Study On GNSS-IR Soil Moisture Based On PSO-BP Neural Network Algorithm

Soil water,also called soil water content,is one of the main indicators of water condition in soil,and it is also one of the most important parts in ecological system.The conventional methods of measuring the water content of soil have their merits and demerits.As GNSS-IR(Global Navigation Satellites System Reflection Interchange Reflection)is developing rapidly,GNSS-IR technique has been widely applied in the world because of its high precision,good continuity and low-cost.But due to the influence of surface roughness,plant cover,snow depth and SNR,there are abnormal features in feature parameters,which influence the precision of soil water retrieval.In order to solve these problems,we present an outlier detection method based on MCD robust estimation,which can reduce the influence of outlier to data quality.On the other hand,machine learning algorithms-BP Neural Network and particle swarm optimization(PSO)are introduced to weaken the influence of environmental factors on inversion accuracy.This paper focuses on the following topics:1.On the basis of present situation of GNSS,a systematic introduction was given to the theoretical knowledge of GNSS reflection signals and the basic principle of GNSS-IR inversion of soil moisture.Firstly,the current GNSS system and the characteristics of GNSS reflection signals were introduced.Then,the GNSS-IR soil moisture inversion process was introduced.Based on traditional methods,a onedimensional linear regression soil moisture model was constructed using characteristic parameters such as amplitude,frequency,and phase.In the inversion model composed of three characteristic parameters,the phase parameter model performs best.The average value of the correlation coefficient R of 13 satellites is0.408,the average value of the root mean square error RMSE is 0.0526,and the average value of the average absolute error is 0.0413.It can be seen that the inversion accuracy still needs to be further improved.2.Solution to anomalous feature parameters due to environment factors such as surface roughness,vegetation coverage,snow depth and the quality of signal to noise ratio data,An MCD robust estimate algorithm is presented.Based on the robust estimation of MCD,a robust amplitude model,a robust frequency model,and a robust phase model were constructed.Among them,the correlation coefficient R of the robust amplitude model is 30.7% higher than that of the traditional amplitude model,the root mean square error RMSE is 6.85% lower than that of the traditional amplitude model,and the average absolute error is 6.19% lower than that of the traditional model.The correlation coefficient R of the robust frequency model is 64.42% higher than that of the traditional frequency model,the root mean square error RMSE is 9.8% lower than that of the traditional amplitude model,and the average absolute error is 8.41%lower than that of the traditional model.The correlation coefficient R of the robust phase model is 24.8%higher than that of the traditional phase model,the root mean square error RMSE is 5.83% lower than that of the traditional amplitude model,and the average absolute error is 5.57% lower than that of the traditional model.In order to make full use of multiple characteristic parameter information,a multiple linear regression model based on MCD robust estimation was established.Among the 13 satellites,the mean R is 0.5623,the mean RMSE is 0.0427,and the mean MAE is 0.0343.In terms of correlation coefficient R,compared with the robust amplitude model,the robust frequency model,and the robust phase model,it increased by 51.17%,70.56%,and 10.44% respectively;In terms of root mean square error(RMSE),it decreased by 17.37%,17.03%,and 13.85%,respectively;The average absolute error MAE decreased by 22.02%,19.61% and 12.11% respectively.It is shown that the MCD robust estimate method can increase the precision of the inverse.3.In view of the influence of environmental factors of stations in soil moisture inversion research,this paper introduces BP neural network and PSO-BP neural network on the basis of using MCD robust estimation to remove outlier.The input layer parameters of the machine learning algorithm are amplitude,frequency,and phase,and the soil moisture value is used as the output layer parameter.We constructed robust BP neural network models and robust PSO-BP neural network models respectively,and compared them with robust multiple linear regression models.The results of 13 satellites indicate that the average correlation coefficient R of the robust PSO-BP neural network model is 0.6928,which is 23.21% and29.93% higher than the robust multiple linear regression model and robust BP neural network model,respectively;The average root mean square error(RMSE)of the robust PSO-BP neural network model is0.0368,which is 14.99% and 22.10% lower than the robust multiple linear regression model and robust BP neural network model,respectively;The average absolute error MAE of robust PSO-BP neural network model is 0.0286,which is 16.62% and 22.28% lower than that of robust multiple linear regression model and robust BP neural network model respectively;Based on the above accuracy indicators,the advantages and disadvantages of the three models are as follows: robust PSO-BP neural network model>robust multiple linear regression model>robust BP neural network model,It is suggested that the Machine Learning Algorithm can reduce the surrounding environment and increase the precision of soil water inversion.