Soil Moisture Retrieval Using GNSS Reflected Signal Based On Machine Learning

Soil moisture is an important hydrological variable that plays an important role in the study of ecosystem water cycle,vegetation water supply,and soil carrying capacity.The L-band signal emitted by GNSS satellite has strong penetrability,and its reflected signal is very sensitive to soil moisture.Therefore,GNSS remote sensing technology based on GNSS reflection signal provides a new and efficient monitoring method for soil moisture.At present,the theory of soil moisture inversion based on GNSS reflection signal has not yet developed a relatively perfect analytical model.Due to the environmental impact factors such as soil surface roughness,vegetation cover and snow cover,large amount of manual measurement data,complex modeling,and the weak generalization characteristics of the model.Therefore,the paper studied how to reasonable and effective suppress environmental impact factors and construct a soil moisture inversion model for GNSS satellite reflection signals,and finally proposed a soil moisture inversion model of GNSS satellite reflection signal combined with machine learning algorithm.The article analyzes the effect of the model through simulation experiments and actual experiments.(1)NDVI is a vegetation index that can effectively reflect the growth state of vegetation and vegetation coverage.Its simple and effective monitoring is helpful for the study of vegetation growth.The paper extracted the pseudorange multipath error MP1 of L1 band and the pseudorange Multipath error MP2 of L2 band from the signalto-noise ratio observation value,and then performed time series analysis,spectrum analysis and correlation analysis with NDVI.The results showed that MP1,MP2 and NDVI have obvious annual periodic and seasonal characteristics,there was a significant linear correlation between the two,and the correlation coefficient was between 0.6084 and 0.7554.At the same time,the entropy method was used to fuse the pseudorange multipath errors of the two frequencies.It was found that the correlation coefficient of the pseudorange multipath error MP and NDVI after fusion was increased by 3%~17% compared with MP1 and MP2,respectively.It was shown that the entropy method can be used to fuse the pseudorange multipath errors of two band,which can significantly improve the NDVI inversion accuracy.(2)In order to invert the complex modeling and the lack of generalization characteristics of the model in the inversion of soil moisture by GNSS reflection signals,machine learning algorithms can not only learn and store a large number of input-output mode mapping relationships,but also has the ability to simulate complex quantitative relationships without knowing this mapping.therofore,paper taked the phase of the SNR observation as the input,the soil moisture as the output,and used the machine learning algorithm to construct the SOM-BP inversion model,the SOM-SVR inversion model and the linear model,respectively.the simulation and measured experiments showed that the SOM-SVR model had the highest correlation coefficient,the root mean square error RMSE and the average absolute error MAE were the lowest,the performance of the inversion model was the best.the linear model correlation coefficient was the lowest,the root mean square error RMSE and the average absolute error MAE were the highest,and the performance of the inversion model was the worst.That indicated the clustering ability of SOM neural network and the good generalization ability of SVR can effectively suppress the influencing factors such as vegetation coverage,which are difficult to collect or model,and improve the accuracy of soil moisture inversion.(3)Using the significant correlation between pseudorange multipath error and NDVI,the fused pseudorange multipath error was taken as the vegetation information,together with the phase of the SNR observation was used as an input,and three inversion models were constructed again.The experimental results showed that after adding the vegetation information,the three inversion models have obtained more environmental characteristics,the correlation coefficient R,the root mean square error RMSE,the average absolute error MAE are about 5% higher than the inversion model without vegetation information.Therefore,when the vegetation information was difficult to obtain,the pseudo-range multi-path error is the most vegetation factor substitute,which can further improve the soil moisture inversion accuracy.