Deep Auto-encoder Framework For SAR Images Change Detection

As the representative of microwave remote sensing,synthetic aperture radar(S AR)not only has the characteristics of wide coverage,large amount of information,fast access to in-formation that other general remote sensing has,but also has the characteristics of all-day,all-weather and others that are not affected by light and climate environment.So it is widely used in many fields such as national defense security construction and national economic de-velopment.Among them,SAR change detection is a key component of SAR image interpre-tation,which has been widely concerned by scholars at home and abroad.However,due to the inherent imaging mechanism of SAR images,speckle noise inevitably exists,which has an important impact on SAR change detection.In order to suppress the influence of speckle noise on SAR change detection and improve the detection accuracy,this dissertation com-bines multi-scale features,extracts discriminant features by using deep auto-encoder model to study SAR image change detection.The main contents of this dissertation are as follows:1.A fast unsupervised deep fusion network(FuDFN)for SAR image change detection is proposed.The main purpose of this method is to generate difference image in feature learn-ing process by using stack auto-encoder.Compared with shallow network,the framework can extract more useful features,which is conducive to better detect change results.In addi-tion,we have found a training subset of the whole samples,which can properly represent the whole samples.It can not only accelerate the training of deep neural network,but also avoid underfitting.Furthermore,we design a fusion network framework,which can combine the ratio operator based method to ensure that the higher level representation is better than the lower level representation.Experiments on four real synthetic aperture radar images show that the network is superior to the traditional ratio method and convolution neural network.2.A supervised variational auto-encoder(SVAE)is designed to study the representation ability of SAR and to detect change results.Firstly,two original images are preprocessed,and the difference image is obtained by logarithmic ratio method.Then,the difference image is analyzed by fuzzy C-means clustering(FCM),and the pseudo labels is obtained.For the input of SVAE,it is directly selected from two SAR images instead of sampling from DI to avoid information loss.Having input and pseudo labels,SAVE can learn latent representations that obey the Gauss distribution.According to the latent representations,SVAE can obtain the final change detection results.Experiments on four data sets show that SVAE can obtain discriminant features which are beneficial to change detection,and the results of it are better than that of other common methods.3.A feature extraction method is proposed,which combines convolution Siamese auto-encoder and recurrent network(RSCAE)to learn temporal and spatial features.The former can produce stronger spatial feature representation to suppress speckle noise,while the latter can effectively simulate the temporal relationship of dual-temporal SAR images.Firstly,the convolution Siamese auto-encoder is designed to learn more powerful spatial feature repre-sentation from a large number of unlabeled data.Secondly,the recurrent network is used to explore the temporal dependence of the two temporal images to extract more discriminant temporal and spatial features.Finally,discriminant spatial-temporal features are input into a Soft-Max layer to predict the final change results.The effectiveness and robustness of the method are evaluated on three sets of real SAR image data sets and one set of simulated images.4.Inspired by human visual cognitive system,a multi-scale visual cognitive(MVC)network is proposed,which detects the temporal dependence between multi-scale spatial features of two temporal images.The network consists of three blocks:unsupervised visual block,read-write based memory block and supervised cognitive block.Visual block can generate multi-scale spatial feature representations stored in memory block,and suppress speckle noise in an unsupervised way.Cognitive block can effectively model multi-scale spatial feature pairs stored in memory blocks to obtain multi-scale temporal-spatial feature relations.The visual block is composed of convolutional autocoder,which aims to learn robust multi-scale spatial feature representation without supervision.Cognitive block is composed of a recurrent network and a Soft-max classification layer.The purpose of this block is to model the multi-scale temporal dependence of two images and predict the final labels.The network is inspired by human visual cognitive system,which not only improves the detection accuracy,but also suppresses speckle noise.Experiments on three sets of real SAR image data sets and one set of simulated images show the effectiveness,robustness and superiority of the proposed network.