Multi-source Remote Sensing Soil Moisture Retrieval Based On Neural Network Algorithm

Soil moisture is a key parameter in hydrology,meteorology and agricultural research and plays an important role in the global water and energy cycle.The use of remote sensing data to monitor surface soil moisture is the focus of current research.However,both optical remote sensing and active or passive microwave remote sensing have their advantages and disadvantages.The combined retrieval of surface soil moisture by multi-source remote sensing data is a trend in recent years.Based on Sentinel-1 SAR data,Sentinel-2 optical data and AMSR2 brightness temperature data,BP neural network model was trained in the Naqu area of the Qinghai-Tibet Plateau and the Salamanca region of Spain to retrieve the surface soil moisture.And the main conclusions are as follows:(1)In the BP neural network,the radar backscattering coefficient,the radar incident angle and the vegetation index were used as input variable of training models,and it was feasible to retrieve the soil moisture.(2)In the combined retrieval of vegetation and soil moisture by active and passive microwave data,the microwave polarization difference index(MPDI)could effectively remove the influence of vegetation on soil moisture retrieval.Adding slope factor expressing terrain elements into BP neural network could train the network better and obtain more accurate retrieval result(RMSE=0.045cm~3/cm~3,ubRMSE=0.043cm~3/cm~3,r=0.767,Bias=0.001cm~3/cm~3).It was indicated that for the special geographical area,factors that expressed geographic topography should be considered in the soil moisture retrieval to improve the accuracy of soil moisture retrieval results.(3)NDVI,NDWI1 and NDWI2 from Sentinel-2 could well express vegetation coverage information in the combined retrieval of vegetation cover soil moisture with optical and active microwave data.NDWI1 based on short-wave infrared SWIR1(center band 1.61 m)had better performance.In addition,Sentinel-1 VV polarization mode data was more suitable for the study of soil moisture retrieval(RMSE=0.049cm~3/cm~3,ubRMSE=0.048cm~3/cm~3,r=0.681,Bias=0.008cm~3/cm~3).Taking the VV and VH polarization backscattering coefficient,radar incident angle,NDVI and NDWI1 as neural network input variables,the soil moisture retrieval results were more accurate(RMSE=0.045cm~3/cm~3,ubRMSE=0.044cm~3/cm~3,r=0.576,Bias=0.011cm~3/cm~3).Indicating that increasing the BP neural network input parameters could be appropriate to compensate for the lack of information caused by a polarization data and a single vegetation index,it was more conducive to retrieve the soil moisture.