Study On Spaceborne GNSS-R Vegetation Parameter Inversion Based On Modified Models And Machine Learning

Since its emergence in 1990s,the Global Navigation Satellite System Reflectometry(GNSS-R)technology has been one of the research hotspots for scholars at domestic and international level because of its abundant signal sources,low cost and strong signal penetration capability.Vegetation monitoring based on spaceborne GNSS-R technology and inversion of several parameters related to vegetation is one of the new application directions in the field of spaceborne GNSS-R.Due to the influence of various factors such as rough and uneven surface of the earth,the probability of vegetation coverage and reflection from water bodies,GNSS signals will be scattered during the propagation process,with different degrees of attenuation,and finally received by the receiver.At the same time,the GNSS reflected signals also carry information such as surface roughness,soil moisture,vegetation and so on.Based on this principle,this thesis firstly analyzes the correlation among three vegetation parameters,and then uses the spaceborne GNSS-R observation data to correct the surface reflectivity on a global scale based on the modified models and machine learning models to eliminate the influence of other factors in the process of GNSS signal reflection,and then inverts three vegetation parameters:vegetation water content(VWC),vegetation optical depth(VOD),and canopy height(CH).The main research results are:(1)The correlations between three vegetation parameters were studied based on the Soil Moisture Active and Passive(SMAP)and the Global Ecosystem Dynamics Investigation(GEDI)datasets.The results show that there are good correlations among the three vegetation parameters.Among them,the highest correlation was found between VWC and VOD with a correlation coefficient of 0.96;the lowest correlation was found between VWC and CH with a correlation coefficient of 0.79.(2)The correlations between surface reflectivity and the three vegetation parameters were analyzed using surface reflectivity calculated from the Cyclone Global Navigation Satellite System(CYGNSS)data over a one-year period.Among them,the correlation between surface reflectivity and CH was the highest with a correlation coefficient of 0.357,and the correlation with VWC was the lowest with a correlation coefficient of 0.258.The results show that the use of surface reflectivity alone is not sufficient to invert the vegetation parameters.(3)Based on the CYGNSS observation data,vegetation observations were obtained based on the modified models based on the surface reflectivity after removing the sampled data in areas where the land cover type was water,urban and so on,and then the vegetation parameters were inverted.Among them,the inversion results of vegetation parameters based on the R-V-R model correction were the best.The Pearson correlation coefficient(R),the mean absolute error(MAE)and the root mean square error(RMSE)in the VWC inversion results are 0.672,0.973 kg/m~2,and 1.176 kg/m~2,respectively;the R,MAE,and RMSE in the VOD inversion results are 0.606,0.127,and 0.149,respectively;and the R,MAE,and RMSE in the CH inversion results are0.565,2.061 m,and 3.16 m,respectively.The results show that adding soil moisture and roughness coefficient data to the surface reflectivity calculated by CYGNSS can effectively improve the inversion accuracy of the spaceborne GNSS-R vegetation parameters.(4)Five data of surface reflectivity,incident angle,soil moisture,roughness coefficient,and the International Geosphere-Biosphere Programme(IGBP)land cover type were used as model input data,three vegetation parameters were used as output data,and three machine learning models,BP neural network,Support Vector Regression and Random Forest(RF),were established by selecting the optimal hyperparameter combination of the models with Bayesian optimization and other methods,and the model results were analyzed.The results showed that the RF inversion of the three vegetation parameters gave the best results.Among them,the inversion results based on the RF method have R,MAE,and RMSE of 0.905,0.317 kg/m~2,and0.547 kg/m~2 with VWC,R,MAE,and RMSE of 0.862,0.054,and 0.079 with VOD,and R,MAE,and RMSE of 0.735,1.423 m,and 2.251 m with CH,respectively.In general,all three machine learning methods effectively improved the inversion accuracy of the three vegetation parameters.This thesis has 29 figures,14 tables,89 references.