Font Size: a A A

Research On Fully Polarimetric SAR Image Processing And Ocean Surface Wind Field Inversion

Posted on:2021-02-15Degree:MasterType:Thesis
Country:ChinaCandidate:Y DingFull Text:PDF
GTID:2370330647452749Subject:Electronics and Communications Engineering
Abstract/Summary:PDF Full Text Request
Synthetic Aperture Radar(SAR)is an effective remote sensing method for marine wind and waves.Current mature wind field inversion models are mainly based on VV unipolar SAR images.The subject is a multi-parameter model that is nonlinearly related to multiple variables at the same time.Wind direction and wind speed cannot be decoupled for single factor inversion,which limits business applications.With the development of multi-polarization and full-polarization SAR,it is found that cross-polarized scattering is weakly correlated with wind direction,which can realize direct inversion of wind speed,and has a broad application prospect.At present,full-polarization SAR is in the exploration stage for sea surface wind field inversion.On the one hand,the number of channels of full-polarization SAR images is doubled,which makes the image preprocessing more complicated.On the other hand,the linear relationship between variables in the cross-polarization inversion model.More research is needed on its universality.In this paper,we use GF-3 fully polarized band 1(QPSI)mode SAR image data to carry out technical research on inversion of sea surface wind field: pre-processing and water segmentation of fully polarized SAR images,using improvements Deeplabv3 + network model,a multi-channel semantic segmentation model suitable for extracting four polarized image features is proposed.For the GF-3 fully polarized SAR data used for sea surface wind field inversion,four commercial models of the CMOD series were used to verify the applicability of the fully polarized mode inversion wind field;the cross-polarized data was analyzed for wind speed,relative wind direction and Based on the correlation of radar incident angles,a multi-stepwise regression method and BP neural network method are used to explore the model of GF-3 cross-polarization data for sea surface wind speed inversion.The specific research content is as follows:1.Analyze the polarization and scattering mechanism of SAR image and its principle for wind field inversion,and compare several current commercial C-band commercial wind field inversion models and their advantages and disadvantages.TheGF-3 image is then pre-processed,including radiation calibration,multi-view processing,geocoding,and coherent speckle noise filtering.Especially for the filtering of coherent speckle noise,which affects the accuracy of wind field inversion,the filtering effect of common filtering algorithms on the data of full polarization is compared.Among them,enhanced Lee filtering is used for the standard deviation of cross-polarized data The indicators on the equivalent and the equivalent views are the best,and Gamma Map filtering is the best on these two indicators for co-polarized data,but in the edge preservation index,the enhanced Lee filtering has an excellent effect on the four polarized images For Gamma Map,enhanced Lee filtering is used to denoise the image.2.Analyze the defects of traditional image segmentation technology for fully polarized SAR images,improve the semantic segmentation model Deeplabv3 +,design the semantic segmentation model of multi-channel input,and juxtapose the four polarized images of fully polarized SAR images as a four-channel image input network Using the improved lightweight network Xception as a feature extraction network,the application of deep separable convolution reduces the parameter setting.The multi-channel model solves the result of inconsistent water segmentation in the same scene due to different polarization modes when each image is separately segmented in a fully polarized SAR image,and improves segmentation accuracy and reliability.The test images were used to compare the segmentation algorithms.The results show that in the water segmentation of each polarized image,the effect of the deep learning segmentation model is better than the traditional method.However,the single channel input FCN and Deeplabv3 + network can achieve 96.67% of the optimal MIo U in the cross-polarization data segmentation of the test image,but the segmentation effect on the co-polarization data is significantly lower than the cross-polarization,the minimum is only 88.99%,the segmentation result With inconsistency,the segmentation result of this method achieves 95.85% MIo U.3.Aiming at the research on the use of Gaofen-3 fully polarized data for sea surface wind field inversion,the applicability of the data inversion for the fully polarized SAR images containing NDBC buoy data was verified by using fourCMOD series business models.It shows that the inversion error of the CMOD5 model is the smallest compared to the buoy observation,the mean square error of the four wind speeds is only 0.31 m / s,and the wind direction error is 4.06 °.In order to explore the correlation between GF-3 cross-polarization scattering and wind vectors,80 SAR images of QPSI mode in the Atlantic,Pacific and other seas were collected,using ERA-Interim data as the initial wind field and verification data,and the cross was analyzed according to the Pearson correlation coefficient The linear correlation between the polarization data and the variables in the wind field inversion,and then the reference model was established by two methods of stepwise regression and BP neural network.The research results of the current data show that the GF-3cross-polarized backscattering intensity has a linear relationship with the wind speed and is almost independent of the wind direction,which is beneficial to simplify the inversion model in the application of wind field inversion.
Keywords/Search Tags:SAR image, full polarization, water segmentation, neural network, wind field inversion
PDF Full Text Request
Related items