SAR Polarization Decomposition And Feature Selection For Snow Recognition

Snow cover is an important component of the cryosphere with multiple attributes,and it is also one of the most active natural factors of the globe's surface.Because snow cover has a high reflectivity to the solar radiation,global snow eover plays an important role in regulating surface energy budget and global or local climate change through a range of complex interactions and feedbacks.Meanwhile,seasonal snow cover is one of the main sources of seasonal streamflow for mountainous rivers in Northwest China and the precious fresh water resources in arid and semi-arid regions.Also,it plays a very important role in the development of the national economy and the coordinated development of the region.Timely and accurately measurements of the seasonal snow cover are needed for climate research,snow melting runoff simulation,and water management.Remote sensing techniques provide unique opportunities for gathering comprehensive information on snow,especially,the development of polarimetric synthetic aperture radar provides a new method for snow cover information acquisition.Polarimetric synthetic aperture radar can provide a large amount of polarization features,which is sensitive to the physical parameters of snow,and provides a new method for snow recognition.This paper is a part of the National Natural Science Foundation "Joint inversion of snow water equivalence based on SAR and high-resolution optical remote sensing"(Grant No.41271353).According to the needs of the projects on snow recognition,the research chooses Manasi River Basin of Tianshan Mountain as study area.This study obtained polarization features by target decomposition based on RADARSAT-2 data,and explored a method on snow recognition based on the polarization features of snow,and the roles of polarization features in snow cover recognition was also discussed.The main research contents and conclusions are as follows:(1)Representation of snow cover in SAR images.Four kinds of backscattering coefficient obtained from RADARSAT-2 data were used to analysis the difference between the snow cover and snow-free.Meanwhile,take three basic scattering components obtained from Pauli decomposition for example,this study analysised the polarization scattering response of the snow cover and snow-free area.The analysis results show that the backscattering coefficient of snow cover is lower than that of snow-free under the same conditions of underlying surface and incidence angle.Also the scattering mechanisms of snow cover and snow-free are different in RGB composition images of Pauli decomposition.It is possible to discriminate the snow-cover from snow-free by SAR data.(2)Polarization feature extraction from SAR data.Polarization scattering matrix and coherent matrix were calculated,and the polarization features of SAR image were obtained using the coherent target decomposition and incoherent target decomposition methods.The results show that the target decomposition methods are effective methods to extract different polarization features of snow cover,which represent different scattering mechanisms.(3)Optimal polarization feature selection for snow cover recognition.The random forest model was trained based on the polarization features,and the variable importance estimations were acquired by the random forest model.According to the variable importance,optimal polarization feature subset was selected.The key point was to discuss the contribution of optimal polarization features to accurate discriminate snow and snow-free area.The results show that the polarization features possess different degree of importance,and optimal polarization feature selection is very necessary.It is found that the optimal polarization feature set selected by random forest is useful to snow recognition and Vvol.Fvol,Kd,T33,P3,H(1-A),DERD,PF,and Span possesses more contribution to the snow recognition in snow accumulation and snow-melt period.The main contributions of this study is the selection of optimal polarization features for snow cover recognition by random forest model,and the exploration of the role of different polarization features in the process of snow identification based on scattering mechanism.The result of this study will be useful to snow recognition by SAR data.