Change Detection Of SAR Images Based On Deep Neural Networks

Change detection in remote sensing images is used to detect changes in the same place at different time periods on the surface of the Earth.Because of the advantages of Synthetic Aperture Radar(SAR),which are not affected by time,weather and other conditions,the change detection technology based on SAR image has import research value.At present,this technology has attracted the attention of more and more researchers,and it has also been widely used in many fields,such as urban planning,disaster assessment,and forest early warning.This paper proposes the idea of combining the change detection of SAR images with the deep neural networks.We not only study the problem of change detection problem based on homologous images,but also the difficult heterogeneous images.The basic knowledge of SAR images is combined with deep learning,and some new change detection methods are proposed.The focus of this paper includes as follows:1.Feature extraction method based on deep auto-encoder has been studied.A change detection model in SAR images based on spatial fuzzy clustering and deep auto-encoder is proposed.In this method,the spatial fuzzy clustering is used to effectively suppress the speckle noise in the SAR images.Deep auto-encoder can automatically extract feature information from the image,extract the feature information of the difference image,and effectively reduce the impact of speckle noise on the images.This method does not require human intervention.The experiment was performed in different datasets and obtained ideal results.2.The convolutional neural network can make full use of the spatial information of the image and extract the deep features of the image.Based on this,an unsupervised change detection method in SAR images based on convolutional neural networks is proposed.This method uses the convolution neural network to directly generate the classification results from the original two SAR images.This method can ignore the process of generating different image,and therefore it will reduce the influence of the different image on the final classification result.The convolution neural network can fully learn the spatial characteristics and deep expression of images,and has stronger robustness.The basic idea is to first generate false label through unsupervised fuzzy clustering.Then,the convolution network is trained by suitable false labels.Finally,the final classification results are obtained through the trained convolution neural network.The entire process of this method is an unsupervised training process,and does not require prior knowledge and human intervention.This method has achieved good results in different datasets.3.Based on the study of heterogeneous change detection,a structure based on a full convolution symmetric neural network is proposed.This method converts the original images into the feature space through symmetric two networks containing only convolutional layers.The feature representation becomes more consistent and a different image is generated in the feature space.Through the iterative optimization of the network,we can get a good difference image.The learning process is completely unsupervised.By experimenting with homogenous datasets and heterogeneous datasets,compared with existing methods,the proposed method has good performance.