| Synthetic Aperture Radar(SAR)is a kind of high-resolution imaging radar system.SAR is operational under almost all weather and illumination conditions,and thus has found important applications in civilian and military fields.Until now,many SAR sensors still work on single-band and single-polarization,obtaining single-channel SAR images.These images have wide coverage range as well as high spatial resolution,which have been successfully applied to resources investigation,disaster monitoring,manmade targets detection and many other tasks.However,there’s only one complex scalar for each pixel of single-channel SAR images,so that its performance for some applications is restricted by the limited information content.Therefore,if the information content of each pixel can be expanded effectively,the performance of single-channel SAR images will be enhanced.Under such background,this dissertation focuses on the expansion of pixel’s information content as well as the representation of SAR images.Researches are carried out for the representation of contextual information on SAR images.The Directional Context Covariance Matrix(DCCM)is proposed,and then applied to SAR image terrain classification,ship detection and change detection.The main contributions of this dissertation are as follows:(1)A novel representation for the contextual information on SAR images is proposed—the DCCM.The target pixel is extended from a complex scalar to a group of matrices in DCCM representation by combining pixel intensity with spatial information in the neighborhood.The information content of each pixel is expanded remarkably through such representation,while the information aspect is increased as well.By obtaining the variance of pixel intensity on each orientation inside the neighborhood,DCCM is able to depict the local contextual information more accurately.Then,DCCM is applied to analyzing SAR images slices of different land cover.Comparing with three typical texture features,the gray level co-occurrent matrix(GLCM),Gabor filters and multilevel local pattern histogram(MLPH),the slices achieve the highest ratio of inter-class distance to inner-class distance in DCCM feature space.The result indicates that DCCM is able to extract information from SAR images more effectively,so that it has an advantage on discriminating different land cover on SAR images.Therefore,DCCM could be a promising method for several SAR image applications,such as terrain classification,ship detection and change detection.(2)A SAR image classification method based on DCCM texture feature is proposed.Firstly,the correlation values are normalized in order to reduce the redundancy.Secondly,logarithmic transform is applied to enhance the contrast of the dark areas on the image,so that the textures inside could better reveal.Thirdly,DCCM texture feature is obtained by selecting some of the elements in DCCM and reconstruct them into a vector.Then,a DCCM based SAR image classification method is proposed.Combining the proposed method with decision tree classifier and support vector machine classifier,respectively,terrain classification experiments are carried out.Classification methods based on GLCM,Gabor filters and MLPH are taken into comparison.Experimental result shows that,the overall classification accuracy increased by 7% with the proposed method on both UAVSAR and AIRSAR datasets.Furthermore,the feasibility of combining DCCM texture feature with deep learning methods is explored.The texture features are taken as the input of a convolutional neural network(CNN)and comparative experiments are carried out.The classification result shows that the CNN with DCCM texture feature as input gets the highest overall accuracies on both dataset,which are 87.07% and 94.19%,respectively.This indicates that the combination of CNN and the proposed method can further improve SAR image classification performance.(3)A constant false alarm ship detection scheme based on the maximum value detector is proposed.First,the maximum value detector is proposed in order to enhance the contrast between ship targets and sea clutter.According to the difference in distribution continuity between target pixels and high-intensity clutter pixels,the maximum value detector improves contrast by means of preserving strong scattering points.The analysis with target-clutter ratio and statistical histogram confirms that the maximum value detector can effectively enhance the contrast between targets and clutter,as well as increase the detectability of targets.In particular,the target-clutter ratio of weak targets is improved by at least 10%,and the overlap area of histogram is cut down by 40%.Thereafter,a constant false alarm ship detection scheme based on maximum value detector is proposed,and comparative studies are conducted.On both Radarsat-2 dataset and Gaofen-3 dataset,the proposed scheme gets the highest figure of merit,reaching 100% and 96.79% respectively.The experimental result demonstrates that the proposed scheme can not only detect ship targets accurately,but also preserve their morphological characteristics well,which can be helpful for the recognition of target type in subsequent applications.(4)A SAR image change detection method based on DCCM and eigenvalue decomposition is proposed.First,elements contain intensity information and spatial information are extracted from DCCM to form a variation matrix for each pixel.Second,eigenvalues of variation matrices are obtained through decomposition,after which the variation detectors are calculated.Then,a series of operation are conducted to enhance variance information,including orientation averaging,two-way processing,median filtering and normalization.Finally,the difference map with spatial information is obtained for change detection.Analysis on multi-temporal UAVSAR dataset shows that the proposed method can well locate the regions with intensity or texture variations on SAR images.Meanwhile,it can also indicate the level of changes.In comparison with the commonly used subtraction operator and log-ratio operator,the proposed method has better sensitivity toward the changed regions with lower intensity as well as the regions with texture variation. |
