SAR Images Change Detection Based On Generative Adversarial Network

Remote sensing image change detection refers to analyzing remote sensing images of the same surface area in different periods to identify change areas.Synthetic Aperture Radar(SAR)image has become an important source of remote sensing data because it can ensure all-day and all-weather imaging and does not depend on the conditions of light and atmosphere.SAR image change detection is widely used in disaster monitoring,land use,military monitoring and urban change analysis.In this thesis,how to combine the generative adversarial network(GAN)to make better use of the difference information between two-temporal SAR images is studied.The main work is as follows:(1)A SAR image change detection method based on dual difference learning generative adversarial network(DDLGAN)is proposed.By the adversarial process between generator and discriminator in DDLGAN,the transition from one temporal image to another temporal image is realized.In the process of transition,the difference information extracted in the transition process can better distinguish the changed class from the unchanged class.This method is the first method to use the transition from one temporal image to another temporal image for change detection.When analyzing the extracted difference information,the parameters can be updated to the original image through the back propagation algorithm,thus avoiding the problem that the original image information will be lost in a class of methods of generating difference map and analyzing.Two generators in the DDLGAN method respectively extract the difference information of bi-directional transition between different temporal SAR images,and send it to the classification network for processing,which further improves the difference between the changed class and the unchanged class.The effectiveness of this method can be verified by the change detection results on multiple SAR image datasets.(2)A change detection method based on contrast constraint enhancement and feature difference fusion is proposed.There are differences in amplitude between two-temporal images due to different imaging conditions or the land cover type is unchanged but the number and state change.Firstly,the contrast constraint generative adversarial network(CCGAN)is used to generate images with smaller differences between unchanged class and larger differences between changed class,at the same time,the detailed information of the generated image is as close as possible to the original image,thus enhancing the contrast between unchanged class and changed class,enhancing the discriminant performance of image level,and making the judgment of unchanged class and changed class easier.The method then extracts features and obtains multi-level difference features by feature difference fusion network,and further fuses the multi-level difference features after preliminary fusion of up-sampling and down-sampling,thus making better use of the multi-level difference features of the deep neural network,enhancing the discrimination ability at the feature level.The experimental results prove the effectiveness of this method in the task of change detection.