Remote Sensing Image Change Detectioin Based On Generative Adversarial Networks

Remote sensing image change detection is a widely used technique that can obtain the changed region by analyzing a set of images taken at the same location but at different times.With the development of remote sensing technology,the number of images that can be obtained is increasing,and the diversity of images is increasing,too.How to process these complex images more intelligently and more accurately is an urgent need for change detection technology.Depending on the source of the image,change detection can be divided into two categories: homologous and heterogeneous.At present,the neural network-based homologous change detection technology generally adopts a discriminant model,which requires a large number of labeled samples for training,and is difficult to apply in practice.Heterogeneous change detection,because the images are acquired by different sensors,has different pixel representations for the same target,so it is not possible to compare directly in the observation space.As a generative model,the generative adversarial network can learn the distribution of data without the need for labeled training samples,and the confrontational learning can well realize the style conversion between two different types of images.So whether it is homologous or heterogeneous,the generative adversarial network can be applied to image change detection to solve some problems faced by the current algorithm.This article mainly contains the following three tasks:1)A hyperspectral image change detection algorithm based on generative adversarial network is proposed.The method can extract important features in the hyperspectral image,and while merging the image features,and then can learn the true difference information between the images from an initial difference image.Thereby,a higher quality difference image can be generated,and the detection accuracy will be improved.The method is completely unsupervised,and the robustness of the algorithm can be improved to some extent by means of the confrontation learning between the generator and the discriminator.2)A heterogeneous image change detection algorithm based on conditional adversarial network is proposed.The method converts an image into an observation space having the same pixel representation as another image by means of a conditional adversarial network’s image style conversion function,so that the heterogeneous images can be converted into homologous images.Then,by comparing and analyzing the difference information of the two homologous images,the change detection result of the heterogeneous image will be obtained.3)A heterogeneous image change detection algorithm based on coupled translation network is proposed.The method is based on the second work,combining the feature extraction of the variational autoencoders with the image style conversion function of generative adversarial networks,using a coupled translation networks,translates the two heterogeneous images in two opposite image translations.As a result,a set of heterogeneous images can be translated into two sets of homologous images,and then the two sets of homologous images are compared and analyzed,and finally the intersection of the two difference image is taken as the final experimental results.The method is completely unsupervised,and the images are translated into a aligned observation space for comparison.Compared with the current algorithms of comparing in the feature layer,this method can achieve higher detection accuracy.These three methods apply the generative adversarial network into homologous and heterogeneous change detection,and gradually improve the accuracy of the detection.The hyperspectral image change detection method based on generative adversarial network utilizes the generative model,which overcomes the problem that the discriminant change detection method requires a large amount of labeled data,and generates a higher-quality difference map unsupervisely.The heterogeneous image change detection algorithm based on conditional generative network is an exploration of applying image translation to heterogeneous change detection.By translating heterogeneous images into homologous,and then comparing the homologous images,the difference of heterogeneous images can be obtained.This method solved the problem that heterogeneous images cannot be directly compared in the original observation space.The heterogeneous change detection algorithm based on coupled translation network is a step further.A coupled translation networks can be used to translate a set of heterogeneous images into two sets of homologous images.The difference information between the two sets of homologous images is complementary and effective,improving the detection accuracy.These three methods proposed in this paper all use the generated adversaril network to solve some problems existing in the current image change detection.Each method is verified in the real data set,and the effectiveness is proved.