Speckle Reduction And Change Detection In Multi-polarization SAR Images

With the development of synthetic aperture radar(SAR)imaging technology,satellite constellation become the new trend of remote sensing.Through the formation of flight and network using different remote sensing satellites,the revisit period of the satellite is greatly shortened by constellation,the multi-temporal observation requirements for key areas are met,and a large amount of SAR images are provided.Among multi-model SAR systems,polarimetric SAR(Pol SAR)which can obtain multi-channel SAR data at the same time through different polarimetric radar waves,enriches the description of object scattering information on the ground.Pol SAR images can be used for continuous and stable monitoring of ground objects in many aspects,so how to combine multi-temporal Pol SAR images with big data technology is a challenging task.Under this background,this thesis focuses on the framework of change detection using dual-temporal Pol SAR images.Three aspects of analyzing Pol SAR images are studied: speckle reduction,unsupervised change detection,and supervised change detection.The change detection performances of the proposed algorithms are verified by Pol SAR images obtained from various satellites.The analyses and applications of multi-temporal Pol SAR images are established based on the proposed algorithm.In details,the research work of this thesis focuses on the following contents:1)Aiming at the problem that the inherent speckle of Pol SAR data will affect the human interpretation of the images,a despeckling method based on the pre-trained convolution neural network models is proposed.Convolution neural networks can automatically mine the deep features of SAR image,and have great advantages in the analysis of speckle noise of SAR image.But it is difficult to obtain the ground truth images of Pol SAR data,so training the network with the corresponding ground truth dataset is impossible.Therefore,this work considers using the pre-trained convolutional neural network models based on Gaussian white noise for Pol SAR images despeckling.Through the multi-channel logarithm with Gaussian denoising framework,the pre-trained models can be transferred to the despeckling of Pol SAR image,so the network parameters do not need to be learned again.This method has good despeckling performance on both mono-polarization and multi-polarization SAR images.The despeckling method can be used in the preprocessing of traditional change detection frameworks.2)Unsupervised change detection directly using the original Pol SAR data,which does not need the assistance of manual annotation information,can detect the changed areas of the image flexibly.In the framework of unsupervised change detection,the generation of difference maps is the most critical step,and the quality of difference map will directly affect the accuracy of change detection.For the unsupervised change detection at the object level,the Wishart mixture model is proposed to describe the statistical distribution of the super-pixel.Then,the similarity of the corresponding regions’ Wishart mixture distributions is measured by Cauchy-Schwarz divergence.Finally,the changed areas are obtained by thresholding.By using the Wishart mixture model,we can obtain a more accurate difference map and a higher detection accuracy in the change detection experiment of Pol SAR images.3)The supervised change detection method can detect the changed areas more accurately,and also can distinguish the types of changes by manually annotating a training dataset.Convolutional neural networks are composed of many independent neurons,which can be trained to distinguish the changed areas by learning the model from the training dataset.To solve the change detection problem of Pol SAR images based on deep learning,a dataset of urban change detection by using sentinel-1 dualpolarimetric SAR images are elaborately annotated,and then three different change detection frameworks based on U-Net are analyzed.Through the proper preprocessing of Pol SAR images,the changed areas can be accurately detected in the testing images after training the parameters of the change detection network with labeling samples.4)Based on the proposed algorithms of speckle filtering and change detection,the applications of multi-temporal Pol SAR images are analyzed.Through the visualization of Pol SAR image,we can clearly observe the changed areas in the multi-temporal data.The change detection vector and matrix of the multi-temporal Pol SAR images are obtained by using the change detection algorithm.The corresponding change detection matrix and vector using the Pol SAR images acquired by Radarsat-2 and Sentinel-1satellites are analyzed to show the change time points,change degrees,and change modes of different regions.