Snow Depth Monitoring Method Of GNSS Reflected Signal Based On Ensemble Learning

Snow is an important freshwater resource on Earth,and obtaining accurate snow depth data efficiently can provide reliable evidence for global climate prediction,environmental management,policy-making,and water resources regulation.Traditional methods of measuring snow depth suffer from disadvantages such as high manpower consumption,limited applicability,low spatial resolution,and high cost.The Global Navigation Satellite System Reflectometry(GNSS-R)technology,which uses passive signals and provides signals all day long,can effectively compensate for the deficiencies of traditional methods and has attracted widespread attention from scholars both domestically and abroad.However,the current utilization rate of GNSS-R snow depth monitoring research data is low,and the inversion accuracy is not high,so further development and application are urgently needed.In this thesis,we conducted multi-frequency GNSS-R snow depth monitoring research based on Global Navigation Satellite System(GNSS)interference reflection and multipath error theory.We carried out theoretical derivation,simulation,and experimental verification of the signal-to-noise ratio method,pseudo-range method,and carrier phase method,and further introduced the integrated learning method to establish a multi-satellite fusion inversion model of snow depth.The main contents are as follows:(1)We summarized the electromagnetic wave scattering theory and introduced three polarization models,explaining the relationship between electromagnetic wave scattering and surface roughness.We studied the basic characteristics of GNSS signals and the principle of Lomb-Scargle spectral analysis method.We also researched the basic principle and method of GNSS-R snow depth inversion,laying a foundation for the various GNSS-R snow depth inversion algorithms proposed later.(2)We constructed a mathematical model of GNSS interference reflection amplitude attenuation and analyzed the frequency characteristics of multipath signals.Through theoretical analysis and simulation experiments of signal-to-noise ratio,pseudo-range,and carrier phase multipath errors,the experiment showed that the main frequency of the three multipath signals is positively correlated with the height of the reflection surface,proving the feasibility of using signal-to-noise ratio,pseudo-range,and carrier phase multipath errors for snow depth inversion.(3)We established a snow depth inversion model of single/dual-frequency carrier phase and pseudo-range combination.We studied the signal-to-noise ratio observation value snow depth inversion algorithm and compared and verified the snow depth inversion error of different frequency band signal-to-noise ratios.We constructed four snow depth inversion models,including the combination of single-frequency pseudorange and dual-frequency carrier phase observation values,the combination of dualfrequency pseudo-range and dual-frequency carrier phase observation values,the combination of dual-frequency pseudo-range observation values,and the combination of dual-frequency carrier phase observation values.We analyzed their multipath error characteristics and analyzed the monitoring results using the observation data of the NWOT station.The experimental results showed that the inversion result of the signal-to-noise ratio multipath error in the GPS L1 band is the best,and the correlation coefficient with the data provided by the Plate Boundary Observatory(PBO)in the United States reaches 0.99.The combination of dual-frequency pseudo-range amplifies noise,and its inversion result is inferior to that of the single-frequency pseudorange combination.The combination of dual-frequency carrier phase observation values has lower inversion accuracy than other methods due to the weak signal quality of the GPS system L2 frequency band and the relatively large signal noise.(4)We propose a multi-satellite fusion algorithm based on ensemble learning to address the problems of low utilization of single-satellite feature data and high weighted error of simple fusion results from multiple satellites in snow depth inversion.We investigate the standardization method of snow observation data and construct a dataset.We use the Dropout-BP neural network to fuse the multi-satellite observation data on the current day for snow depth inversion,and use the AttentionLSTM neural network to fuse the multi-day and multi-satellite observation data in the past for snow depth prediction.We then combine them using the Stacking ensemble learning method to fully utilize the observation data and accurately monitor snow depth.Experimental results show that the Dropout-BP network has high inversion accuracy,with RMSE errors reduced by 0.2-2cm in three frequency bands of signal-tonoise ratio.However,different combinations of pseudorange and phase frequency bands result in RMSE errors ranging from 3cm to 6cm.The Attention-LSTM network has RMSE errors below 8cm for snow depth inversion,but the RMSE errors of signal-to-noise ratio data in each frequency band are generally 2.3cm higher than those of the Dropout-BP neural network.This algorithm can fully utilize historical observation data to achieve overall better snow depth inversion,but the historical observation data has a lag,resulting in suboptimal inversion accuracy.The multi-satellite fusion snow depth inversion algorithm based on Stacking ensemble learning can further reduce RMSE errors by about 0.1-0.6cm based on the better results of Dropout-BP and Attention-LSTM,achieving effective fusion of current and historical data.The Thesis has 57 figures,22 tables,and 71 references.