GNSS-IR Soil Moisture Inversion Model Based On Noise Correction Of Low Vegetation

Soil moisture content is a key parameter in the study of the global water cycle and atmospheric cycle,and an important carrier of soil matter and energy cycle.It is of great significance for agricultural and forestry development,drought and flood forecast,forest fire warning,and so on.With the advantages of strong penetration,wide coverage and easy acquisition of L-band electromagnetic wave signals emitted by navigation satellites,GNSSIR(Global Navigation Satellite System-Interferometric Reflectometry)technology based on L-band reflected signals is gradually applied to the inversion monitoring of surface physical quantities such as vegetation biomass,snow depth and soil moisture.In view of the lack of soil moisture inversion model considering the effect of vegetation in GNSS-IR technology,this study introduced multiple linear regression and different machine learning algorithms to construct a GNSS-IR soil moisture inversion model based on the modified effect of low vegetation,which significantly improves the accuracy of soil moisture inversion..The main research contents and achievements of this paper are as follows:(1)A GNSS-IR soil moisture inversion model based on multiple linear regression and machine learning algorithm was constructed.Multiple linear regression,BO-GPR,ANN and CART algorithms were used to establish inversion models by using satellite SNR characteristic parameters and soil moisture measured values,and the inversion accuracy of different models was compared.The results show that the inversion accuracy of machine learning model is better than that of multiple linear regression model,and when there are many hyperparameters in the machine learning model,an appropriate automatic optimization algorithm can be used to optimize the model.R of the inversion model based on the multiple linear regression is 0.64～0.75,RMSE is 0.061～0.072,and MAE is 0.04～0.049.Compared with the multiple linear regression model,the accuracy of the inversion model constructed by machine learning algorithm is significantly improved,and the improvement effect is BO-GPR >ANN > CART.Compared with linear regression model,R of BO-GPR model is increased by21.3%-24.3%,RMSE is decreased by 16.7%-37.7%,and MAE is decreased by 10.2%-36.5%.(2)The effectiveness of using NDVI to quantify the effect of low vegetation was systematically evaluated.The results show that the phase values of SNR after the modification of vegetation noise effect decrease to a certain extent,which is consistent with the conclusion that the phase of SNR was very sensitive to the change of vegetation biomass.Compared with the phase values before and after modification of low vegetation noise effect on PRN9,23,27 and 31,it is found that the modification effect is more obvious between the day of year190～240.This is because the shrubs around the GNSS antenna at P038 site grew well during this period,and the increase in vegetation biomass led to a more significant impact on reflected signals.(3)A soil moisture inversion model was established based on normalized NDVI to correct the effect of low vegetation.By using the phase of SNR and its residual characteristic parameters after correcting vegetation noise effect,a soil moisture inversion model based on normalized NDVI was constructed by using three machine learning algorithms,including XGBoost,SVM and MLP.An index to measure the indirect proximity between soil moisture inversion value and measured value(positive contribution rate)was proposed.The results show that the accuracy of the three soil moisture inversion models has been significantly improved after correcting the effect of low vegetation.The model accuracy is XGBoost >SVM > MLP(the overall positive contribution rates are 75.0 %,66.0 %,63.2 %,respectively).Among them,R of XGBoost is increased by 1.1 % ～ 4.8 %,RMSE is decreased by 10.7 % ～34.7 %,and MAE is decreased by 9.1 % ～ 39.0 %.