area Based On Machine Learning And Multi-source Remote Sensing Data

As a basic state variable that controls a wide range of processes occurring at the land-atmosphere interface,soil moisture affects many processes such as infiltration,runoff,evaporation,heat exchange,solute infiltration and erosion.At the same time,soil moisture is an important parameter for hydrology,meteorology,and agro-environment research,and it is also one of the important indicators of water resource cycling.Due to extremely shortage of water resources,desertification and salinization have been very prominent ecological problems in arid areas.As a key variable affecting the water and energy balance in arid area,soil moisture is the most effective driving factor for regional hydrological and vegetation processes.Therefore,accurate monitoring of soil moisture is of great significance for ecological environment protection and sustainable development in arid regions.However,due to the complicated spatial and temporal variations of soil moisture,accurate assessment of soil moisture is still a challenging task,and the rapid development of remote sensing technology may provide the possibility of accurate soil moisture estamation in a large-scale.Many scholars have integrated different remote sensing data to retrieve soil moisture in previous studies.Among them,Landsat-8 and Radarsat-2 are typical representatives of optical and radar data.As we know that both optical and radar remote sensing contain soil moisture information,each has its own advantages and disadvantages.Finding and establishing a method or model that can synthetize the information inherent in both optical and radar images are of great significance for improving the estimation accuracy.Owing to the higher adaptivity,machine learning models provide a new clue to achieve this objective.Juyanze,which is located in western Inner Mongolia,is selected as the study area.Based on Landsat-8 and Radarsat-2 imagery,research on soil moisture inversion using machine learning algorithms,including Random Forest（RF）,Support Vector Machine（SVM）and Back Propagation Artificial Neural Network（BP-ANN）,combined with soil moisture impact factors and in situ measurements,was carried out in this paper.First,Radarsat-2 image was processed using standard intensity and phase processing,and Cloude-Pottier decomposition as well as Yamaguchi decomposition,to obtain SAR backscatter coefficients and multiple polarimetric parameters,which are moisture related and are used as radar impact factors.Based on Landsat-8,multiple band reflectances are acquired and optical indices such as NDVI are calculated as optical impact factors.Second,the importance of soil moisture impact factors and their correlation with soil moisture are quantitatively analyzed.Different parameters are combined to form multiple combination schemes,which are used as input factors respectively to construct different machine learning models.Based on k-fold cross-validation,all the models are comprehensively evaluated according to their precision,accuracy,and stability.Finally,the model with excellent performance was used to retrieve the regional soil moisture,and random forest classification was performed to acquire the current land use information.This study will provide reference for the modeling of soil moisture inversion in arid areas by machine learning method,and provide a scientific basis for understanding regional soil moisture dynamics and regional ecological environmental protection.The main conclusions are as follows:1.Among the microwave impact factors,mean scattering anglehas the greatest influence on the model accuracy,followed by entroy H and anisotropy A.Among backscattering coefficients,cross-polarization coefficients have higher important scores than those of co-polarization,and the correlation coefficients with soil moisture are also higher.even scattering1)1)_{（9（9） and volume scattering1）1}have much higher significance than surface scattering1)1)and even scattering1)1)₍(9（9）.Parameters derived from Cloude-Pottier decomposition have distinct contribution to the inversion results,and their importance scores and correlation coefficients are significantly higher than backscatter coefficients and Yamaguchi decomposition parameters.Among the optical impact factors,correlation between optical reflectance and soil moisture varies from the highest to the lowest in the sequence of TIRS,SWIR,RED,and NIR.The corresponding optical products also show a consistent arrangement rule,indicating that soil moisture is more sensitive to thermal infrared and short-wave infrared band.2.For the three machine learning methods,the performance of the soil moisture inversion model coupled with multi-source remote sensing data is significantly better than the model using only radar or optical remote sensing data sources.V-R² is increased by an average of 25.58%,and V-RMSE is decreased by an average of 11.93%.Results from the cooperative inversion model integrate the spatial pattern of both radar optical models.Comprehensive application of multi-source remote sensing data assigns the model a greater choice of information space,indicating that the combination of multi-source remote sensing data can improve the model performance to some extent.3.Considering R² and RMSE,the RF model is more suitable for soil moisture inversion in arid areas than SVM and BP-ANN.Regardless of the combination scheme,e.g.sole radar scheme,sole optical scheme,or the cooperative scheme,the RF model can always achieve the lowest V-RMSE and the highest V-R²,while SVM and BP-ANN perform relatively poorly,indicating the great potential of RF model in soil moisture retrieval in arid regions.Among the 69 models,the best performance model is a RF model trained with the combination of BC+Cloude+Yamaguchi+OI.The V-R² and V-RMSE values are 0.86 and 5.06%with V-SD_R² and V-SD_RMSE being 0.15 and2.85%,respectively.This model can account for 91%variations in soil moisture.4.Soil moisture content is low in most part of the study area,and the average value is only8.29%.The soil moisture content in the Swan Lake and the ancient lake core area is much higher than other areas.Spatial variations in regional soil moisture are mainly related to soil texture,thermal conditions,and surface coverage.Vegetated areas and tidal flat areas have higher soil moisture due to vegetation cover and water replenishment.The combined effect of the"back salt"phenomenon and salt crust reflection results in higher soil moisture in some saline-alkali areas.Bare soil texture is more sticky,which is benefit to moisture retention,so the moisture content is higher compared to sandy soil.Model inversion results demonstrated a high consistency with the actual conditions,and can provide valuable spatial moisture information for the whole region.