| As an important part of the earth’s ecological environment,snow accumulation is closely related to the survival and development of human society.Therefore,the rapid acquisition of snow depth data with high accuracy and spatial and temporal resolution is not only beneficial for surface climate analysis and hydrological studies,but also for snowmelt runoff forecasting and snow disaster control.Traditional snow depth measurement methods,such as manual measurement method,laser scanner measurement method,airborne laser scanning method,etc.need to consume a lot of human and material resources,and in terms of efficiency,accuracy,operational difficulty and cost can not meet the requirements of large-scale applications,for the time-sensitive snow depth monitoring is obviously unreasonable,especially in the remote areas,manual operation is more difficult.Global Navigation Satellite System Interferometric Reflectometry(GNSS-IR)technique based on signal-to-noise ratio(SNR)observations is widely used for snow depth inversion because of its easy implementation.The GNSS-IR technique has the advantages of high resolution,low cost,all-weather,and stable remote sensing signal,which fills the gap of existing medium-scale resolution monitoring means.This thesis aims to systematically study the key technologies and methods in GNSS-IR snow depth inversion applications.Through theoretical derivation,modeling analysis,and experimental comparison,this paper conducts a series of research to address the problems of poor signal separation effect,less satellite observation data of single constellation,and poor inversion accuracy of snow-free state,including snow depth inversion research based on variational mode decomposition,the study of snow depth inversion based on variational mode decomposition,the study of multi-constellation combination strategy based on entropy value method,and the study of snow depth inversion method based on machine learning optimization.The main research contents and contributions of this paper are as follows:(1)The research background and development status of GNSS-IR snow depth inversion technology are summarized,and the key technologies and remaining problems of GNSS-IR snow depth inversion are analyzed.On this basis,the relevant basic theories involved in GNSS-IR snow depth inversion technology are elaborated in detail,including a detailed introduction of the basic characteristics of GNSS reflected signals,the characteristics of SNR observations combined with the physical model of multi-path effect,and the correlation between the two are analyzed.And the types of SNR observations for each frequency band of the four major systems recorded in the current RINEX 4.0 format are listed.Finally,the principle of GNSS-IR altimetry is introduced in more detail.(2)To address the problems of of poor signal separation and poor robustness of least squares fitting method,a study of GNSS-IR snow depth inversion based on variational mode decomposition is proposed.Firstly,the variational mode decomposition algorithm is introduced and the method of determining the number of decomposition layers is described,and then the GNSS-IR snow depth inversion experiments are conducted to evaluate the performance of the variational mode decomposition algorithm in separating the trend terms.The experimental results show that the variational mode decomposition algorithm can effectively separate the reflected signals from the snow layer in the SNR observations,especially in the case of small snow depths.Compared with the least squares fitting method,the accuracy of the inversion results based on the variational mode decomposition algorithm at P351 station and AB33 station is improved by about 54% and 15%,respectively.(3)In response to the problem of unsatisfactory accuracy and spatio-temporal resolution of snow depth inversion results for single constellation,a study of snow depth inversion based on the entropy method for a multi-constellation combination was carried out.Firstly,the spatial constellation composition of the four GNSS systems is introduced,and then the inversion performances of the variational mode decomposition algorithm in different constellations are evaluated,followed by the analysis of the connection between the snow depth and the main spectrum amplitude,which is used as the additional reference data source to analyze the snow depth inversion results.Finally,the main spectrum amplitude is used as the evaluation index in the entropy method to determine the weights of each constellation to obtain the combined snow depth inversion results.The experimental results show that the combined strategy can introduce more GNSS observations and improve the accuracy and spatio-temporal resolution of the snow depth inversion results.The content of this study effectively complements the algorithm library of GNSS-IR snow depth inversion and provides theoretical and technical support to realize snow depth inversion by making full use of multi-constellation observation data.(4)Aiming at the problem of low inversion accuracy in the snow-free state in the traditional model,machine learning algorithms are introduced into the GNSS-IR snow depth inversion study,and the detection of snow state was added before the snow inversion:(a)A signal classification model based on support vector machine(SVM)with SNR sequences as input samples is introduced to detect snow states on the ground.The experimental results show that the SNR sequences from snow-free state and snow-covered state can be well detected with the help of SVM,and the detection accuracy of the samples can reach 96%.Within the set threshold,the snow state detection accuracy can reach 98% for a single day during the experiment,and the root mean square error can be reduced from 20 cm to 15 cm with the help of the snow state detection results.(b)To deal with the problems of low snow depth inversion accuracy in snow-free state,topographic bias in antenna height,and limited snow depth monitoring range,this paper proposes a multi-feature snow depth inversion model based on Back Propagation(BP)neural network optimized by genetic algorithm.The model uses the frequency,amplitude and phase of the SNR multipath oscillation term as the input features to perform snow state detection and snow depth regression prediction step by step using BP neural network.The experimental results show that the accuracy of snow state detection can reach 96.4%.Compared with the conventional model,the root-mean-square error of the results of snow depth regression prediction combining the detection results is reduced by about 23.0%,and the correlation coefficient is also improved.In addition,the model can obtain snow depth data without a priori data such as antenna height,which provides the possibility of snow depth inversion for mobile receivers. |
PDF Full Text Request