| Synthetic Aperture Radar(SAR)is a kind of high resolution imaging radar.It has the characteristics of all-weather,all-time and non-environmental impact.The image it produces is called SAR image.Because of the advantages of SAR,the research of SAR image has become a hot spot in the field of remote sensing image.At the same time,with the continuous development of spaceborne and airborne platforms of synthetic aperture radar,it can easily imagine multiple time periods in various regions.Through the analysis of SAR images in different time periods in the same area,the process of detecting the change area of ground objects is called SAR image change detection technology.This technology can be widely used in many production and construction activities,such as agriculture monitoring,plant growth analysis,urban planning,natural disaster assessment and so on.Therefore,SAR image change detection technology has become a research hotspot in the field of remote sensing image processing.Deep learning simulates the working mechanism of human brain and establishes artificial neural network to process abstract data such as image,voice and video in real life.In recent years,the deep convolution neural network in deep neural network has been widely used in image processing because of its excellent image feature extraction ability.In this paper,convolutional neural network is used to extract features of two-period SAR images.Combining supervised and unsupervised classification algorithms,two kinds of SAR image change detection methods based on residual network and generative adversarial networks are proposed.The main points of this paper are as follows:1.In the process of SAR image change detection,it is necessary to filter SAR image and construct difference image.This paper presents a method of SAR image change detection based on residual network and fuzzy clustering,which avoids these two steps.By constructing a convolutional residual network,the neighborhood windows centered on the pixels of two SAR images are directly input into the residual network to extract the difference characteristics of the pixels of two-phase SAR images.Secondly,in view of the need for supervised training of training residual network,a label selection method based on spatial distance weighting is proposed to construct training samples and labels for training residual network.Then,the residual network and the fuzzy clustering algorithm are combined to classify the difference features of SAR image pixels in two periods learned by the residual network using the fuzzy C-means clustering algorithm.Experiments on different standard datasets show that this method has good detection performance.2.In view of the large number of noise points in SAR images,it is necessary to select appropriate denoise algorithms.Combining with the model of generative adversarial networks,a SAR image change detection method based on feature learning of generative adversarial networks is proposed in this paper.The method avoids the process of filtering operator and constructing difference image.By constructing two convolutional neural networks as the discriminator and generator to the generative adversarial networks model,then using the discriminator to extract the high-dimensional features of the two-phase SAR image in an unsupervised way.At the same time,combined with the proposed SAR image pixel point space distance weighted label selection method,the supervised training classifier is used to extract the high dimensional features of the discriminator.In view of the need to learn an effective generating network to generate high-quality false images,this method directly uses discriminators to extract high-dimensional difference features of neighborhood information of two-phase SAR image pixels,so it can suppress SAR image noise.Experiments on different standard data sets show that the proposed method achieves good performance with good anti-noise performance. |
