SAR Image Change Detection Based On Deep Learning And Conditional Random Field

Change detection of remote sensing images means to detect changes of image intensity or textures that occurred in the same area between different acquisition dates,and to get change information of interested ground object in shape,position,quantity and other property.Nowadays,change detection becomes an efficient data-analysis tool in application such as land use and land cover,environmental monitoring,disaster assessment,urban planning and monitoring,militarily strategic target monitoring and so on.Compared with optical multispectral data,synthetic aperture radar(SAR)is not sensitive to atmospheric and sun-illumination conditions,so multi-temporal SAR images were played an important role.However,SAR images are typically affected by multiplicative speckle noise due to its imaging mechanism.This poses challenges in the task of SAR images change detection.To solve the problems in SAR image and improve the SAR image change detection accuracy,this thesis complete the following three work:1.A new change detection method for SAR images based on multi-scale guidance image(MGI)and Stacked Denoised Auto-encoder(SDAE)is presented.Firstly,a new difference image is produced.Considering the heterogeneous and homogeneous information of the difference image,and using mean-ratio images with different neighborhood,the differences in the difference image can be highlighted effectively in our method.Then,the features of two SAR images and MGI are learned by use of SDAE.The two SAR images features are subtracted and the detection results are classified by use of FCM.The method has been demonstrated on six real SAR images that the segmentation accuracy has improved.2.A novel change detection method of SAR images based on manifold learning and stacked semi-supervised adaptive DAE is presented.First,a new semi-supervised Denoised auto-encoder based on manifold learning is proposed.Manifold learning has advantages to learn the nonlinear structure information in data.Partial label information is fused and dimension of hidden unit is reduced by use of LPP,so the features learned by SDAE are distinguishable.Then,combine the adaptive auto-encoder and semi-supervised auto-encoder,on one hand,the features extracted by semi-supervised auto-encoder is distinguishable,on the other hand,utilizing adaptive auto-encoder eliminates the difference in SAR images acquired by different SAR sensors.Finally,the two SAR images features acquired by semi-supervised adaptive SDAE are subtracted and the detection results are classified by use of FCM.This method has been demonstrated on six real SAR images,and the number of missed pixels and false pixels in final result is reduced.3.A novel change detection method for SAR images based on local characteristic Conditional Random Field is presented.First,a new multi-scale multi-direction gradient fusion distance image is produced by calculating the two SAR images,then it is fused with four direction CKLD,and a novel multi-scale and multi-direction gradient cumulant kullback-Leibler Detector(MDGCKLD)is produced,then as a local characteristic detector the MDGCKLD is added to conditional random field,and the change detection result is acquired by this modified CRF.Furthermore,log-ratio difference image is pre-classified by K-means,and the uncertain pixels in pre-classification results using the result achieved by the local characteristic conditional random field,and the certain pixels in pre-classification results using the result achieved K-means.The results on six real SAR images demonstrate the efficiency and effectiveness in the change detection task.