Soil Moisture Retrieval Method Based On Multi-source Remote Sensing Collaboration In The Southern Slope Of Qilian Mountains

Soil water is an important part of land water cycle,controlling surface runoff and surface water evaporation.It plays an important role in hydrological prediction,surface carbon cycle,surface evaporation and vegetation transpiration.Qilian Mountain area is a typical ecologically fragile area and an important climate sensitive area.Soil water is also the main body of ecosystem circulation.Macro-monitoring the spatial distribution and regularity of soil water in this region is of great theoretical significance for rational allocation of water resources,ecological environment monitoring and agriculture and animal husbandry development planning.In this paper,the Amidongsot watershed located on the southern slope of Qilian Mountains was selected as the research area.Sentinel-1 SAR data and Landsat 8 optical image data were used to carry out soil water collaborative inversion,and water cloud model and VWC model were used to construct soil water inversion model.Logistic regression decision tree model,support vector machine and BP neural network were selected.Three machine learning algorithms were used to estimate the surface soil moisture,compare the accuracy differences of different methods,select the best inversion method and analyze the importance of each influencing factor.The main research results are as follows:(1)In the process of removing vegetation water influence by water cloud model,in Sentinel-1 SAR data,VH backscattering coefficient inversion of soil water influence is superior to VV backscattering coefficient inversion under VV and VH polarization modes.In Landsat 8 optical images,among the three planting coverage indexes of NDVI,NDWI and NDMI,NDMI was more sensitive to the analysis of vegetation water content and contributed a lot to the estimation of soil water content.(2)Combined with the vegetation index of water cloud model and the radar backscattering coefficient under different polarization modes,the data with better fitting effect,soil backscattering coefficient and incidence Angle were input into three machine learning algorithms,and the output data was the corresponding soil water value.The results of soil water inversion by logistic regression tree R~2=0.8243,RMSE=0.0551 were obtained.ub RMSE=0.1062,bias=0.0030;Support vector machine inversion results R~2=0.6696,RMSE=0.0642,ub RMSE=0.1042,bias=0.0041;BP neural network inversion results R~2=0.9043,RMSE=0.0365,ub RMSE=0.1038,bias=0.0058.The results show that the three machine learning algorithms all have certain fitting effects and the three models can invert the soil moisture in this study area,among which the BP neural network has the best estimation effect for this area.(3)In the Amidongsoe Watershed,which has the typical characteristics of the southern slope of Qilian Mountains,the overall soil water content distribution is good,except for the water body,the higher soil water is mainly distributed in the low altitude area and the mountain top area;In addition to the influence of elevation,the soil moisture around the gullies in the region is relatively high and distributed perpendicular to the elevation gradient.Due to the difference of evapotranspiration and water conservation function,the difference of soil moisture on the negative slope and the sunny slope is large and the boundary is clear.The spatial pattern is influenced by the gradient change of elevation and the distribution of different microgeomorphic types.