SAR-based Snow Cover Discrimination And Snow Depth Inversion Method Integrated Scattering Characteristics And Machine Learning

Snow is an essential component of the terrestrial ecosystem and one of the most positive elements of the cryosphere,with an important feedback effect on the global climate regime.Snow cover and its melting process can store water for hydropower generation and agricultural irrigation in mountainous areas,being one of the important water resources;at the same time,natural disasters such as avalanches and spring snowmelt floods can also cause serious damage to public facilities.The precise retrieval of snow depth(SD),snow-covered area(SCA),and dry and wet snow discrimination has research significance for snowmelt runoff monitoring,regional climate simulation and water resources management.Synthetic aperture radar(SAR)has received widespread attention and rapid development because of its unique advantages of being all-weather,penetrating,independent of light,and less affected by clouds.Since the backscattering information received by the radar is more sensitive to the dielectric properties and geometric structure of snowpack,it provides the possibility for SAR-based data to extract the physical parameters of the snowpack,distinguish dry and wet snow cover,and invert the snow depth.In view of the shortcomings of the current SAR-based algorithms for dry and wet snow discrimination and snow depth inversion,the exploration on extracting extent of snow cover,distinguishing dry and wet snow,and retrieving snow depth is carried out by integrating snow scattering characteristics and machine learning(ML)techniques.The main research contents and conclusions of the dissertation are as follows.(1)The effects of local incidence angle(LIA)and underlying surfaces on the backscattering coefficients of different polarizations are analyzed.,and a SCA extracted method using multipolarization SAR has been developed.Due to the increased liquid water content and dielectric constant,wet snow has a significant change in its backscattering value,compared to the snow-free surfaces.Taking advantage of this characteristic,the dissertation develops a SCA extraction method based on multipolarization SAR.Firstly,the variation of the backscattering values under different polarization for wet snow and snow-free surface with LIA has been investigated,and a weighting function considering LIA is constructed to combine the backscattering of co-polarization and cross-polarization to generate multipolarization backscattering ratio images;secondly,adaptive thresholds are calculated for segmenting the ratio images by distinguishing the types of underlying surfaces,and the decision rules established by the terrain information are introduced to extract snow coverage.The assessment shows that compared with the overall accuracy(OA)of 43.5%obtained by the single polarization-based method with fixed empirical thresholds,the method proposed in this dissertation is more advantageous in extracting snow cover,and the OA value has been improved to 80.8%.(2)With respect to the dry and wet snow delineation,the supervised DSVM-MRF method and the unsupervised DSAE-WFCM method have been developed by joining the dual-polarized SAR polarization decomposition and ML.Most of the existing algorithms are proposed with commercially available and expensive quad-polarized SAR as the study data,and the applicability of these algorithms is limited.In contrast,the dual-polarized SAR with larger coverage,longer time series,and freely available data has more application advantages.In this dissertation,a variety of polarization decomposition parameters were obtained by deriving the expression of H-α polarization decomposition on dual-polarized SAR data.The DSVM-MRF method selected the optimal polarization parameters as the input features of the Support Vector Machine(SVM)through the J-M distance,and to ensure the iterative convergence to the local optimal solution,the discrimination results obtained by the SVM were used as the initial value of Markov random field(MRF)model,and then the iterated conditional method(ICM)iteratively outputted the discriminated results.In order to improve the ability of dual-polarized SAR data to be classified on wet and dry snow,and meanwhile reduce the manual interpretation caused by outlining training samples,an unsupervised DSAEWFCM method is also developed.The DSAE-WFCM method utilized the sparse autoencoder(SAE)to build a deep network based on the pixel neighborhood to extract the effective polarimetric information,and the reconstructed polarization parameters were fed into the feature-weighted fuzzy C-mean clustering(WFCM)with different weights by distinguishing different underlying surfaces.Validation by field measurement data shows that the two algorithms developed using dual-polarized SAR data consider pixel neighborhood information and reduce the influence of speckle noise,and the OAs of DSVM-MRF and DSAE-WFCM reach 84.5% and 88.8%,respectively,which are similar to the overall accuracy of 90.0% obtained by the quad-polarized SAR method.(3)Based on the co-polarization phase difference(CPD),a snow depth inversion method that integrates snow scattering characteristics with machine learning techniques has been developed.Since snow ice particles with different structures cause refractive index differences in the HH and VV polarization,the phase difference formed by the delay in the occurrence of different polarization can be characterized using CPD,in which the axis ratio of ice particles was used to quantitatively describe the anisotropic structure of snow ice particles.Using the relationship between incident wavelength and phase,a geometric equation was constructed by the propagation paths of the two polarization signals within the snowpack,and the CPD can be established as a function of snow depth,density,axis ratio of ice particles,radar incident wavelength,and radar incident angle.By extend FAST sensitivity analysis(EFAST),it has been found that the CPD of horizontally aligned ice particles is larger than 0 and more sensitive to snow depth,and the snow depth inversion has been achieved by fitting the CPD as a function of snow depth;while the CPD of vertically aligned ice particles is smaller than 0 and more sensitive to the axis ratio of ice particles,and the CPD has been first fitted as a function of axis ratio,and then the CPD,snow density,axis ratio and radar parameters were used to inverse snow depth;thus,the snow depth inversion method has been obtained by distinguishing the structure of snow ice particles.Meanwhile,K-nearest neighbor regression,support vector regression and random forest regression were employed to retrieval snow depth.Finally,the decision rules with optimal accuracy were constructed to form a snow depth inversion algorithm that integrates snow scattering characteristics and machine learning.The validation results show that the accuracy of the snow depth inversion obtained by combining snow scattering characteristics and machine learning is better than that of the other two single methods,with the root mean square error(RMSE)of 4.9 cm and the correlation coefficient(R)of 0.7.